Implementation evidence

Telegram work-buddy personality (2026-10-18)

The assistant now loads SOUL.md into its private, group and scheduled reply prompts. It defines a warm, dependable work partner, natural language matching, optional “小伙伴” / “buddy” forms of address and sparse humor/emoji. Failures, risks, approvals and formal reports use a calmer tone. The base policy permits friendly phrasing while retaining concise answers, evidence requirements, approvals and private/group boundaries.

Fixed queue, investigation, implementation, repair, waiting and generic failure notices use conversational wording. Confirmed-state routing, controls, deduplication and delivery timing retain their existing contracts. The Markdown is copied beside bundled Node entrypoints and loaded using the module URL; source processes load it directly. Restart source processes, or rebuild and restart bundles, after editing it. No dependency, setting or migration was added.

The UI preference, user guide and runnable pilot acceptance examples are updated. Validation passed Biome, strict TypeScript, production build, 887 deterministic tests across 101 files in the shared working tree, the host Node runtime contract and the Node 24 runtime contract in the built application image with networking disabled. Compose config also validates. Coverage includes prompt integration, factual notices, safe failure guidance and the existing delivery/permission/cancellation contracts. Live model tone and native Telegram desktop/mobile display, including waiting, failure and final states, remain unverified. No live model call, Telegram send, account connection or deployment was performed.

Code Truth repository descriptions and automatic IDs (2026-10-18)

Code Truth repository modals now ask only for the GitHub URL and an optional description of up to 2,000 characters. Internal IDs are generated automatically; existing IDs survive edits. Saved cards show the GitHub repository name and description as text. Network/branch controls and their nested modal were removed. Descriptions and repository URLs reach the assistant through list_code_targets so it can select generated target IDs without asking users for them.

Indexing resolves the repository's current default GitHub branch on each refresh and retains its actual branch name and immutable commit as provenance. The upstream MCP adapter retains one internal snapshot selector. Legacy saved network mappings remain readable, are omitted from settings responses and are removed on save; they no longer select branches. Legacy duplicate-URL records remain editable. Matching app/worker and Code Truth service builds are required for this behavior. The runtime skill, API example, setup docs and scoped UI preference were updated.

Validation passed Biome, strict TypeScript for both packages, production build, Node runtime contract, 882 deterministic tests across 100 files, and 25 admin and repository browser scenarios. The four focused Code Truth browser scenarios passed again after final spacing cleanup. Desktop/mobile saved displays, editors, loading/empty, pending, validation and recoverable error states were visually inspected. The real CodeGraph/MCP smoke passed with automatic branch discovery. No dependency or SQL migration was added; live GitHub calls and deployment were not performed.

Cloudflare connection discovery (2026-10-18)

Cloudflare OAuth now starts without account/zone IDs. The callback discovers authorized accounts, automatically selects a sole account for Pages-only access, and asks for a named workspace account when several are available. Configuration uses named account choices and searchable, paginated domain checkboxes in the shared modal. Saved summaries show verified names; manual API-token setup remains an explicit fallback. No domains are granted automatically. Operators still register/configure their OAuth client once.

Inventory endpoints project only names/IDs and require the authenticated workspace operator, Origin, CSRF and current revision. Provider responses are fenced against disconnect, role/credential revocation and settings changes. Saving a new account or domain selection verifies provider access and each domain's account before the local grant changes. Discovery works while configuring a disabled plugin and does not expose inventory to model tools. Missing Zone Read still allows an explicit Pages-only selection. UI rules and runnable API examples are updated.

Validation passed Biome, strict TypeScript, production build, Node runtime contract, all 875 deterministic tests across 98 files, and 22 Cloudflare browser scenarios at 1280px/390px. Desktop/mobile saved displays, editors, account choice, loading/empty, pending, failure/recovery and dark theme were visually inspected. No dependency, migration, live Cloudflare connection, provider mutation, Telegram send or deployment was introduced; live Cloudflare consent remains an acceptance gate.

Automatic Codex PR publication (2026-10-18)

Removed the publication checkbox from the Codex configuration dialog and its Enabled/Disabled summary row. Verified implementations automatically publish PRs ready for review. Policy parsing fixes publishByDefault to true, including saved false values, so removing the control leaves no hidden opt-out. Task authority snapshots retain the fixed value; existing tasks with a different snapshot stop at the configuration fence instead of acquiring new publication authority. Analysis, questions, evidence checks, current GitHub permissions and verification requirements retain their existing boundaries.

Validation passed Biome, strict TypeScript, build, 867 deterministic tests, 35 admin/loading browser scenarios and 20 plugin permission/modal scenarios. Desktop/mobile screenshots cover saved summaries, light/dark editors, loading/empty states, pending saves and recoverable errors. Regression tests cover automatic publication without an explicit PR request, duplicate polling, read-only analysis and cancellation of retained tasks with the old opt-out authority snapshot.

GitHub App setup without icon uploads or a welcome modal (2026-10-18)

GitHub App creation no longer offers an icon download, upload instructions or an upload acknowledgement in setup or Manage GitHub. Setup proceeds directly from manifest registration to account authorization. Failed authorization remains retryable, and an already authorized or connected workspace does not restart authorization. Owners may customize the default badge independently in GitHub. Once repository connection succeeds, setup opens the workspace directly; the Welcome to RepoDesk modal, Get started action and confetti were removed. The unused icon component, asset, styles and build copy were removed. The current UI preference supersedes the earlier guided branding step.

Validation: Biome, strict TypeScript and production build passed. All 27 admin and focused onboarding browser scenarios passed with fake GitHub transports. Desktop (1280px) and mobile (390px) creation/saved displays, loading, authorization pending/error, installation waiting and direct workspace entry were visually inspected. The final full Bun run passed 863 of 864 tests in the concurrently edited workspace; the unrelated Review Bot test "Telegram follow-ups wait for a Review Bot fix on the same PR" reproducibly received cancelled instead of queued. Earlier access-link and Cloudflare OAuth failures passed in isolation and on the final full run. No live GitHub account was connected for validation.

v0.1.40 release preparation (2026-10-18)

Version advances to 0.1.40, covering continuous Codex conversations and the single coding workflow, scoped long-term memory, conversation-thread history and statistics, and the Cloudflare plugin with browser authorization.

Release preflight found remaining admin/loading browser assertions for the removed execution summary and Reviewed/Direct selector. They now check the publication summary, absence of the selector and persistence of the independently saved PR publication setting. Application behavior is unchanged by this correction. Local Biome, strict TypeScript, build, Node runtime contract, 863 deterministic tests and 35 admin/loading browser scenarios passed. Stable publication requires the full Verify workflow to pass for the release commit.

The Reviewed removal is intentionally breaking: executionMode, issue-only runner starts and the separate developmentTasks response collection are removed without compatibility or migration. App/worker/supervisor/job builds must be upgraded together. See the current Codex workflow and the release notes before upgrading.

One Codex task workflow (2026-10-17)

Codex now uses one continuous task workflow. Removed Reviewed proposals, the coding_task approval kind, legacy workspace task model/executor, and the old Issue/codex exec runner path. There is no migration or compatibility bridge. The strict settings contract rejects the removed executionMode field; publication by default remains independent. Requests without that optional publication setting can start source-bound intake normally. Existing GitHub issue/PR action approvals remain separate application features.

The runner requires a continuous task envelope and rejects issue-only starts before reserving execution. Questions and ordinary chat cannot grant implementation or publication. Verified actor/repository access, original request evidence, audience, cancellation, retention and durable publication checks remain enforced. Review Bot fixes retain their own enabled setting and GitHub write checks without a Codex mode dependency. A regression checks revocation before execution.

The admin configuration summary/editor has no mode selector or task-mode badges. GET/PUT/logout return the same tasks collection, also used by Overview counts and Stop. Saved summaries, editing, loading/empty, pending and failed-save recovery were checked at 1280px/390px, including dark theme. The shared modal and Stop confirmation contracts remain in use. Documentation, bundled tool guidance and examples agree.

Validation: Biome, strict TypeScript, build and 829 deterministic tests across 95 files passed in the isolated feature checkout with disposable PostgreSQL. The merged shared checkout also passed Biome, types, build and 863 tests across 98 files, including the independent Cloudflare changes present there. The relevant 40 browser scenarios passed; the final configuration copy/assertions and Stop flow were rechecked independently. App, supervisor, Docker job and OpenSandbox job images built; the final app passed the Node 24 runtime contract. Both Compose overlays validate. The full Docker lifecycle smoke passed with continuous provider/device tasks, live questions across restart, checkpoint recovery, repair, publication fencing, auth continuation, usage, erasure and cancellation. A fresh one-iteration Docker benchmark fixture also completed with no failures.

No dependency or database migration was added. Upgrade app/worker/supervisor/job builds together. No live model, GitHub/Telegram write or deployment was performed.

Conversation-thread history and Telegram identity (2026-10-17)

The admin history list now shows continuous conversation threads at /admin/threads, grouped by valid saved routing identity before pagination and ordered by latest activity. Equal timestamps follow retained request order. Cards show the initial subject, latest preview, request count, status, Telegram names/usernames/IDs, group titles/IDs, topic IDs and participants. Missing profiles use IDs. Legacy requests without a valid thread and scheduled requests stay separate; ignored or unaccepted automatic follow-up classifiers are excluded.

Thread detail pages show chronological request/reply pairs, whole-conversation statistics and per-request links to existing run diagnostics/actions. Thread and request pages retain independent pagination through navigation and refreshes. Sidebar/Overview use Threads and the matching count; old run-list URLs redirect with their query parameters. Existing run-detail URLs and the request-summary API remain available. Both new read endpoints preserve workspace admin/operator access, including private histories; non-admin members, revoked members and foreign workspaces remain denied. Telegram metadata is saved display data, not authority. No dependency, database migration or live Telegram profile lookup was added.

Validation: bun run check, bun run typecheck, bun run build, all 869 deterministic tests across 97 files with disposable PostgreSQL, and all 45 browser scenarios across Threads, run details, Overview loading and admin passed. Desktop/mobile screenshots at 1280px/390px were reviewed in light/dark themes, including conversation identity, ordered requests/replies, pending/partial usage, loading, empty and error/retry states. Checks cover pagination before/after grouping, equal timestamps, legacy isolation, shared participants, profiles/ID fallbacks, unlinked operator reads, member/revocation denial, cross-workspace IDs and retained run navigation. No deployment, live model call or Telegram message was performed.

Cloudflare plugin (2026-10-17)

Plugins now includes Cloudflare with a saved summary, title switch and shared modal editor. Operators configure an encrypted write-only workspace token, account and up to 20 explicit zones. Private interactive Pi requests can query selected zones/DNS records and account Pages projects/deployments. Owners/admins can propose exact DNS create/update/delete requests and URL/whole-zone cache purges. Existing actor-bound approval controls authorize the reviewed payload; application services reserve and execute writes. Duplicate workers cannot repeat mutations, and ambiguous outcomes stay unknown without automatic retry. Pages environment variables and configuration secrets are excluded. Current permissions/scope apply to reads, approvals, execution, pending delivery and retained/derived evidence, including group-handoff exclusion.

Cloudflare saves preserve legacy file-plugin fallbacks and other built-ins; Code Truth saves now preserve other built-in settings too. No package, migration or new environment setting was added. Configuration changes apply without restart. Workers, deployment creation, secrets and arbitrary API access remain outside scope.

Validation: Biome, strict TypeScript, production build, Node runtime contract and all 863 deterministic tests across 96 files passed with disposable PostgreSQL, including 20 focused Cloudflare tests. All 8 Cloudflare browser scenarios passed at 1280px/390px; saved/editor, empty, loading, pending, error/recovery and dark theme screenshots were inspected. The expanded browser run passed 49 of 50: an existing Codex mobile loading test still measures its unchanged “Deployment runtime” value at 48px against a 32px limit, reproduced in isolation. No live Cloudflare/model/Telegram calls or deployment were performed. See configuration, API contracts and limits and runnable examples.

Cloudflare browser authorization (2026-10-17)

New Cloudflare connections default to browser OAuth; existing API token connections retain their authentication method. Operators register a Cloudflare client once, then save its write-only secret, client ID and API scopes in the shared Authorization editor. Connect uses Authorization Code, S256 PKCE and session/workspace/revision-bound state. The callback verifies the configured account before enabling the plugin, including the same tool-collision checks as manual enablement. Existing account, zone, actor, destination and exact-action approval restrictions still apply.

Access/refresh tokens are encrypted per workspace and connection generation. Durable refresh claims serialize workers without changing the plugin revision; failed, ambiguous or interrupted refresh requires Reconnect without token replay. Disconnect removes local authority before provider revocation, reports unconfirmed revocation and fences late refresh results. Client/scopes/authentication changes require disconnect first. Client registration, public-client domain verification where applicable and live Cloudflare consent acceptance remain operator prerequisites.

Validation: all 863 deterministic tests across 98 files, including 34 focused Cloudflare tests, and 14 Cloudflare browser scenarios passed. Desktop/mobile saved summaries, both editors, loading, pending, failure, reconnect and disconnect states were visually inspected, including dark theme. Biome, strict TypeScript, production build and Node runtime contract passed. No package, migration or environment variable was added. No live account connection, Cloudflare API mutation, Telegram send or deployment was performed. See the registration and connection guide.

Codex task conversation (2026-10-17)

Direct continuous development tasks have a durable conversation separate from implementation requirements. Chat Agent can formulate factual questions with ask_development_task, read bounded history and receipts with development_task_conversation, and relay new requirements through the existing source-bound input tool. Questions never advance requirement revisions or grants. Requirement receipt, dispatch, completed-turn reading and verification are separate.

Active implementation processes use an attempt-scoped mailbox and pinned app-server turn/steer. Only explicit read/reply/progress tools cross this channel; other server requests remain unsupported. Tokens stay out of the Codex child and other phases, survive supervisor restart and rotate between launches. Claims persist before steering. Idempotent mailbox acknowledgement retries cannot resend native input; unconfirmed native delivery stays unknown. Replies and meaningful progress arrive asynchronously in the original audience, with history available to Chat Agent.

Completed/waiting tasks can answer through a separate read-only execution without changing their result, product question, PR or authorization. The existing unique live-attempt constraint and a conversation pointer serialize subsequent coding behind cleanup. Usage settles once. Frozen attempt identities handle unknown starts; durable erasure tombstones also fence starts delayed until after cleanup. Source expiry/deletion covers the dialogue and all recorded executions.

Validation on an isolated v0.1.39 checkout plus this feature passed Biome, strict TypeScript, production build and 807 deterministic tests across 91 files with disposable PostgreSQL. This avoids unrelated concurrent UI, memory and Cloudflare work in the shared checkout. The real pinned CLI 0.155.1 passed a local fake-provider proof for initial/live questions, reading, answers, progress, token accounting and restored tools on session resume. App, supervisor, Docker job and OpenSandbox job images built; the app passed its Node 24 contract and both Compose overlays validated. The full Docker lifecycle smoke passed, including duplicate questions and an active conversation surviving supervisor restart. Earlier smoke failures exposed a fake GitHub transport intercepting the private channel; the fixture now permits only those two scoped channel endpoints through its real local transport.

No dependency or database migration was added. Matching app/worker/supervisor builds are required. Live model quality, Telegram delivery and a deployed OpenSandbox repeat remain separate rollout checks. Reviewed legacy task coverage is not implied by this Direct-task evidence. See the conversation journey.

Private and shared long-term memory (2026-10-17)

RepoDesk now extracts source-linked historical memory from successful idle conversations and consolidates a small recall index. Personal recall crosses the same user's private Topics; shared recall crosses participants only within the same bot/group/Topic. Workspace, actor, bot and source permissions remain enforced. Existing discussion records and context compaction remain separate.

The independent worker checks once per minute, with a 30-minute idle window, bounded tool-free Pi extraction/consolidation, durable leases and checkpoints, hourly retry backoff, and no automatic replay of unknown provider outcomes. Charges use the originating request's pinned model and remaining dollar budgets and are recorded as separate memory attempts. Recent input is bounded while full retention dependencies remain tracked. State uses existing workspace JSONB; there are no added dependencies or SQL migrations.

query_memories supports bounded historical recall; direct manage_memory requests remember the human's actual message, forget a selected record, or toggle scoped recall. Shared settings and deletion of another speaker's or mixed-speaker records require a workspace owner/admin. Source edits, retention, forgetting and workspace deletion invalidate derived context. Content-free tombstones prevent forgotten sources from being automatically re-extracted. Memory cannot grant external-write authority or become an approved instruction.

The 31 dedicated deterministic memory tests passed with disposable PostgreSQL. An isolated v0.1.39 checkout containing only this feature passed all 820 tests, Biome, strict TypeScript, the production build and the Node runtime contract. Coverage includes JSONB identity stability, private/shared isolation, current-user intent, ownership, source edits/deletion, long input, redaction, budget failures, concurrent claims, usage replay and checkpoint recovery. No live model call, Telegram send or deployment was performed. Recall fidelity and live Telegram acceptance remain release gates. See design and limits and examples.

Run detail statistics (2026-10-17)

Individual run detail pages now lead with message/tool-call counts, model attempts, reported total tokens, cache hit rate and recorded actual USD spend. Message counts share the displayed transcript's entries; tool calls merge retained call IDs with durable records without counting tool results again. Tokens and charges aggregate all recorded attempts, including follow-up and compaction. Cached reads and writes count toward input; only cached reads count toward the hit rate. Missing and partially reported usage/charges remain explicit. No API, dependency or migration was added. Runs remain individual requests; continuous chat threads are distinct.

Validation: strict TypeScript, production build, scoped Biome checks, all 794 deterministic tests with disposable PostgreSQL, and all 15 Runs browser scenarios passed. Desktop/mobile screenshots were reviewed for populated, loading, error, pending, partial and empty usage states, including light/dark themes. Repository-wide Biome reported formatting/import issues in concurrently edited Codex conversation files outside this change. No deployment or live provider/Telegram call was performed.

v0.1.39 release preparation (2026-10-17)

Release preflight caught Docker-specific credential ownership checks leaking into the simulated runtime under the CI runner's UID, plus browser fixtures missing the new runtime endpoint and expecting the previous configuration label. Ownership preparation now belongs to the Docker driver; real Docker retains its private credential permissions and job UID requirements. Browser fixtures supply a valid runtime response and assert the current deployment-runtime label.

Biome, strict TypeScript, the production build and all 789 deterministic tests across 90 files passed with disposable PostgreSQL. The 32 runner tests also passed in a network-disabled UID-1001 container, and 23 browser scenarios passed for loading, connection-error recovery and session expiry. Full GitHub Verify and exact-image checks remain separate release gates until their results are recorded in the published release notes.

Telegram PR code and review evidence (2026-10-17)

The default GitHub extension now registers query_github_pull_request alongside the existing metadata tool. Members can read each connected-repository PR's changed files/patches, review records and inline comments for implementation comparisons. The bundled GitHub skill directs comparisons to actual code evidence and distinguishes a summary-tool capability gap from a failed GitHub permission check. Missing installation grants return specific recovery guidance.

Reads request only a fresh, single-repository Pull-requests-read installation token. Existing tenant, member, upstream repository, private-chat and approved workflow-source boundaries apply throughout network reads. Each page checks PR identity and head/base commits before and after fetching; later pages require the returned commit pins. Changes reject mixed-commit evidence. Response budgets, missing patches, pagination and the 3,000-file/150-page limits remain explicit; review records are live and retain their own commit provenance.

Validation: Biome, strict TypeScript, production build and the Node runtime contract passed. All 789 deterministic tests across 90 files passed with disposable PostgreSQL, including ten new PR-detail scenarios and default-agent tool registration. Coverage includes differing PR patches, read-only token scope, pagination and missing evidence, commit changes, permission revocation at each network stage, tenant/member/private-group boundaries and missing-grant guidance.

No dependency, migration, write grant or new deployment configuration was added. Live GitHub/model comparison quality and rollout to the existing app/worker remain release gates. See setup and limits and the runnable comparison journey.

Admin-panel Codex runtime selection (2026-10-17)

Plugins → Codex → Runtime → Edit selects Docker or OpenSandbox for the whole deployment, or follows CODEX_CONTAINER_ENGINE. The saved panel override survives supervisor restarts. Optional per-engine images, OCI sockets and shared sandbox paths prepare alternate host profiles; the panel never returns or edits their management credentials.

Operator/session/Origin/CSRF checks protect changes, followed by the supervisor's authenticated management API. Revision checks reject stale writes; task starts and switches are serialized. A different engine must pass readiness before saving and cannot replace an engine with retained tasks/checkpoints, including completed work. Corrupt, oversized or symlinked saved selection files fail closed. Unconfirmed saves, including a failed audit acknowledgement after the runner commits, instruct the operator to read the current selection before retrying.

Validation: Biome, strict TypeScript, production build and 789 deterministic tests across 90 files passed with disposable PostgreSQL. 36 relevant browser scenarios passed across runtime selection, account authentication, Codex Stop, overview and logs. Runtime saved/editor, pending, loading, error and unavailable states were inspected at 1280px and 390px, including dark mode. The mobile Edit action stays at the top right. Final app/supervisor images built; Node runtime and authenticated runtime API/persistence/restart image checks passed. Docker Compose validation and the full Docker smoke passed, including checkpoint repair, publication fencing, device continuation, erasure and cancellation. No application deployment or production backend switch is implied. microsandbox and Podman were removed at the user's request, including adapters, profiles and panel options; removed environment/saved selections fail closed and browser coverage asserts that the choice is absent. See configuration and limits.

The default and local startup script use Docker. Shared container helpers moved to container.ts; no Podman executable is installed in the supervisor image. Its uncompressed size dropped from 951,122,785 to 849,540,010 bytes (about 97 MiB). Native UID 1000 supervisors retain private task credentials without a root-only ownership change; the root container path preserves the job UID before copying. The runner guide is now codex-runner.md. Existing approval/protocol wire values remain stable, with no Podman execution path or automatic data migration.

Landing-page live-test environment (2026-10-17)

The dedicated landing repository has a committed baseline-v1, a published Cloudflare page and a manually triggered branch preview. Local/CI checks and desktop/mobile browser acceptance passed. The existing VPS Codex job image also passed a credential-free fresh-clone/install, check, typecheck, unit-test and build preflight in disposable isolated storage. The existing GitHub App can access the new repository; a host-operator action selected it for the workspace and synchronized the existing verified user account.

The requester approved the concrete proposal through the existing Reviewed flow. The real Docker task produced PR #2 with passing Verify jobs and a native Cloudflare github:push preview. The earlier automatic-preview gate is resolved by that observed deployment. Additional VPS OpenSandbox acceptance produced PR #3, also with passing Verify/Cloudflare checks and desktop/mobile deployed-preview acceptance. The five-trial fixture comparison, failed first OpenSandbox attempt, prepared-browser retry and microsandbox exclusion are recorded separately in the VPS comparison. See the acceptance receipt.

Alternative Codex sandboxes and benchmark harness (2026-10-17)

The local supervisor supports deployment-selected OpenSandbox alongside Docker through a typed phase-runtime interface. Existing workspace policy, authentication ordering, task fencing, repair and publication remain in the supervisor. The OpenSandbox SDK dependency is pinned to 1.1.1 and loaded only when selected. OpenSandbox uses private authenticated management and shared host storage.

Fresh phase environments retain task-scoped work storage, read-only inputs, implementation-only account credentials/model tokens and separate publication storage. Durable sandbox/job identities survive restart; uncertain launches do not replay. Cleanup retries after failures, and retained records reject backend switches. OpenSandbox has a dedicated execd-compatible image target, pinned companion images, deny-default egress and documented SELinux storage labeling.

bun run benchmark:sandboxes runs a credential-free Git/file/check/patch and simulated-publication fixture across selected engines. It records phase and operation timings, P50/P95, host/SDK metadata and explicit failures, and cleans only its own task identities. Real Docker and self-hosted OpenSandbox have each completed the entire fixture on this host. The earlier microsandbox experiment was removed at the user's request after the VPS lacked KVM; its SDK, adapter and configuration are absent. This does not measure live model execution, dependency downloads or remote GitHub publication. Prewarmed pools, memory snapshots and cross-task dependency caches are not enabled. See setup, commands and verification boundaries.

Validation: Biome, strict TypeScript, production build and all 767 deterministic tests across 87 files passed with disposable PostgreSQL. The 48 focused runner tests include new phase/credential isolation, restart/deduplication, uncertain launches, cleanup retry, artifact bounds/symlinks, backend mismatch and SDK command identity/ownership checks. Initial full runs encountered two unchanged flaky fixtures; their focused reruns and the final full suite passed. That initial local implementation pass performed no migration, production backend switch, live model/account call or GitHub/Telegram write. Subsequent authorized VPS acceptance is documented separately above.

Final app, supervisor, ordinary job and OpenSandbox job images built. The app's isolated Node runtime contract, Docker Compose validation and the real Docker smoke passed, including fake provider/device auth, parallel capacity, restart, question reconstruction, repairs, publication fencing, erasure and cancellation. The initial local sandbox comparison also passed without failures; its small-sample timings and limits are recorded in the sandbox guide. Temporary test database, task containers, comparison network and OpenSandbox server were removed/stopped.

Invite links use a circular countdown (2026-10-15)

The invite card replaces its visible countdown text row with a small animated circle immediately after the title, keeping the final word and circle together when the title wraps. The ring counts down the existing ten-second refresh, spins while refreshing and pauses on errors. Screen-reader text and a hover tooltip retain timing/status; reduced motion disables transitions and rotation.

Verification: Biome, strict TypeScript, production build and all 751 deterministic tests passed with disposable PostgreSQL. Eight focused browser checks passed at 1280px and 390px, covering placement, countdown/reset, copying, member updates, loading, pending, errors/retry, navigation cleanup and reduced motion. Light/dark saved cards and desktop/mobile loading, pending and error previews were inspected.

Compact skill instruction previews (2026-10-15)

Preset and custom skills share a compact saved preview: labeled settings in a responsive grid, required tools as wrapping tags, and a short empty-workflow message. Complete instructions and populated dependent-workflow details remain readable. Management actions stay at the top right when titles wrap; custom editing remains in its modal and presets retain toggle-only controls.

The same Team recap browser fixture shrank from 756px to 393px at a 1280px viewport (48%) and from 968px to 625px at 375px (35%). Both themes, expanded and collapsed previews, pending/error toggles, long content, empty tools/dependencies and saved display/editor separation were checked with deterministic fixtures. Biome, strict TypeScript, the production build, all 751 deterministic tests with disposable PostgreSQL and 48 browser scenarios across skills, GitHub imports, workflows and admin management passed. A broader run encountered a missing trace artifact during browser cleanup; the 26-scenario admin/preset rerun passed with an isolated server and output directory. No dependency or database migration was added.

VPS deployment of invites and member account updates (2026-10-14)

The VPS at https://dev.stonegate.me serves package v0.1.37 from clean commit 1afd6f5f58b2a5b31a9e1006ec6bcdc7ffadd526, deployed through manual archive bundle 1791965539905-97510167. This includes single-account invites with the ten-second link countdown, corrected GitHub member sources/custom usernames and the member GitHub View text button.

The commit's Verify run passed. The exact release images also passed local Node runtime, Compose health, migration rerun, protected backup and disposable restore checks before transfer. The VPS verified the image/source checksums and coding checkpoints, created and verified its protected pre-release backup, and completed migration/startup checks. All four production services are healthy. Public health, readiness and admin requests returned HTTP 200; all three served JavaScript/CSS assets matched the verified image. The bundle retains CI, source and post-deployment receipts.

Runtime configuration, encryption key, runner token, persistent database/runner volumes and the existing PostgreSQL/CI runner container identities were preserved. No migration was added. This verifies a manual archive cutover; the published release updater's direct local-image path remains a separate live gate.

Single-account invites and link countdown (2026-10-14)

Members & access now shows a signed Telegram invite that refreshes every 10 seconds with a countdown. Its first actual Telegram sender receives regular membership automatically, without an approval decision. Each token binds to one account in the workspace's locked inbox transaction. Concurrent claims and duplicate events cannot enroll another account; replay after revocation cannot restore access. Existing active roles are preserved, inactive administrators rejoin as members, and revoked owners still require recovery. Bot identity, tenant, private-chat sender, signature and expiry checks remain application boundaries.

Copied links remain valid for 24 hours independently of the display refresh. Only hashed claims are persisted; expired claims and workspace purge remove them. Redemption records membership/audit and default private selection, resolves a prior pending request, and queues the fixed welcome reply without model/tool execution. It also works before activation. Ordinary requests and old workspace-ID links still require manual approval.

The read-only link card retains its monospace URL and one-second copied icon, keeps the URL visible during refresh, and pauses with an error toast and Retry on failure. Membership changes discovered by refresh update the member table.

Verification: bun run check, bun run typecheck, bun test against isolated PostgreSQL (751 passing tests), and bun run build; eight focused Playwright checks cover countdown/rotation, copying, member refresh, loading, pending, error/retry, navigation cleanup and existing manual decisions. Rendered states were inspected at 1280px and 390px. No live Telegram calls were used.

GitHub member choices and custom usernames (2026-10-14)

Member lookup now returns the personal installation owner directly or organization members from GET /orgs/{org}/members. It no longer inspects repositories or their collaborators. Personal lookup requests no installation access token. Organization lookup retains Members-read permission checks, pagination, stable-ID deduplication and the cancellable 60-second directory budget.

New/Edit member dialogs also offer Enter a username, usable during directory loading, empty results or lookup failures. Saving resolves the username through GitHub's public user API and stores its canonical login and numeric ID as an unverified association. Invalid/unavailable users and organization names retain the draft. Duplicate identities, stale membership/connection revisions, changed workspace authority and replacing verified identities are rejected. Saved custom accounts can be reopened, reselected and cleared; assignments create no repository grants.

Biome, strict TypeScript, the production build and 751 deterministic tests with disposable PostgreSQL passed. All 22 member browser scenarios passed, including custom input, failed lookup/save recovery, draft retention, Cancel, verified-account protection and pending-save locks. Rendered New/Edit and saved View dialogs were inspected at 1280px and 390px, including loading, empty, pending and error states. GitHub calls used fake transports; live GitHub acceptance remains pending. No new dependency or database migration is needed.

VPS deployment of skills and provider metadata (2026-10-14)

The earlier VPS rollout deployed package v0.1.36 from clean commit 2dd613aae9fe14abf07fcbc15abfc4d87e940571, deployed through manual archive bundle 1791958130084-55435898. The bundle retains the checksum-verified source archive, image metadata and local verification report. This rollout includes GitHub skill imports, read-only presets, typed provider metadata, completed-setup navigation and the saved-card action layout changes documented below.

Biome, strict TypeScript, the production build, 737 deterministic tests with disposable PostgreSQL and all 239 browser scenarios passed. All three release images built locally. The exact application's Node 24 runtime contract, isolated Compose configuration and service health, migration rerun, backup and restore rehearsal passed. Disposable test containers and volumes were removed.

The protected cutover waited for durable coding checkpoints and verified the mode-600 backup /opt/repodesk/backups/pre-release-1791958130084-55435898.dump. All 19 existing migrations remain applied; this release adds none. App, worker, Codex runner and PostgreSQL are healthy. Public health, readiness, the admin page and its scripts and stylesheet returned HTTP 200. Running app, worker and supervisor image IDs, and the runner's configured task image, match the imported release images.

Runtime configuration, the encryption key and runner token passed preservation checks. Database and runner-state volumes, PostgreSQL and all three Actions runner container identities were preserved; the host updater remains active. The source archive transfer resumed after an SSH disconnect while production remained healthy. This verifies the manual archive rollout, without completing the published-release local-image updater gate or live GitHub/provider/Telegram acceptance checks.

Preset skills use enable/disable only (2026-10-14)

Team recap and Follow-up draft are read-only presets. Their cards show a Preset badge, saved description/instruction preview and one enable/disable action, with no editor, archive, publishing, policy test, revision counts or rollback controls. Catalog pages containing only presets omit the sample policy form. Custom, conversation-created and GitHub-imported skills retain their editors and version management.

The API rejects preset edits, tests, publishing, rollback and archive with preset_skill_read_only. Admin authorization and optimistic revisions still guard enable/disable, and disabling keeps existing run/workflow revocation behavior. Internal published snapshots and pins remain intact. New seeds are marked as presets; unchanged legacy seeds are recognized without rewriting their contents or tool grants. Previously customized or archived seeds remain custom, including when a customized draft is restored to the original text.

Validation passed Biome, strict TypeScript, production build, the 737-test full suite, and focused policy/API tests after the legacy classification check was strengthened. 46 browser scenarios passed across presets, existing skill search, GitHub imports, admin management and team workflows. Seven new preset scenarios cover toggle-only controls, read-only previews, pending locks, errors and retained custom editing. Desktop (1280px) and mobile (375px), light/dark presets and pending/error states were inspected. The initial full run had three unrelated budget/model test timeouts; the isolated 51-test suite and full rerun passed. No dependency or SQL migration was added. See preset API examples.

GitHub skill imports and source updates (2026-10-14)

Agent skills imports accept GitHub repository, folder and SKILL.md links. The modal discovers nested skills, reports incompatible files and imports the selected files as disabled, unpublished drafts. Connected private repositories use the workspace operator's GitHub App with a repository-restricted Contents-read token; public reads carry no credentials. No dependencies or SQL migrations were added.

Saved cards show repository/ref/path and imported commit. Check for updates compares the tracked file, previews current and incoming instructions, and saves an update draft. Local edits need explicit replacement. Stale source commits, stale skill revisions, removed files and revoked connections preserve saved content. Published versions, enablement and scheduled version pins remain intact. Updates are initiated in the panel; no automatic syncing or publication is claimed. Imports retain the existing Markdown/frontmatter/tool limits and do not include supporting bundles or executable files.

Validation passed Biome, strict TypeScript, production build and 732 deterministic tests against disposable PostgreSQL. 22 browser scenarios passed across the GitHub imports, existing skills catalog and team-workflow suites, including eight new GitHub scenarios. Desktop (1280px) and mobile (375px) saved summaries, import discovery, pending locks, update review and failures were inspected; update review and saved updates also passed dark-mode inspection. GitHub API responses were deterministic fixtures; live GitHub App acceptance remains unverified. See skill design and runnable API examples.

Typed provider pricing and context discovery (2026-10-14)

Provider New/Edit dialogs now include a type: automatic detection, OpenRouter, self-hosted New API, OpenAI, DeepSeek, Mistral, Groq, xAI or another compatible API. The standalone model metadata module fetches model IDs and available API context/output/token-price data on provider saves. OpenRouter token prices are normalized to USD per million tokens. New API adds anonymous public pricing discovery, preserves proxy prefixes and requires the key's configured billing group before applying token/group/cache ratios. Price failures preserve model discovery; per-call and expression billing remain unknown.

Official APIs without price/capacity data use provider-specific bundled Pi catalogs, clearly labeled separately from API data. Current DeepSeek Flash/Pro aliases retain both peak/off-peak prices and separate cached/uncached input rates from an official reference checked 2026-10-14. Peak is the default conservative estimate; advanced chat settings can explicitly select Off-peak estimates. This choice is not automatic clock/holiday billing. Manual input/output/cache-read/cache-write overrides remain independently optional and support the existing currency conversion contract.

Saved summaries and modal editors show automatic information separately from overrides, with sources and units. Tier comparisons expand on demand; missing metadata is explained without inventing free prices or context caps. A blank context override still defers enforcement to the provider. Chat runs pin fetched metadata, the selected price tier and overrides; provider updates never revalue history. Reservations cover the highest input/cache rate. Ownership, encrypted keys, version checks, bounded responses and redirect denial remain enforced. Model lists still do not prove inference protocol support.

Validation passed Biome, strict TypeScript, production build, the Node runtime contract and 732 deterministic tests against disposable PostgreSQL. All 20 model-provider browser scenarios passed, covering type/group selection, cached price overrides, tier persistence, cancellation, pending locks, failed saves, loading/empty states and the existing setup/currency/service choices. Saved summaries, editors and pending/error states were inspected at 1280px and 390px in both themes. No dependency or database migration was added. No deployment or live provider inference, account connection or Telegram send was performed.

Completed setup navigation (2026-10-14)

Deployment navigation hides Setup when the authenticated progress API reports an active deployment. Inactive deployments retain the resumable link. The entry is absent until the activation state is known; navigation rechecks it and ignores responses from earlier routes or sessions. Deployment pause does not reopen setup. New workspace and direct setup routes remain available. Ordinary members do not request operator-only setup progress.

Validation passed Biome, strict TypeScript, the production build and 721 deterministic tests against disposable PostgreSQL. 29 browser scenarios passed, including 11 focused setup-navigation checks and existing sidebar, sign-in and session-expiry checks. Desktop (1280px) and mobile (390px), both themes, complete/incomplete, pending and failed status views were inspected. No dependency, migration or deployment was added.

Model-price currency selection (2026-10-13)

Bot chat model price overrides in setup and Model settings now support USD, CNY, TWD, HKD, EUR, GBP, JPY and SGD. Existing prices default to USD. Non-USD selections require an explicit positive usdRate (USD per currency unit), retained until edited. The runtime converts only entered overrides, including input overrides used for cache pricing; catalog fallback costs already use USD. Budgets and usage totals remain USD. Original prices, currency and rate are pinned in each run, so later edits do not revalue history. Prices remain independently optional, with unknown custom prices distinguished from zero. No automatic exchange-rate lookup is used.

The shared dropdown and modal editor retain entered prices when currency changes, clear the previous rate, and show currency codes and the saved rate in summaries. Cancel discards drafts; failed saves retain them; pending saves lock dismissal and duplicate submission. Model save errors scroll into view with Save and Cancel on mobile.

Validation of the isolated currency change passed Biome, strict TypeScript, the production build and 704 deterministic tests against disposable PostgreSQL. All 16 model-settings browser scenarios passed, including non-USD setup, reload persistence, validation, cancellation, save failure and pending locks. Saved summaries/editors were inspected at 1280px and 390px in both themes; loading, empty, pending and error states were also inspected at both widths. No dependency or migration was added, and no deployment or live provider/Telegram call was performed.

VPS deployment of GitHub plugin permissions (2026-10-13)

The recorded VPS serves package v0.1.36 from manual archive bundle 1791874171029-72300379. Its immutable source snapshot is bd6c3dc446f33b72e3bd68793b31769de0617a4f, based on 68630a5 plus the GitHub permission changes below. The bundle retains its source archive and image checksum metadata; this snapshot does not identify a published GitHub release.

The app, worker, Codex runner and PostgreSQL are healthy, and https://dev.stonegate.me/readyz returned HTTP 200 after cutover. Running app, worker and supervisor image IDs match the verified snapshot. Migrations 018 and 019 are applied; no workspace retains legacy Codex or Review Bot repository arrays. The deployment waited for durable coding checkpoints and verified the mode-600 backup /opt/repodesk/backups/pre-release-1791874171029-72300379.dump before migration. Existing configuration, secrets, volumes, updater and all three Actions runners remain in place.

All three release images built locally. The Node runtime contract, isolated Compose app/worker health and migration rerun, offline Open Code Review preparation and deterministic Docker Codex checkpoint/authentication/publication smoke checks passed. Local fixture containers and volumes were removed. These checks do not complete the live Telegram/model pilot or the published-release local-image updater gate.

GitHub permissions for Codex and Review Bot (2026-10-12)

Codex and Review Bot no longer have repository lists, Add/Edit/Remove actions, maintainer grants or automatic review owners. Their saved configuration summaries open shared modal editors for plugin-wide workflow settings. GitHub owns repository permissions; Codex uses the repository's GitHub default branch. Verified write access is required for coding. Tagged fixes use the PR’s actual base/head instead of a separate configured branch. Tagged reviews/answers require verified read access; fixes require write access. Automatic reviews use the enabled operator-owned App policy. Workspace membership, tenant connection scope, request approvals and scoped App tokens remain enforced before execution/publication. Task initiators can still stop their own work after losing GitHub access.

Migration 019_github_plugin_permissions.sql removes current plugin repository grants, advances affected revisions and preserves historical task/approval evidence. Mixed legacy execution choices become Reviewed; publication and review options remain enabled only when all previous targets agreed. Fixes also require the resulting Direct policy. Conflicting Review Bot destinations retain the lowest workspace ID and disable the others for operator review. Later connection overlaps route once to the same stable workspace choice. Tasks in other workspaces reauthorize before further execution or publication. Reapplying the migration makes no further change. Older clients cannot restore repository grants on save.

Validation passed Biome, strict TypeScript, production build, the host Node runtime contract and 707 deterministic tests against disposable PostgreSQL. 82 browser scenarios passed across the affected plugins, authentication/loading/task/card flows and the existing admin journey. A final 44-scenario pass verified the Codex feedback placement and device/authentication flows; two focused Review Bot page-error recovery checks also passed. Desktop (1280px) and mobile (390px), both themes, saved summaries, editors, loading, empty, pending, errors and conflict recovery were inspected. Save failures retain drafts and keep feedback, Save and Cancel visible together. No dependency was added. No deployment, live account connection, GitHub/Telegram send or model call was performed; existing live release gates remain.

Optional assistant limits and model settings (2026-10-12)

Model calls per run has been removed from Team summaries, editors, workspace settings and ordinary assistant execution. Follow-up classification and conversation compaction no longer consume a workspace call quota; each internal tool-free pass still has its own one-call bound. Migration 018_remove_model_call_limit.sql removes current saved overrides and advances affected workspace versions once. Older clients' maxTurns is discarded on save. Historical run snapshots remain audit evidence and do not enforce call quotas.

Run and monthly USD budgets default to None (null), meaning unlimited. Clearing a field restores None; entered zero permits only zero-cost requests. Existing explicit budgets remain configured. Atomic reservations still enforce configured caps, including current and pinned run budgets, concurrent requests and unknown charges. Summary/editor, Usage and Telegram accounting show unlimited values clearly.

Context window, maximum output and both token prices are independently optional. A blank context window leaves enforcement to the provider for known and custom models; there is no invented app context-window cap. Catalog context metadata remains available for conversation-memory planning, while the provider-context choice is pinned separately on runs. Explicit context/output overrides retain local validation. Output allowance uses configured/catalog values or the documented automatic allowance for unknown models. Missing custom prices remain unknown, retain token usage and are excluded from dollar totals. Finite dollar budgets require known prices and reconciliation of earlier unpriced charges. Verified charge reconciliation retains authorization and replay protections.

Team save errors now appear above the primary action and Cancel, keeping both buttons together on mobile. Version conflicts preserve the reload action and draft. Local verification passed lint, strict TypeScript, production build, 701 deterministic tests against disposable PostgreSQL and 19 focused browser scenarios on an isolated port and fixed UI build. Desktop (1280px) and mobile (390px) summaries, editors, loading, empty, pending, error, clearing, zero and conflict recovery states were checked. No application deployment or live provider/Telegram calls were made by this change.

Automatic activation after model setup (2026-10-12)

Model settings now contains provider and chat-model summaries with their shared modal editors. The separate Activate the bot card, activation action and duplicate inline access form are removed. Members & access retains enrollment, revocation and access requests; Operations retains optional Telegram account verification. Setup and token-replacement guidance describe automatic activation.

A successful workspace chat-model save attempts activation on the server. Saved bot credentials, a ready receiver and an enabled published skill in a workspace with a valid operator-owned model selection remain prerequisites. If Telegram or skills complete later, receiver/worker recovery activates automatically, including after restart. Legacy model credential saves also reconcile readiness. Activation locks and rechecks configuration, commits its audit once, preserves deployment pause and accepts empty membership without granting anyone access. Failed writes leave activation inactive and retain the saved model for recovery. The manual API remains compatible and validates receiver readiness before returning idempotently. Webhook registration remains explicit in Site domain; no model generation call is used as a setup test.

Validation: lint, strict TypeScript, production build and all 701 deterministic tests across 79 files passed with disposable PostgreSQL. 27 admin/provider browser scenarios passed in webhook mode and five focused setup/loading/save scenarios passed in polling mode. A final 13 focused model/setup browser scenarios also passed before commit. Desktop (1280px) and mobile (390px) saved, empty, editor, loading, pending and error screenshots were inspected. This change adds no dependency or migration and performs no deployment, live account connection, Telegram send or model call.

Three isolated Actions runners (2026-10-10)

The Actions Compose deployment now provides runner, runner-2 and runner-3. Each service has separate registration/work and Docker-data volumes, so up to three jobs can execute concurrently without sharing Docker daemons, CI database ports or disposable stacks. The original service and volume names are preserved. Each runner has a two-CPU ceiling and 2 GiB memory limit, for a combined 6-GiB memory ceiling on the recorded four-CPU, 8-GiB VPS. CPU capacity remains shared with production; the browser suite still executes on one Playwright worker.

Validation: lint, strict TypeScript, build, 683 deterministic tests against disposable PostgreSQL, the Node runtime contract and shell/Compose checks passed. The runner image built locally and passed Docker-in-Docker resource-limit and cleanup smoke tests locally and on the VPS under the new 2-GiB limit. All three deployed daemons passed simultaneous HTTP checks on the same internal port and nested CPU/memory/PID-limit checks. See runner operations.

The recorded VPS now has repodesk-vps (original ID 21), repodesk-vps-2 (ID 22) and repodesk-vps-3 (ID 23) online with healthy Compose containers. The original registration and Docker volumes remain. Production readiness is HTTP 200. No application release or workflow dispatch was performed by this change; complete Verify timings with three runners remain unmeasured.

Operations summaries and modal controls (2026-10-10)

/admin/operations now shows labeled health values, deployment activity, workspace setup status and operator audit records without inline forms. New panel accounts, identity verification, workspace access recovery and redacted run inspection open focused dialogs. Account and verification results stay in separate readable summaries. Cancel discards drafts; failures retain input; pending actions show operation-specific progress and lock dismissal/duplicates.

Deployment pause/resume requires confirmation with the loaded revision. A conflict keeps the error in the dialog and requires reloading before confirming again, with the primary action and Cancel kept together. Initial loading keeps labels and skeleton values visible; failed loads show unavailable states and a retry toast. Workspace-dependent actions require loaded, non-removed choices. Background refresh retains content and the dialog trigger's focus.

Validation: lint, strict TypeScript, build, 683 deterministic tests against disposable PostgreSQL and 18 Operations browser scenarios passed. Desktop (1280px) and mobile (390px) summary/editor, loading, empty, pending and error screenshots were inspected, including light/dark layouts. No dependency, migration, deployment, live account connection or external message was added.

In-card log pagination and refresh menu (2026-10-10)

Runtime logs now keep the shared pager inside the entries card, below the table and outside its horizontal scroll area. Blue chevrons surround a regular 14px current range; the separate Page label is removed. Cursor history, filter/search resets, loading locks and empty/error recovery remain supported. The API still has no total count, so the reference's total is not invented.

The five-second checkbox is replaced by an Auto-refresh menu with 10-second, 30-second, one-minute (default), five-minute and Off choices. A left checkmark identifies the selection; Refresh now sits below a divider and reloads the current page. Latest logs still returns to the newest page. Changing intervals replaces the timer without fetching immediately; pending requests and older pages pause automatic refresh. Refresh now disables during requests. The shared popup supports a bounded wider menu and upward placement when needed on mobile; its existing dropdown defaults remain unchanged. Arrow/Home/End navigation, Escape, Tab and outside dismissal preserve focus and touch targets.

Validation: lint, strict TypeScript, production build and all 683 deterministic tests passed with disposable PostgreSQL and host proxies disabled. 39 browser scenarios across Logs, Runs, dropdowns and admin passed, including a corrected timer-test selector; the focused dropdown/log run passed all 12 scenarios. Populated and open-menu screenshots were inspected in both themes at 1280px and 390px, plus loading/error/empty states at both widths. No dependency or API change was added; no deployment was performed.

Automatic Docker storage retention (2026-10-10)

Added an optional daily host cleanup timer and a dry-run-first release image policy. It retains the current release, two complete app/job/supervisor image sets, fresh builds and every existing container reference. Image aliases share protection; changed/newly referenced images are rechecked. Host containers, volumes, credentials, backups and metadata are outside the deletion policy. Unused host build cache has a 4GB retention target. Cleanup and verified host builds share a storage lock; direct cutovers retain their deployment lock.

The Actions runner image enables Docker-driver cache GC with a 2GB target and a bounded synchronous post-job hook. Only a daemon labeled repodesk.ci=true can remove old unused CI images, unused build cache and unreferenced disposable CI volumes. The outer runner/registration and production state stay intact. See retention and activation.

Verification: lint, strict TypeScript, build, 683 deterministic PostgreSQL tests, Docker runner/application builds, Node runtime contract, Compose configuration, shell syntax and systemd units passed. The real isolated runner smoke removed an unreferenced test volume and retained a stopped container's volume/image. A read-only VPS preview found 50 eligible old release tags. This initial validation did not delete VPS resources or restart services.

The operator then approved activation and immediate cleanup. Installed matching host scripts using the VPS's existing Node binary and verified their hashes against the checked-in working files. The daily timer is enabled and active; host cleanup removed the 50 eligible tags and completed with exit status zero. The updated runner image passed its real isolated VPS smoke before replacement while GitHub reported it idle. Runner ID 21 remains online with its existing registration and both outer volumes. Its labeled Docker 29.9.0 daemon and configured post-job hook passed immediate live CI cleanup. The filesystem now has 60GB free, 39% used, compared with 7.9GB free before this rollout. Core container identities, volume names and the current-release marker are unchanged; all containers are healthy and public readiness is HTTP 200. A future naturally completed GitHub job remains the live event check for automatic hook invocation.

An unexpected VPS reboot occurred during CI image export. Services recovered after boot, and the new image subsequently passed smoke and installation. The available journals do not establish its cause; no reboot command was issued by this workflow. See the reboot incident.

In-app update disk-space recovery (2026-10-10)

The first v0.1.34 host update built and smoke-tested its images but failed while creating an image archive on a full VPS disk. It never entered cutover, so the app/worker stayed healthy on v0.1.33. Capacity recovery removed the incomplete archive, unused old build cache and obsolete unreferenced RepoDesk release image tags while retaining current/recent/requested images and all container references. The operator's signed-in Retry update action completed a fresh verified attempt, 065610d4-768a-4f7f-8517-46b7b0a2bf12, installing release bundle 1791614204914-36610498. The reloaded browser shows v0.1.34 and the new member GitHub View action. App, worker, PostgreSQL and Codex runner are healthy; public readiness returns HTTP 200. The updater is active, the backup is mode 600, and 7.9GB remains free. See the incident report.

The local correction deploys same-daemon images by recorded immutable IDs, rejects retagging before writers stop, checks a minimum 1GiB reserve before checkout/build/cutover, and reports disk-space failures explicitly in the modal. Manual archive deployments keep checksum/import verification. This correction requires a future published release; the v0.1.34 recovery uses its original archive-based updater.

Validation: lint, strict TypeScript, build, 673 deterministic tests, five focused update browser scenarios, Docker app build/Node runtime smoke, Compose configuration and host-script syntax passed. Desktop/mobile error and retry states were inspected. Tests cover local-image cutover, changed app/task image IDs, low disk space, duplicate/interrupted jobs and existing backup, migration, checkpoint and readiness boundaries.

Member GitHub details behind View (2026-10-10)

The Members & access GitHub column now contains only a View button for every member. Its shared read-only modal shows the account/profile link, connection or verification status, sync time and complete saved repository access list. Unlinked accounts show Not linked; verified accounts with no repositories show an explicit empty state. Shared repository cards preserve permission labels, wrapping names and external links. X/Escape closes the dialog and restores focus to View. Opening it makes no additional GitHub request or write. Add/Edit retains its separate account selector. Member-load failures now use the shared compact retry toast.

Validation: lint, strict TypeScript, build and all 669 Bun tests passed against disposable PostgreSQL. The combined member, repository-card and admin browser run passed 34 scenarios; the final member suite passed 13 scenarios, including load failure/retry, unlinked/pending accounts, zero/one/five/six repositories, the complete 60-repository list, permission labels, focus and editing. Desktop/mobile display and dialog screenshots were inspected in both themes, plus loading, empty, pending and error states at 1280px and 390px. This replaces the earlier five-entry member preview. No dependency, migration or deployment was added.

Shared runtime log pagination (2026-10-10)

Runtime logs now reuse the Runs pagination component below the table, with previous/next chevrons, a current range/page summary and 12px row/pagination text. Cursor history supports returning through older pages without changing the API or inventing a total count. Filters, search and Latest logs reset history; automatic updates pause on older pages and resume on return to the latest page. Loading disables both navigation actions, failed pages retain previous-page recovery, and empty pages keep the pager visible. Log request errors use the shared retry toast; mobile admin toasts sit below the header to avoid covering pagination.

Deterministic browser coverage includes desktop/mobile cursor navigation, filter/search resets, loading/error/empty recovery, keyboard pagination, 44px action targets and paused/resumed polling. No dependency or API change was added.

Validation: bun run check, strict TypeScript, production build and all 669 deterministic tests passed with disposable PostgreSQL. All 50 browser scenarios across Logs, Runs, admin and loading/recovery passed with isolated server/artifact settings after a shared-output run lost a Playwright trace during cleanup. Populated light/dark and loading/error/empty screenshots were inspected at 1280px and 390px; error toasts leave the mobile pager accessible. No deployment was performed.

Retryable GitHub verification for coding tasks (2026-10-10)

Temporary GitHub permission-sync failures no longer permanently cancel Reviewed or Direct coding tasks. Unavailable/expired verification blocks new execution, auth resume, result processing and publication while retaining the task attempt. Already-running isolated work can finish under its original grant; successful permission sync resumes progression without repeating start or publication. Rate limits are distinct from permission denials, respect server cooldowns and increase sync backoff after repeated failures. Last confirmed permission data remains comparison-only while unavailable and keeps its original timestamp.

Confirmed permission loss atomically fences pending durable tasks for the actor and tasks containing that actor's inputs. Stop, membership and configuration revocation remain effective during recovery. Completed review-state results and confirmed PRs retain their saved status; follow-ups still reauthorize access. The panel and Telegram give specific verification/denial guidance. A safe github_user_sync_failed runtime event records the failure category without raw upstream diagnostics. Old cancelled records are not automatically restarted. See the incident evidence and recovery journey.

Validation: Biome, strict TypeScript, production build, Node runtime contract and 669 deterministic tests passed with disposable PostgreSQL and host proxies disabled. 10 coding-task browser scenarios passed at 1280px and 390px; desktop/mobile permission recovery and denial displays were visually inspected. Tests cover unchanged-attempt recovery, duplicate polls, publication blocking, undispatched reservation cleanup, rate-limit cooldown/backoff, explicit Stop, confirmed loss, contributor boundaries and preserved completed PR status. No dependency or migration was added. No deployment, live model call, GitHub mutation or Telegram send was performed.

PR #9 conflict resolution (2026-10-10)

Reconciled PR #9 (a325040) with main (13d7932), retaining the mobile navigation sheet and responsive version placement alongside main's compact notifications and below-list run pagination. Both implementation-history entries are preserved. Toast layout coverage now checks clearance above the desktop version footer and below the mobile version header; the UI documentation records that responsive distinction.

Validation: lint, both strict TypeScript checks, 661 deterministic tests with disposable PostgreSQL, production build and Node runtime checks passed through bash scripts/verify-skills.sh quality. All 158 browser tests passed across the three release-N-browser phases (58/54/46). Two initial browser processes were terminated before completion; the complete rerun in terminal sessions passed without skipped tests. Desktop/mobile screenshots were inspected for navigation, version placement, loading/empty states, toast errors, sign-out pending/error states and update dialogs. Physical mobile browsers remain unverified. No remote publication, deployment or .github/ changes were made.

Compact shared notifications (2026-10-10)

Shared admin toasts now use a neutral surface, subtle border/shadow and small colored status icon. Recovery actions show Retry inline beside the message, with descriptive accessible names; retry and close controls keep 44px touch targets. Network failures read “Couldn’t connect to RepoDesk.” Long messages wrap on narrow screens, and notifications leave clearance above the version/update control. Error persistence, manual dismissal and confirmation auto-dismiss/hover/focus behavior are retained. Dialog errors remain inside their editors.

Validation: lint, strict TypeScript, production build and 661 deterministic tests passed with disposable PostgreSQL. 45 focused browser scenarios passed across loading/recovery, Review Bot, provider/configuration editors, site settings, connected services, session expiry and shared save confirmations. Network-error coverage includes keyboard retry/dismissal, persistent errors, long messages, pending retry, recovered saved summaries and retained values after refresh failure at 1280px, 390px and 320px in both themes. Desktop/mobile screenshots were reviewed. No dependency or migration was added. No deployment was performed.

Mobile admin navigation and version placement (2026-10-09)

At widths up to 800px, the admin shell shows a compact persistent header and opens navigation in a scrollable left-side sheet. It reuses the shared modal's focus boundary, Escape handling and plain X; backdrop taps, navigation and workspace selection dismiss it. Desktop keeps its persistent sidebar. Resizing back to mobile starts with the sheet closed. Workspace permissions, account controls and update authorization remain enforced by the existing flows.

The mobile admin version and update indicator sit at the header's right edge, clear of page content. Mobile sign-in/setup use a normal-flow footer; desktop retains the bottom-right label. Both use the same version and update portal.

Deterministic browser coverage in sidebar.e2e.ts exercises 320px, 390px, 800px and desktop layouts, focus/dismissal, workspace selection, short-screen scrolling, empty workspaces, restricted navigation and page loading/error states. Account tests open the mobile sheet before exercising its controls. The sign-out fixture covers pending dismissal locks, visible failures and retry; error text wraps below the identity/actions without squeezing the username into a narrow column. Verification uses bash scripts/verify-skills.sh quality and all three bash scripts/verify-skills.sh release-N-browser phases (N = 1, 2, 3); these prepare dependencies and disposable PostgreSQL without model or GitHub credentials. Physical mobile browsers and deployment are outside these local checks.

GitHub release checks and host updates (2026-10-09)

Signed-in pages check the configured repository's latest stable release and show an arrow after the version label when a newer numeric version is available. The shared modal shows formatted changelog/update notes and an operator-only Update now action. Installation queues an atomic host request after checking release identity, notes, commit, session/origin/CSRF and updater heartbeat. Durable reservations and audit fence concurrent/replayed/uncertain requests. Only the separate host daemon builds images and invokes the backup/migration/readiness cutover. It checks successful Verify evidence for the pinned commit. The GitHub deployment workflow is removed; no Actions write credential is required. Workspace accounts can review notes. See installation and recovery.

2026-10-10: The version update indicator uses a compact 28px circle, 16px arrow and 5px status dot, with its 44px tap target and keyboard focus retained. Focused browser coverage checks both themes at desktop and mobile widths, opening notes from the transparent padding and with the keyboard. Desktop/mobile screenshots were reviewed in both themes; lint, strict TypeScript, build, 669 deterministic tests and 13 focused browser scenarios passed.

Direct VPS deployment was verified on 2026-10-09 for cfffdfc (v0.1.32) at the recorded deployment host. App, worker, Codex runner and the host updater are healthy; public /readyz returns ready and /admin serves the new version and indicator mount. The app's UID-1000 process verified queue read/write access, a fresh updater heartbeat and migration 017. The pre-migration backup is protected with mode 600. The updater uses a private Node v24.21.0 binary extracted from the verified app image, preserving the host's existing global runtime configuration. Installing a subsequent published release through the modal remains a live staging gate. Deterministic tests and browser fixtures do not perform deployment.

Validation: lint, strict typecheck, production build and 661 deterministic tests passed with disposable PostgreSQL and host proxies disabled. Nine browser scenarios passed for updates and session expiry. Desktop/mobile screenshots were reviewed for version placement, loading, populated/empty notes, pending locks, request errors and offline guidance. The Docker app image built and passed its Node runtime smoke; Compose configuration, host-script syntax and the systemd unit validated. Host tests cover pinned builds, failed verification, interrupted cutover fencing and a real Node daemon heartbeat without network/Docker work.

Explicit application test discovery in Verify, 2026-10-09

Quality now runs bun test ./tests. Bare bun test recursively discovered the optional Code Truth service tests after its dependency installation was removed from Verify. In the first simplified run, two service test files could not load tar and another service test timed out. The earlier local check retained the service's installed dependencies and missed this clean-checkout failure. Browser and application container jobs passed in that run; the final gate correctly rejected the failed quality job.

Validation used an isolated checkout with only the root frozen-lockfile install and no Code Truth dependencies. The missing-package error was reproduced with a service test. The scoped CI command then passed all 616 application tests in 70 files, with no service tests discovered. Biome, strict TypeScript, production build and the Node runtime contract also passed there. Application behavior, dependencies and test implementations are unchanged.

Simplified Verify workflow, 2026-10-09

Verify now runs quality/database checks, the complete browser suite in one job, and application container checks. Code Truth's separate service installation, typechecking and image smoke, plus the Codex image/lifecycle job, were removed. The browser matrix was removed; its single job has a 30-minute budget. The existing required checks gate waits for all three remaining verification jobs and rejects failed, skipped or cancelled results. Release deployment separately retains its Codex image builds and lifecycle smoke before transfer.

Validation: workflow configuration/dependencies, Biome, strict TypeScript, production build, all 645 deterministic tests and all 137 browser tests passed locally. The application image built and passed its Node runtime contract, isolated Compose migration/readiness/setup checks, and backup/restore rehearsal. No application or test implementation changed.

Independent model-provider browser fixtures, 2026-10-09

The model-provider browser fixture now marks its disposable deployment claimed after inserting its administrator. Previously, a fresh CI shard displayed the first-run Create administrator form while the tests waited for Sign in; a full local browser run hid this dependency on earlier setup tests. The original Browser (3/4) failure was reproduced with one test against fresh PostgreSQL. All six affected tests now pass as a standalone file at 1280px and 390px. Application behavior, dependencies and deployment configuration are unchanged.

Validation: Biome, strict TypeScript, production build, Node runtime contract, all 645 deterministic tests and all 40 tests in browser shard 3/4 passed against disposable PostgreSQL. No production rollout or live external action was performed.

No fixed chat tool-call quota, 2026-10-09

Normal Pi workspace requests no longer stop after eight tool calls. The worker omits the call-count cap, and the runtime treats an omitted cap as unlimited while still accounting for calls. Workspace model-turn/spending limits, timeout, cancellation, permissions and approval pauses remain enforced. Internal summarization/classification stays tool-free; deliberate internal/evaluation caps remain available and have a specific Telegram failure message. No workspace setting, dependency or migration was added.

Regression tests complete twelve reads in one batch and across several model turns, recover from a failed lookup, and stop on revocation after the ninth read. The real worker fixture finishes twelve durable tool outcomes in two model turns and queues one answer; duplicate execution does not replay it. Existing tests cover cancellation, budgets, model-turn exhaustion, approvals and tool-free passes.

Validation in an isolated checkout of this fix: Biome, strict TypeScript, production build, Node runtime contract and 628 deterministic tests passed against disposable PostgreSQL. The application image built and passed isolated migration/health/setup smoke; base Compose validation passed. Concurrent GitHub action changes in the shared checkout were preserved and excluded from this verification. No production rollout, live model call, GitHub write or Telegram send was performed.

Default GitHub tools/skill and approved PR actions, 2026-10-09

Connected workspaces now receive app-owned repository lookup, issue/PR metadata, issue proposals and merge/close proposals by default, alongside the bundled repodesk-github skill. Existing workspaces need no plugin manifest or skill-catalog migration, and custom skill records remain intact. Registered tools describe the current build even when retained conversations contain older replies about missing capabilities. Optional Code Truth and coding execution retain their own configuration requirements.

PR actions require an active owner/admin or configured coding maintainer, current write/admin access for linked GitHub accounts, a selected repository and a separate requester approval for each PR. Telegram previews identify the exact action, PR, target branch, and merge method/head commit. The worker rereads the PR, rechecks authority and deployment state, commits a reservation, then attempts one scoped mutation. Changed/draft/closed PRs fail safely; merges include the reviewed head SHA. Concurrent workers, duplicate approvals and restarts cannot repeat a reserved write. Uncertain outcomes require GitHub inspection. Safe tool/result guidance explains permission updates and PR-state failures without exposing upstream diagnostics.

The App manifest already requested Contents and Pull requests write permissions; older installations still require GitHub permission-update acceptance. Running app and worker processes must be updated through the release flow to expose the new tool. This adds no dependency or database migration. No deployment, live GitHub write, Telegram send or model call was performed. Native Telegram desktop/mobile rendering, real installation-grant changes and live PR outcomes remain staging gates. See setup and limits and the runnable journey.

Validation: bun run check, strict typechecking, production build and all 642 deterministic tests passed against disposable PostgreSQL. The 17 PR-action checks cover default-agent tool/skill exposure for existing workspaces, unchanged custom catalogs, disconnect, exact approval, actor/tenant/privacy boundaries, permission revocation, changed PRs, scoped token grants, duplicate sends and safe unknown outcomes. The Node runtime contract passed on the host and in the built application Docker image, confirming the bundled skill and core tools are packaged.

Actionable coding checkout failures, 2026-10-09

Coding preparation now distinguishes a missing configured base branch from other clone failures. The task container records only an allowed failure code; the supervisor preserves it through restart. Telegram outcomes, requested status and the saved admin task display name the authorized repository and missing branch, explain that Codex did not start, and point to Plugins → Codex → Repositories before a new request. Missing PR branches do not claim that the base branch is absent; uncertain publication retains its existing reconciliation behavior. Old generic failures cannot reconstruct discarded Git diagnostics.

Saved task rows use labeled details on mobile so the longer explanation remains fully readable. Existing task actions and Stop confirmation remain in place. The UI rules record the user's request for specific failure causes and recovery. No dependency, database migration or deployment configuration was added.

Validation: Biome, strict TypeScript, build and 625 deterministic tests passed against disposable PostgreSQL, plus the Node runtime contract and 8 browser checks. Saved failures were visually inspected at 1280px and 390px. Application, task and supervisor images built; application migration/health/setup smoke, the full fake-provider Docker lifecycle (including a real local missing-branch clone and restart), and Docker-overlay Compose validation passed. Tests cover one scoped failure notice, duplicate polls, unknown/missing markers, tenant isolation, both Reviewed and Direct workflows, and no implementation/publication after checkout failure. No live provider call, GitHub write, Telegram send or production rollout was performed.

Shared model providers and GitHub App icon, 2026-10-08

Setup now includes Workspace → Model → Telegram → GitHub. Operators save an OpenAI-compatible provider URL and encrypted key once; saving fetches its model list. Providers are reused across that operator's workspaces. Bot chat, Codex and Review Bot save independent model selections from the same catalog. Chat execution pins the selected provider and workspace model revision; Codex's task-scoped proxy binds the chosen endpoint/model and keeps provider keys sealed in the supervisor. Legacy configuration remains readable, and ChatGPT device-code authentication is still available. Provider discovery does not make a paid inference request.

Provider records use saved summaries and New/Edit dialogs; service editors contain provider/model selectors. Tests cover ownership, CSRF, encrypted storage, URL/key binding, failed discovery, stale saves, unavailable models, key rotation, sharing, independent selections, and task-proxy routing after restart. Workspace deletion removes its model selection; shared operator provider records remain configuration. No dependency or database migration was added.

The original GitHub creation flow offered the RepoDesk icon as a 200×200 PNG under 1 MiB and an owner-appropriate App settings link, then waited for an upload acknowledgement before authorization. That manual icon step was removed on 2026-10-18; see the current entry above.

Validation: Biome, strict TypeScript, production build and all 619 deterministic tests passed with disposable PostgreSQL. All 41 affected browser scenarios passed: admin (17), provider/model/icon (6), Codex device (8), and Review Bot (10). Desktop/mobile saved displays, editors, pending/error states and the icon handoff were visually inspected. App, job and supervisor images built; full fake-provider Docker lifecycle, offline OCR preparation and Docker-runner Compose checks passed. Tests used isolated loopback ports/artifacts and fake provider/GitHub transports. No live provider inference, GitHub logo upload, account connection or deployment was performed. Real provider protocol/model compatibility remains pilot acceptance.

Shared repository card styling, 2026-10-08

Repository summaries now reuse one card component and list styles in Members & access, Manage GitHub, Codex, Review Bot and Code Truth. Compact and detailed cards share their surface, border, corners, typography, badges and action layout. Long names wrap; managed cards give names the full heading width on mobile. External GitHub links use accessible labels and 44px targets. Existing saved details, New/Edit dialogs, the five-member-repository preview limit and GitHub's expandable connection list retain their behavior. The UI rules record this app-wide preference.

Biome, strict TypeScript, production build and 611 deterministic tests passed in an isolated checkout containing only these UI changes, against disposable PostgreSQL. 44 browser checks passed across repository styling, member previews, repository search/refresh, Review Bot editing/errors and loading/recovery. New checks compare card styling across all five views at 1280px and 390px in light and dark themes, verify long names, saved details, safe links and edit/cancel behavior. Rendered cards were visually inspected at both widths and in both themes. The complete working tree also passed 619 deterministic tests during implementation. No dependency, migration or deployment was added.

Compact member repository previews, 2026-10-08

Members & access now shows at most five repository permission entries in each member's GitHub column. Lists longer than five offer Show all with the total count, opening the complete saved list in the shared read-only modal. Long names wrap; the modal scrolls, traps focus, closes with its X or Escape and returns focus to the triggering button. Opening it performs no additional GitHub request or write. Member editing and verification-pending account displays retain their existing flows.

Validation: Biome and production build passed; all 619 deterministic tests passed against disposable PostgreSQL. All nine focused member browser checks passed, including zero/one/five/six-entry boundaries, a 60-repository list, permission labels, keyboard behavior, member editing and lookup failure recovery. Preview and modal rendering, including the final long repository name, were visually inspected at 1280px and 390px. Full typechecking remains blocked by an unrelated unresolved advanced reference in the concurrently edited web/model-providers.tsx. No dependency, migration or deployment was added.

Telegram PR controls, 2026-10-08

Reviewed and Direct coding task messages with a confirmed PR now use native Review/Merge URL buttons. Review opens changed files; Merge opens the PR on GitHub for its permission checks and confirmation. Outcomes, waiting questions, blockers, follow-up progress, cancellation receipts and requested status keep the PR actions. The existing progress message receives the same controls through its durable edit. Before publication, existing task controls remain applicable. Older Status taps return current PR actions without granting a review or merge operation in Telegram.

Deterministic tests verify sent and edited button payloads, the waiting-after-PR case, confirmed publication, follow-up work, old controls and revoked delivery. Biome, strict TypeScript, build and all 611 deterministic tests passed against disposable PostgreSQL. Live Telegram desktop/mobile rendering remains unverified; no deployment or live Telegram/GitHub action is part of this change.

Open Code Review primary framework, 2026-10-08

Review Bot now uses Open Code Review (OCR) delegation as its primary review framework. The task image pins the official CLI at 1.12.12. Credential-free setup prepares a commit-pinned file/rule plan, validates JSON compatibility and complete rule coverage, and stops on preparation failures before model execution. Codex performs read-only analysis using its existing workspace credentials. Review instructions require file coverage and explicit exclusion/skipping reasons; that reported inspection coverage remains model-generated. Answers and authorized fixes retain their existing workflow. No migration or new model account is required.

Biome, strict TypeScript, build and all 592 deterministic tests passed against disposable PostgreSQL. App/job Docker images built, migrations/API health and base plus Docker-runner Compose validation passed. The pinned OCR CLI's real offline preparation passed inside the job image, including exact refs, grouped rules, exclusions and failure behavior. The repeatable smoke script uses disposable Docker volumes and no model/GitHub credentials. Rebuild existing task images to use this integration. Live semantic review quality and intended-host Codex sandbox compatibility remain release gates. See setup and boundaries.

Overview coding-task count, 2026-10-08

The Overview Coding Tasks card now includes both Reviewed tasks in workspace state and Direct tasks in the workspace-scoped task table, using the same list query as the Codex page. Waiting, completed and cancelled tasks count once, independently of their execution attempts. The card opens the Codex task list instead of the Plugins catalog. Existing count visibility and tenant permissions remain in effect.

Regression coverage checks empty, Reviewed-only, Direct-only and mixed counts, other-workspace isolation and workspace-administrator access. Browser checks use the real local API with deterministic database fixtures at 1280px and 390px, verify the count and destination rows, and cover the existing pending, failed-load, retry and saved-value refresh states.

Validation: Biome, TypeScript and build passed; all 586 Bun tests passed against disposable PostgreSQL, and all seven focused browser checks passed. Rendered Overview counts and the matching Codex list were inspected on desktop and mobile.

GitHub App onboarding defaults, 2026-10-08

New GitHub App registrations now suggest repodesk in setup, the creation dialog and API requests that omit a name. Their manifests preselect pull_request, issue_comment and pull_request_review_comment; public HTTPS domains enable the operator-specific webhook automatically, including a configured admin site domain. Local, IP-address and non-HTTPS origins retain an inactive placeholder with the subscriptions selected. GitHub's generated secret is required at conversion and stored encrypted for Review Bot's existing operator-scoped fallback after restart. Existing App records and explicit secret replacements remain supported. Review Bot repository policies and verified member requirements remain unchanged.

Biome, strict TypeScript, build and 584 deterministic tests passed with disposable PostgreSQL and loopback proxy bypass. Added coverage checks name defaults, public and local origins, custom domains, encrypted secret retention, signed webhook intake, invalid conversions and operator isolation. All 17 admin browser scenarios passed, including setup name submission, failed-registration recovery and the creation dialog's default name and posted manifest. Desktop/mobile creation and saved App summaries were visually inspected. No dependency, migration, live GitHub/Telegram call, production configuration change or deployment was added.

Codex follow-up intent routing, 2026-10-08

Ordinary messages in a private or linked-group Topic now enter Pi even when a Codex task is working, waiting or ready for review. Pi chooses between answering with permitted read tools, continuing clear changes to the matching task/PR, and starting a separate coding goal. PR reviews and explanations no longer resume Codex automatically. Metadata alone is not presented as a code review.

Direct ingress remains for explicit task controls, replies to the confirmed notice for the current Codex question, and edits/duplicates of accepted task inputs. Unanchored answers receive Pi intent handling. Explicit Pi continuations record their task handoff so native Stop and completion notices follow the correct task. Actor, workspace, bot, chat/Topic, source and repository boundaries remain checked.

Biome, strict TypeScript, production build and the full deterministic suite passed with disposable PostgreSQL and loopback proxy bypass. Coverage includes private and group Topics, ordinary/old-result replies, duplicate events, question/media answers, task-input edits, original-wording continuation on the same PR, multiple tasks and cancellation. These checks establish routing and authorization, not live model intent accuracy. No dependency, migration, deployment or live external call was added. See routing contract and staging journey.

Codex PR readiness, 2026-10-08

Successful Reviewed and Direct coding publications now create PRs ready for review. Verified same-PR follow-ups mark existing drafts ready before confirming success. Readiness checks bind the PR to the configured repository, branch, base and verified commit. Lost acknowledgements are not replayed; read-only recovery requires an open, non-draft PR at that commit. Review Bot fixes retain their existing draft policy. Settings, approval previews, task status, Telegram notices and current guides use the updated publication contract.

Biome, strict TypeScript, build, 577 deterministic tests with disposable PostgreSQL, all 17 admin browser scenarios, Docker job/supervisor builds, full fake-provider Docker lifecycle smoke and Compose validation passed. Mobile settings were visually checked. Localhost was excluded from the host proxy for local service tests. No live GitHub, Telegram or model calls, or deployment, were performed.

Review Bot, 2026-10-07

Review Bot is implemented locally: selected-repository automatic PR reviews, verified GitHub mentions, read-only answers, checked same-PR fixes, immediate status, cancellation, question continuation and durable reconciliation. Webhook secrets/receipts use migration 016_review_bot.sql; task configuration and records remain tenant-scoped. Repeated events, source/permission changes, stale heads, branch rewinds and uncertain outcomes are covered deterministically.

Biome, strict TypeScript, build, 569 deterministic tests with disposable PostgreSQL and 97 browser scenarios passed. App/job Docker builds, migration, API health, fake-Git PR preparation, Node 24 Pi contract and Compose validation passed. No live GitHub, Telegram or model call was part of these checks. The development host's Codex command-sandbox probe failed at bubblewrap namespace/devpts setup; live review acceptance requires a compatible runner. See the linked setup and acceptance journeys.

This dated log records local implementation and verification, including the P0 bot/admin/worker foundation and bounded GitHub paths. Some changes were deployed to a local Podman stack, but the primary GitHub journey has not been validated end to end with live GitHub, model and Telegram operations or accepted by a pilot team. Each entry states its own test scope and remaining gates.

Implemented

Parallel Codex task execution (2026-10-07)

Replaced the single global active-task guard with CODEX_RUNNER_CONCURRENCY (default four, range 1–32), wired through Docker and Podman Compose. Capacity is reserved before container launch, restored from durable records after restart and shared across repositories/workspaces. Independent API-key tasks run Codex in parallel; attempts for the same continuous task remain serialized. User-input and authentication pauses still release capacity, while excess tasks use the existing queued retry paths.

Tasks sharing one workspace's managed ChatGPT cache serialize implementation, repair and auth resume to preserve refreshed credentials. Preparation, setup, checks and publication can overlap; separate workspace caches run independently. Queued repairs do not consume an attempt before they can launch. See the runner guide for configuration and per-task resource limits.

Validation: Biome, strict TypeScript, build and 517 deterministic tests across 62 files passed with disposable PostgreSQL. Six added runner tests cover capacity, duplicate starts, restart/cancellation, task and tenant isolation, credential-stream ordering, auth resume and repair scheduling. App, task and supervisor Docker images built; host/container Node runtime contracts and all three Compose configurations passed. Docker and rootless Podman smoke verified two simultaneous implementation containers, capacity rejection and supervisor restart, plus the existing task/auth/ publication lifecycle. External services were fake; no live account, model call, GitHub write, Telegram send or deployment was performed. No dependency or migration was added.

Repository reports, conversation skill drafts and work handoffs (2026-10-07)

Implemented the accepted R01–R03 plan. GitHub reads use repository-scoped Issues/Pull requests read tokens, bounded pages, canonical links, retrieval times and merge-time windows. Repository schedules pin source IDs and connection revision; private group reports require admin audience approval. Previous complete local calendar days include 23/25-hour DST transitions. Access changes block reads and queued delivery, including indirect evidence from older assistant reports.

An owned successful assistant run can produce a sanitized skill proposal. Requester approval creates a disabled admin-visible draft; existing admin publish/enable controls make it reusable. Unpublished drafts depend on retained sources; published procedures have an independent approved-instruction lifecycle. Handoffs batch-read durable task inputs/context, preserve questions/checkpoints/verification uncertainty, and include permitted discussion notes. Personal and group/topic boundaries remain distinct. Added tool context is covered by a compaction regression fix and test.

Validation: Biome, strict TypeScript, build and 502 deterministic tests across 61 files passed with disposable PostgreSQL. 19 browser scenarios passed, including two new desktop/mobile feature checks with visual inspection. The built host Node runtime contract passed. Added 30 unit/integration tests and two browser scenarios; these features add no dependency or SQL migration. The final pass includes concurrent membership/GitHub account edits without reverting them. External services were fake; no deployment, account connection, live Telegram send, model call or GitHub write was performed. Live report usefulness, teammate reuse and handoff quality remain pilot checks in the runnable examples.

Telegram progress and cancellation feedback (2026-10-06)

Implemented the friend beta feedback plan. Telegram uses readable task stages and actionable failures, with durable phase notices, coalescing and suppression of obsolete progress. Status/cancellation target replies or the current conversation, with authenticated selection for multiple tasks. Queued selectors and actions recheck current permissions. Cancellation separates stopping from confirmed termination; missing runner records and publication in flight retain uncertainty. Native Stop and original-request cancellation reach a linked coding task under existing permissions. Pi handoff responses and previews cannot claim queued coding work is complete. Ordinary cancellation deliveries are included in scoped admin run details. The Codex panel uses the same readable stages.

The 2026-10-07 follow-up adds inline Status/Cancel controls to coding notices and long private queue notices, removing repeated command reminders. Longer stages use short, natural follow-ups. Button taps bind to confirmed sent messages and recheck current task permissions; acknowledgements clear the native spinner silently. Local button delivery, routing, isolation, revocation and replay tests passed as part of the 526-test suite, alongside Biome, typecheck and build. New button behavior still needs live Telegram acceptance.

The later 2026-10-07 correction makes routine progress edit the original task message, including delayed and repeated repair stages. Necessary questions, actionable blockers, requested receipts/status and final outcomes remain new replies. Waiting/terminal states close the existing progress message. Durable edits retain scope and bot checks, serialize, and retry known-message edits safely; confirmed deleted messages can be replaced once. Unknown original sends are not replayed. Verification guidance requires complete bounded commands or a repository script and explains how to report an irreparable frozen-plan blocker. See the incident diagnosis. This correction is local; the diagnosed live task still has its old frozen plan.

Validation: Biome, strict TypeScript, build and 451 deterministic tests passed with disposable PostgreSQL; 25 affected browser scenarios passed with desktop and mobile inspection. Application/supervisor/job images built, Node runtime contract and Compose validation passed, and deterministic container smoke covered the runner lifecycle with fake external services. Fresh migration/API/worker smoke passed. No new dependency or migration was added. Live Telegram/model/GitHub beta acceptance and deployment remain separate steps; see the manual acceptance procedure.

Automatic Codex environment preparation and checks (2026-10-05)

Removed setup/check command configuration from both repository forms and the execution contract. Codex prepares the environment and selects a bounded verification plan from repository instructions. The runner captures the plan, replays it without model/GitHub credentials and checks patch integrity. Repairs retain the original commands; missing plans and failed checks block publication. Migration 013 removes stored overrides and invalidates affected old grants. Older client command fields are discarded rather than becoming hidden overrides.

Local validation before publication: Biome, TypeScript, build, 354 Bun tests, 23 browser tests, pinned CLI protocol, image builds, Node runtime contract, Compose and real Podman smoke with fake external services passed. Live model selection quality and the Telegram/GitHub pilot remain separate release gates. The user requested proceeding to release without additional verification.

Continuous Codex collaboration (2026-10-05)

Implemented the accepted plan locally. Pi relays original source messages; Codex interprets intent, makes technical decisions, asks product questions, implements and repairs failed checks. Repository direct policy is opt-in; existing settings retain reviewed approvals. The application persists tenant-scoped inputs, grants, fenced attempts and delivery events, checks authority again before execution/publication, and enforces cumulative budgets. Telegram replies and Topic follow-ups continue one task and PR. Waiting releases the runner; restarts reconstruct retained context. Ambiguous publication outcomes are reconciled without replaying writes, and expired/deleted sources trigger private-state erasure.

Validation: Biome, TypeScript, Bun build and 351 Bun tests passed with disposable PostgreSQL. All 19 browser tests passed, including reviewed defaults, direct policy persistence and desktop/mobile layout. The pinned Codex 0.155.1 protocol passed on the host and in the isolated job image with a local fake Responses provider. Application, supervisor and job images built; the isolated Node runtime contract and Compose configuration passed. Real rootless Podman smoke with fake external services covered provider/device credentials, questions, checkpoint reconstruction, bounded repair/exhaustion, fresh and same-PR publication, remote-head fencing, cancellation and erasure. No live model call, Telegram message, GitHub write or deployment was performed. The live pilot and semantic intent evaluation remain release gates; see the plan for their acceptance journey.

Release-driven VPS deployment (2026-10-05)

Added a stable GitHub Release workflow that reuses CI verification, builds the release event's immutable commit and transfers its image over verified SSH. The host script locks deployment, checks runtime secret access, stops writers, validates a protected pre-migration backup, runs migrations and verifies both container health and application readiness before recording success. Failed cutovers stop writers and require deliberate schema-aware recovery.

Validation: Biome, TypeScript, Bun build and runtime contract passed; all 317 Bun tests passed against disposable PostgreSQL, including seven deployment tests. Actionlint, Bash syntax and Compose validation passed. The application Docker image built and passed its isolated Node runtime contract. Disposable Docker stacks passed a fresh install and an upgrade to a distinct image ID; both app/worker changed images, database test data persisted, backups validated and host /readyz returned HTTP 200. Temporary containers and volumes were removed. No live VPS deployment, account connection, webhook registration or message sending was performed. See VPS setup for required host provisioning and GitHub secrets.

Local Codex runner reconciliation (2026-09-29)

The live Podman supervisor was healthy with no active coding containers. Two earlier implementation jobs had failed and their work volumes remained within the configured 24-hour retention period. The supervisor now removes a stale credential-free export container before retrying a patch export, so an interrupted read-only export can recover after a restart. Periodic timeout reconciliation now records coding_task_timeout for the task status. The running supervisor image was not replaced during this check.

Validation: Biome, TypeScript, Bun build, the 10 runner unit tests, coding integration tests, Compose configuration and a real Podman smoke with fake Codex and local Git passed. The full Bun suite reached 301 of 302 passes with a reproducible timeout in the unrelated native Telegram Stop test; the coding integration suite passed. The smoke resources were cleaned up.

Codex local-only execution (2026-09-29)

Removed the GitHub Actions coding dispatcher, repository workflow template and backend selector. New approvals run only through the local Podman supervisor. Operators set local check commands for each coding repository; old Actions configuration cannot authorize new tasks. Historical Actions tasks and links stay visible, but the worker does not dispatch, poll or cancel them. Queued and issue-created tasks are cancelled locally; tasks that may have started remotely become unknown and require inspection in GitHub.

Validation: Biome, TypeScript, Bun build, all 301 Bun tests against disposable PostgreSQL, Compose configuration, isolated app image build and Node runtime contract passed. The coding panel browser flow passed in the full browser run. That run stopped at a separate run-summary UI assertion after 15 passes; two later tests did not run. No app service was redeployed or live coding task started.

Overview run and workflow visibility (2026-09-28)

Overview count cards now open their corresponding member, assistant-run, scheduled-workflow and Codex-task views. Counts match the records in those views. Initially, unlinked deployment administrators could inspect only read-only run and workflow metadata. As of 2026-10-06, the workspace deployment administrator and eligible workspace owners/admins can read every retained run's request, result, transcript and delivery text, including other users' private runs. Deployment administrators need no linked Telegram identity for this view. Access remains workspace-scoped; Telegram execution and mutation checks remain in effect. Workflow content keeps its existing identity restrictions. The Codex task table also shows a recorded thread ID when available.

Validation for the 2026-10-06 run-message change: Biome, TypeScript and build passed; all 385 Bun tests passed against disposable PostgreSQL. The setup and run-message browser checks passed for linked and unlinked administrators, with mobile dialog inspection. Permission tests cover tenant boundaries, ordinary members, revoked workspace admins and unchanged private-run cancellation checks.

Validation: Biome, TypeScript and build passed; all 301 Bun tests passed against disposable PostgreSQL, and all 17 browser tests passed with fake external services. The live Clawearn task check failures are documented in the incident record. Future failed local Codex tasks now retain their thread ID and a safe failure stage when the job supplies one; the panel renders the stage in plain language. The local Podman app, worker and Codex supervisor were rebuilt and restarted with the verified images. The job image was prebuilt for future tasks. The protected, verified pre-cutover archive is backups/repodesk-pre-overview-20260928T104445Z.dump; prior images remain tagged rollback-pre-overview-20260928. Compose validation, the isolated Node runtime contract, app health and asset smoke, and real Podman success/failure smoke passed. All three services became healthy, /readyz returned ready, and Telegram polling reconnected. No live Telegram message or new paid Codex task was sent for this release check. Historical failed tasks remain failed with their original generic error code.

Setup flow and panel model settings (2026-09-27)

The authenticated wizard now follows Workspace → Telegram → GitHub App → Enter panel. It reuses guided App registration and repository connection, returning GitHub callbacks to the wizard when started there. GitHub access can be finished later from Plugins. Model credentials and explicit bot activation moved to the operator's Model settings page; entering the panel leaves the bot inactive.

Validation: Biome, TypeScript and build passed; 296 Bun tests passed with an isolated PostgreSQL test database, and all 16 browser tests passed. Browser tests used fake Telegram, GitHub and model transports. This change was not deployed or tested against live providers.

Group names and local deployment (2026-09-21)

Group access cards show the Telegram group title above its numeric ID, with an ID fallback for unknown titles. Linking and incoming group updates persist titles; Recheck bot visibility also fetches the current title for existing bindings. Inaccessible groups retain their last known title.

Validation: 236 deterministic tests with PostgreSQL, 16 browser checks, Biome, TypeScript and build passed. Deployed app and worker to local Podman using localhost/deepx-agent:group-names-20260921 (c8ef95b29cd6), also tagged local. Compose validation, migrations, isolated Node runtime and HTTP smoke checks passed. Both live containers passed health checks; /healthz and /readyz returned 200, Telegram polling connected, and the served admin bundle includes the new card UI. PostgreSQL and Code Truth containers were preserved. Database backup: backups/deepx-before-group-names-20260921.dump (archive verified). Rollback image: localhost/deepx-agent:before-group-names-20260921.

Core implementation

• Opt-in local Telegram long polling with durable per-bot offsets, single-receiver advisory locking, rate-limit backoff, cancellation and shared ingress/owner validation. Setup and readiness checks distinguish polling from the default webhook transport.

• Pi 0.85.1 with configurable OpenAI-compatible endpoints, custom model IDs and thinking levels, bounded sequential tools, fake-stream tests, cancellation, durable checkpoints and safe transcript recovery.

• PostgreSQL versioned migrations, tenant row locks, inbox deduplication, transactional outbox/pg-boss enqueue, fenced run leases and recovery after worker loss.

• One-use deployment claim, scrypt passwords, revocable sessions, CSRF/origin checks, encrypted write-only settings, resumable setup and verified Telegram identity links.

• Telegram webhook validation, command/reply routing, group-admin verification, explicit active membership policy, consented received-message collection and scoped sources.

• Manual requests/recaps, source-ID validation, usage reservations, cancellation, versioned instruction/workflow/skill snapshots and separate delivery outcomes.

• Canonically hashed approvals, daily/weekly recurrence with DST policy, unique occurrences under transaction locks, missed-run notices and workflow lifecycle controls.

• Approved corrections, personal/workspace/workflow instructions, forget/edit, skills catalog/import/drafts/settings/publication/enablement/rollback and deterministic tests.

• React Router admin screens for setup, settings, access, members, chats, workflows, skills, instructions, runs, usage, audit, privacy and operator controls.

• Retention/deletion tombstones, unknown-charge/delivery reconciliation, health/heartbeat, CI, pinned base images, local Docker checks and backup/restore automation.

Validation evidence

• 48 Bun tests: 26 local/unit tests plus 22 tests against real PostgreSQL 17.
• 3 Chromium browser tests: first-claim/resume, write-only credentials, concurrent save conflict, skill publish/test/enable, workflow pause/audit, sign-in and mobile layout.
• TypeScript strict checking, Biome and production frontend/backend build.
• Docker Compose migration/API/worker/PostgreSQL health; /readyz and admin deep links.
• Container restart preserved a claimed deployment/workspace; a custom-format backup restored successfully into a disposable database with three migrations and the fixture.
• Pi contract in the built Node v24.21.0 image: actual fake-provider tool flow, event ordering, bounded turns, cancellation and restored tool-result continuation.

The PostgreSQL tests found and fixed a canonicalization defect: JSONB's key ordering must not change approval hashes. They also cover claim/approval/dispatch races, rollback, independent connections, stale settings, revoked sessions, source isolation, charge reservation races, duplicate workers, rate handling, unknown sends, deletion, version pins and awaiting-approval worker release.

Run the exact commands in setup. Tests need no real Telegram/model secrets. Browser artifacts are generated into ignored test-results/; they are not release assets.

Deliberate implementation choices

The pilot uses typed workspace JSONB aggregates and short SELECT FOR UPDATE transactions instead of a separate SQL table for every nested domain record. Relational constraints cover global identity/binding/inbox/outbox uniqueness. See architecture for limits and the path to normalize high-volume records before scaling.

Skill/workflow editing uses labeled form controls validated by shared server schemas. Skill testing is a deterministic policy/prompt preview. Live model-quality testing is an explicit, separately capped operator evaluation; it has not been performed. Optional Telegram web login remains deferred; verified Telegram linking plus local account sessions is the implemented authentication path. Group migration suspends and requires relinking rather than automatically authorizing a new chat.

Readable admin interfaces (2026-09-21)

The remaining JSON displays and editors have been removed from setup, access previews, approvals, workflows, skills, instructions, run details/recovery and operator tools. Structured results use labeled details and lists; operator audit and workspace status use tables. Explicit form controls retain hidden concurrency versions and bound IDs, mask passwords, and preserve failed drafts. Approvals refresh automatically after proposal creation. Copyable commands and Markdown remain literal content; API serialization is unchanged.

Validation: lint, TypeScript, build, all 217 deterministic tests with PostgreSQL, and all 14 browser tests passed. Browser checks include skill publication, workflow payloads and version conflicts, modal locking, automatic proposal refresh, cross-page JSON removal, and desktop/mobile layouts.

Deployed to the local Podman app and worker on 2026-09-21 with image localhost/deepx-agent:readable-ui-20260921 (99cb37e7d5df). The built image passed the isolated Node runtime contract and HTTP/UI smoke checks against disposable PostgreSQL. Compose validation and migrations passed; both deployed containers passed health checks, /healthz and /readyz returned 200, and the worker reported telegram_polling_connected. Existing PostgreSQL and Code Truth containers were preserved. The verified database backup is backups/deepx-before-readable-ui-20260921-160639.dump; the previous image remains tagged localhost/deepx-agent:before-readable-ui-20260921 for rollback.

Borderless icon buttons (2026-09-21)

All admin icon buttons are borderless, including destructive, sidebar and modal actions. Hover backgrounds, keyboard focus outlines, 44px targets and disabled/busy states remain. Lint, TypeScript, build, 217 deterministic tests and 14 browser tests passed; browser style checks also verified normal, destructive, sidebar, disabled, busy and hover states plus keyboard focus.

Deployed image localhost/deepx-agent:borderless-icons-20260921 (b1c0fbfc56d5) to the local Podman app and worker. Compose validation, isolated runtime contract and HTTP smoke checks passed. Both containers are healthy, /readyz returned 200, Telegram polling connected, and the live stylesheet contains the borderless rule. No database schema changes were needed; PostgreSQL and Code Truth were preserved. Verified backup: backups/deepx-before-borderless-icons-20260921-164108.dump. Rollback image: localhost/deepx-agent:before-borderless-icons-20260921.

Modal action layout (2026-09-21)

All create/edit dialogs now place Cancel immediately after their primary action using the shared ModalActions row. Nested dialogs retain independent cancellation, focus return and pending-save locks. Lint, TypeScript, build, 217 deterministic tests and 14 browser tests passed; desktop/mobile dialog screenshots were inspected. Deployed with the refresh loading update below.

Refresh loading feedback (2026-09-21)

Refresh controls now track data loading through completion or failure instead of only the click callback. Operations waits for both health and setup requests; session refresh also shows the shared rotating icon. Busy controls prevent repeat clicks and respect reduced-motion preferences. Lint, TypeScript, build, all 217 deterministic tests and 15 browser tests passed, including delayed success/error responses, actual rotation, reduced motion and combined refreshes.

Both updates were deployed to the local Podman app and worker on 2026-09-21 with image localhost/deepx-agent:modal-refresh-20260921 (77333bd2bae8). Compose validation, the isolated runtime contract and HTTP/UI smoke checks passed. Both containers passed health checks; /healthz, /readyz and updated UI assets returned 200, and the worker reported telegram_polling_connected. No schema changes were needed; PostgreSQL and Code Truth were preserved. Verified database backup: backups/deepx-before-modal-refresh-20260921-165545.dump. Rollback image: localhost/deepx-agent:before-modal-refresh-20260921.

Workspace settings heading (2026-09-21)

Workspace settings now places refresh at the top right and displays its saved version as a badge after the page title. The former bottom version text and reload control have been removed. Desktop/mobile screenshots were inspected; lint, TypeScript, build, 217 deterministic tests and 15 browser tests passed, including the badge updating after save. The refresh test now waits for sign-in completion before navigation.

Deployed to the local Podman app and worker on 2026-09-21 with image localhost/deepx-agent:settings-heading-20260921 (e770fad107b0). Compose validation, the isolated runtime contract and HTTP/UI smoke checks passed. Both containers passed health checks; /healthz, /readyz and updated UI assets returned 200, and Telegram polling connected. No schema changes were needed; PostgreSQL and Code Truth were preserved. Verified backup: backups/deepx-before-settings-heading-20260921-171210.dump. Rollback image: localhost/deepx-agent:before-settings-heading-20260921.

Unified members and access (2026-09-21)

The sidebar now has one Members & access entry. Its searchable, paginated table shows Telegram names/usernames and numeric IDs, role, membership and effective access. Access requests share this page. As of 2026-10-07, all active members have bot access; the whitelist, access mode and policy preview/apply controls and APIs are removed. Old access-policy browser URLs redirect while preserving workspace query parameters; search/pagination also retain those parameters. The member API provides a membership revision for concurrent edits.

Profiles are optional fields in workspace JSON: trusted Telegram sender updates refresh them, and approved requests populate them. Older approved requests provide a read-time fallback until a fresh profile arrives. Removing a Telegram username clears the displayed value, and changing roles preserves the profile. Authorization continues to use numeric IDs; profiles never enroll users or change permissions. No extra Telegram calls, dependency additions or schema migration are required.

Validation: lint, TypeScript, build, 235 deterministic tests and 16 browser tests passed. Checks include duplicate events, tenant isolation, anonymous senders, profile removal/fallback, member edits, mode changes, revocation, last-admin protection and the legacy URL redirect. Desktop/mobile screenshots were inspected. Deployed to the local Podman app and worker on 2026-09-21 with image localhost/deepx-agent:members-access-20260921 (5a9d0fc514dd). Compose validation, the isolated runtime contract and HTTP/UI smoke checks passed. Both containers passed health checks; /healthz, /readyz, the members page and updated UI assets returned 200, and the worker reported telegram_polling_connected. PostgreSQL and Code Truth were preserved. Verified database backup: backups/deepx-before-members-access-20260921-173543.dump. Rollback image: localhost/deepx-agent:before-members-access-20260921.

External release gates still required

Supply a dedicated staging bot/group, model account and spend cap, HTTPS host/domain, protected secret/backup storage, retention agreement and support owner. Perform the credentialed evaluation and complete the staging demonstration. The initial model is provisional until that evaluation passes. No claims about live recap quality, real webhook delivery or pilot readiness follow from fake-provider tests.

Model settings verification (2026-09-19)

Deterministic tests cover a local streaming Chat Completions server receiving the custom endpoint, bearer key, model ID and all seven thinking choices. PostgreSQL coverage verifies operator-only writes, version conflicts, encrypted/redacted keys, retaining a key, requiring a new key when changing endpoints, and worker settings. Browser coverage saves and reloads the custom model, URL and thinking selection. Live provider capability/price validation remains pending.

LLM extensions / plugins

Implemented: explicit Pi 0.85.1 extension-file loading for granted workspaces, registered tools and the documented headless hooks, per-run factory state, limits, configuration pinning and durable tool outcomes. Deterministic tests exercise loading unchanged Pi files, vetoes, unsupported features, grants, revocation, collisions and unknown outcomes. See the compatibility contract. Executable code is operator-reviewed and runs in-process. The Plugins panel manages installed-file registrations, versions, tool/workspace grants and enable/disable/remove actions. Settings, revisions and audits persist in PostgreSQL; workers recheck changes between execution steps. Sandboxing, arbitrary external writes and web package/code installation remain outside this implementation.

On 2026-09-29 the Plugins landing page was reorganized into compact Installed and Markets cards. Code Truth, Codex and file plugin controls now open on dedicated detail routes. Markets shows three dated picks from Pi's package catalog with upstream links; packages are not installed from the panel and compatibility is not asserted. Browser tests cover the catalog, detail navigation, registration, revision conflicts and workspace isolation.

Verification on 2026-09-20: lint, typecheck and build passed; 90 deterministic tests passed with isolated PostgreSQL and NO_PROXY=localhost,127.0.0.1. Both Compose configurations validated. The Docker image passed the Pi extension runtime contract on Node 24.21.0 with network access disabled, including legacy package aliases, and returned HTTP 200 from /healthz with no configured live services.

Plugin UI verification on 2026-09-20: 96 tests passed against isolated PostgreSQL, including operator/tenant/CSRF boundaries, stale revisions, preserved file hashes, worker execution of saved plugins and mid-run revocation. All 5 browser tests passed, including registration, editing, grants, toggles, removal, persistence, conflict handling and desktop/mobile layouts. Lint, typecheck, build, both Compose configurations and the Docker Node 24 extension runtime contract passed. The image serves /admin/plugins and its plugin-management JavaScript bundle successfully without live services.

Predefined local Code Truth verification (2026-09-20)

Implemented the private local MCP service from the sibling Code Truth core, a named Pi extension with its companion skill, and the Plugins repository editor. The UI supports GitHub URLs, network/branch mappings, per-repository workspace grants, enable/disable, removal, and indexing status with commit provenance.

Verified locally:

• Biome, application and service TypeScript checks, and the application build pass.
• 125 deterministic tests pass, including the original source-query tests, real CodeGraph indexing of a local archive fixture, authenticated HTTP MCP isolation, actual Pi factory/tool/schema execution with a fake model, source output, config digests, cancellation, operator/CSRF/grant/revision checks and tool collisions.
• 6 browser tests pass, including repository configuration persistence, branches, grants, enable/disable, removal, unavailable/ready status rendering and responsive desktop/mobile checks. Ready browser status is a deterministic intercepted fixture; real service indexing is checked separately through MCP.
• Both Docker images build. The Code Truth container indexes and queries a local fixture through real CodeGraph and MCP with a read-only filesystem and no external network. The service entrypoint passes liveness and rejects unauthenticated management requests. The Node runtime extension contract also passes.
• Compose configuration validates with both Code Truth and file-extension overrides.

No real repository download, GitHub token, live LLM, Telegram call or production rollout was used. Setup explains service secrets, supported GitHub repositories and separate local-index retention/cleanup.

Independent workspace plugins (2026-09-20)

File plugins and Code Truth configuration now live on each workspace aggregate, with independent revisions and workspace audit entries. The panel uses the selected workspace and discards its local drafts when switching. Saving a workspace’s plugin configuration does not revoke another workspace’s runs, including those owned by the same operator. Executable plugin management remains operator-only.

Migration 006_workspace_plugins.sql splits legacy operator registries by existing grants, preserves pinned hashes and disabled entries, and removes the old deployment registry. Apply migrations before starting the updated API and worker. The API now uses /api/admin/workspaces/:id/plugins and its Code Truth subroutes.

Verification: Biome, TypeScript and build pass; 128 deterministic tests pass with isolated PostgreSQL, including migration, fallback behavior, cross-workspace access, concurrent revision conflicts and same-owner runtime isolation. 7 browser tests pass, including independent plugin versions/repositories, draft reset on workspace switch, and desktop/mobile layouts. No running application data was migrated.

Workspace GitHub App connections (2026-09-20)

The Plugins panel now supports GitHub App authorization, installation and repository selection, connection status, reconnect and workspace-only disconnect. OAuth uses PKCE, one-use state and session/workspace binding. Temporary user credentials are encrypted; ongoing Code Truth access uses short-lived, repository-restricted, Contents-read installation tokens. Connection revisions invalidate pinned runs. Explicit disconnect prevents legacy deployment-token fallback for that workspace. See registration and deployment setup.

Verification: Biome, both TypeScript checks and build pass. 133 deterministic tests pass against isolated PostgreSQL, including authorization boundaries, forged selections, expiry/replay, token redaction, scoped JWT/token requests, credential rotation and independent snapshot namespaces. 8 browser tests pass, including the callback returning to the original workspace, repository selection, persistence, workspace switching, disconnect and mobile layout. GitHub HTTP calls in these tests use deterministic fixtures; real App OAuth remains unverified until the deployment has registered App credentials.

Both images built and passed isolated container smoke tests: Node runtime contracts, application health/UI, and real CodeGraph indexing plus MCP query. Compose validates with and without the GitHub App override. The local Podman stack was updated after a verified database backup; migrations 006 and 007 are applied. App, worker, PostgreSQL and Code Truth are healthy, and /readyz returns 200. Existing private devnet-develop and testnet-develop indexes are ready using legacy credentials. The new Connect button remains disabled until App credentials are configured.

Guided GitHub App creation (2026-09-20)

The GitHub card now offers App Manifest registration for personal or organization accounts. It preconfigures read-only source permissions and callbacks, asks the user to confirm creation on GitHub, and exchanges the returned code server-side. The App credentials are encrypted in an operator-scoped registry, reusable across that operator's workspaces without automatically connecting them or granting repositories. API and worker read saved credentials without restart. Existing environment/file App configuration remains supported. Migration 008_github_app_registration.sql adds this registry and expiring, session-bound registration flows.

Verification: lint, TypeScript and build pass; 139 deterministic tests and 9 browser tests pass, including ownership/CSRF boundaries, cancellation, expiry, replay, session revocation, invalid conversion responses, concurrent creation, credential redaction/persistence, worker access and complete create/connect/disconnect browser flows. GitHub exchanges are fixtures; no real GitHub App was created by tests. Mobile creation layout was inspected. The container passed the Node runtime contract and health/UI smoke tests, and both Compose variants validate. The local Podman app and worker run the tested image with migration 008 applied; health and readiness return 200 and the served UI bundle includes the creation control.

GitHub Manifest correction (2026-09-20): GitHub rejected the local origin in the inactive hook URL. The hook now uses a reserved public example URL, remains disabled, and subscribes to no events; browser callbacks still use the deployment origin. All 139 tests and 9 browser tests pass with regression assertions, along with lint, typecheck, build, Compose validation and the container runtime smoke. App/worker were updated. A real submission through the user's browser now advances past the prior manifest error to GitHub's Confirm access (sudo authentication) page. Completing that verification and creating the App remain user steps; no App was created by this check. See incident notes.

Model-aware execution budgets (2026-09-20)

Workspace run/month budgets no longer have fixed pilot dollar ceilings. Output and context limits come from bundled Pi catalog metadata or explicit operator overrides for custom models. Workspace language and input/output budgets are automatic. Migration 010 removes retired controls from current settings; historical run snapshots remain intact. Runtime checks happen before each reservation/dispatch, using conservative UTF-8 byte estimates, and model capacity is pinned on runs.

Deterministic coverage includes custom-model capability requirements, higher output budgets, oversized tool-context continuation, pre-dispatch rejection, zero-cost reservations, field-level API validation, authorization and version conflicts. Browser coverage verifies settings errors, saving above old caps and desktop/mobile layouts. Provider capability values are not verified with live calls; large Telegram replies still follow the existing delivery truncation. See configuration.

Telegram access requests (2026-09-20)

Implemented unauthorized-user request buttons, scoped request links, deduplicated pending requests and admin approval/rejection under Members. Approval atomically creates/reactivates regular membership for the verified Telegram ID and queues a confirmation with normal delivery authorization checks. Requests use workspace JSONB; migration 009_access_requests.sql scopes fixed access-help deliveries.

Deterministic PostgreSQL checks cover duplicate updates/requests, concurrent approvals, CSRF/admin/tenant boundaries, stale decisions, rejection retry delay, revocation replay, private/group routing, forged actors, deletion and retention. Telegram calls use test transports; no live Telegram message or production migration was performed.

Verification: lint, strict typecheck and production build pass; 150 automated tests pass with isolated PostgreSQL, and all 11 browser tests pass. Browser coverage includes stale-policy recovery, approval/rejection, member-list refresh, persistence and desktop/mobile layout. Both access-request screenshots were visually inspected.

Local deployment update (2026-09-20)

Model calls per run accept integers from 1 to 20, default 3. Boundary tests reject 21, zero and fractional values. The current source passed Biome, TypeScript, build, 150 deterministic tests against disposable PostgreSQL, and 11 browser tests. The Docker-format image passed the Node 24 runtime contract and isolated HTTP/UI smoke checks. Compose configuration was validated with the existing Code Truth override.

The local app and worker were updated after a protected database backup and migrations; both containers are healthy, /healthz returns ok, /readyz returns ready, and Telegram polling reconnected. No model request was sent for deployment verification. Custom model capability values must be configured before new model runs; service readiness does not validate these provider limits.

Automatic response settings (2026-09-20)

Reply language, Input byte budget and Output token budget have been removed from workspace settings and their API output. Older clients' retired fields are discarded on save. Migration 010 removes existing overrides and advances affected workspace versions without changing historical run snapshots. No workspace language is injected into the prompt; the model's response is returned in its chosen language.

Per-call output automatically fits the model's maximum, remaining context capacity and remaining run/month USD budgets. Atomic reservations still enforce concurrent budget limits. Tests cover removal, legacy migration/idempotency, preserving model text, context/budget adaptation, zero-cost calls, and pre-dispatch budget denial.

Local verification: Biome, TypeScript, build, 154 deterministic tests and 11 browser tests passed. Desktop/mobile settings and the Docker image were inspected. The local app and worker were redeployed after a protected backup; migration 010 is applied, no current workspace retains the retired controls, and liveness/readiness are healthy.

Run deadline diagnosis (2026-09-20)

The executor now distinguishes deadline expiry (run_timeout) from cancellation and worker shutdown (worker_shutdown), preserving the first abort source in logs and run details. RUN_TIMEOUT_SECONDS defaults to 300 seconds and accepts 1–1800; new run queue jobs get an additional 60 seconds for finalization. Migrations update existing queue defaults. Conversation leases continue to renew while running. Unknown reservations remain unchanged and terminal runs are not automatically replayed. Deterministic regression tests cover timeout, shutdown, cancellation, retained reservations, duplicate jobs, queue expiration and configuration validation.

Paginated usage table deployment (2026-09-20)

Usage & budget now renders accounting attempts in a paginated table with status, reserved/actual USD, model/run ID, dates, record counts and mobile table scrolling. Local verification passed Biome, TypeScript, build, 159 deterministic tests and all 12 browser tests, including pagination, empty/error states, refresh and mobile layout.

The Podman Docker-format image localhost/deepx-agent:usage-table-20260920 (b15143713176) passed the isolated Node runtime contract and HTTP/UI smoke checks. Compose configuration validated; a protected database backup was verified and the migration command completed before app/worker replacement. Both containers, PostgreSQL and Code Truth are healthy; liveness/readiness return 200, polling reconnected and the served bundle includes the usage table. The previous image is retained as localhost/deepx-agent:before-usage-table-20260920 for rollback.

Telegram GitHub issue submission (2026-09-20)

Implemented a built-in propose_github_issue tool for workspaces with a connected GitHub installation. It shows the complete repository/title/body in Telegram and requires the requesting actor's explicit approval before the application worker creates an issue and returns its link. Repository selection, connection revision, actor access, cancellation, expiry and pause controls are rechecked before sending. Durable reservations prevent concurrent workers or restarts from replaying a POST; ambiguous outcomes require checking GitHub before requesting another draft.

Guided App creation now requests Issues write permission. Existing Apps/installations must approve that permission update on GitHub. Code Truth tokens remain Contents-read; issue tokens are limited to Issues-write on the single approved repository. There is no panel issue form and no new dependency or database migration.

Verification: Biome, TypeScript and build pass; 173 deterministic tests pass with isolated PostgreSQL and 12 browser tests pass. New tests cover the real Pi factory, complete approval previews, tenant/actor/repository boundaries, concurrent workers, revocation during token minting, permission/error outcomes, ambiguous recovery and workspace deletion. GitHub and Telegram calls use fixtures. No live issue was created, no App permission was changed, and this change has not been deployed.

Rich replies and native Stop (2026-09-20)

Implemented Bot API 10.3 rich model replies in private chats, groups and schedules, with native headings, lists/tasks, quotes, code, dividers, compact aligned tables and safe inline formatting. Private interactive replies stream with native Stop enabled; partial drafts are not automatically saved. Previews are throttled, source-checked, fenced and exclude thinking/tool arguments. A persisted bot/chat/topic/draft/fence binding routes Stop independently of workspace selection, through authenticated, deduplicated ingress. The run guard aborts generation, blocks pending publication and keeps unknown provider reservations. Confirmed rich-send rejection falls back to text; ambiguous sends are never blindly replayed.

No new dependency or SQL migration is required. Existing webhook subscriptions must be explicitly re-registered to include the Stop update; polling subscribes after worker restart. See manual staging example.

Verification: Biome, TypeScript, production build and the full deterministic suite pass against disposable PostgreSQL. Coverage includes native block rendering, preview timing/retries, duplicate/stale/cross-tenant Stop events, revoked access, fencing, paused execution, provider cancellation/accounting, pending final delivery, authenticated webhook handling and polling subscriptions. Telegram/model calls use fixtures. Local deployment is recorded below; end-to-end live Telegram behavior remains unverified.

Internal model cost query (2026-09-20)

Implemented the skill-granted query_model_cost application tool: private-chat queries for the actor's usage or owner/admin workspace totals, grouped by model over today, this month or retained history using UTC. Settled costs, reservations and unknown amounts remain distinct; workspace reports include the current monthly budget. Reports use existing tenant accounting and do not call provider billing APIs. Existing published skills need an explicit grant; new starter skills include it. No dependency or migration is added. See usage examples.

Verification: Biome, strict TypeScript and production build pass; 197 tests pass against disposable PostgreSQL. New coverage exercises UTC boundaries, zero costs, unresolved charges, per-model totals, tenant/actor/admin/private-chat boundaries, grants, cancellation, fencing, revocation, replay argument conflicts and a complete fake-provider Pi execution with durable accounting and duplicate-run recovery. Included in the local Podman deployment recorded below; live Telegram/model behavior remains unverified.

Local rich replies / Stop deployment (2026-09-20)

Updated the local Podman app and worker to image localhost/deepx-agent:rich-stop-20260920 (484928d97634), also tagged :local. The image includes the current verified workspace source, including rich model replies, private streaming and native Stop. The existing Code Truth service and PostgreSQL volume were retained. Migration completed with all ten migrations applied.

Pre-release verification: lint, strict TypeScript, build and 197 deterministic checks passed; all 12 browser tests passed against disposable PostgreSQL. The Docker-format image passed the isolated Node 24 runtime contract and HTTP/UI smoke checks, and the existing Compose configuration with Code Truth validated. A protected custom-format database backup was created and its archive verified at backups/deepx-before-rich-stop-20260920T102943Z.dump (mode 0600). No runs or sends were active at replacement. The previous image is retained as localhost/deepx-agent:before-rich-stop-20260920 (34c183fc43a4) for rollback.

Both deployed containers are healthy and use the tested image. /healthz returns ok, /readyz returns ready, and Telegram polling reconnected. The running worker contains rich-send/draft support and the native Stop update subscription. This stack uses polling, so no webhook registration was needed. No test Telegram message or paid model request was sent for release verification. Temporary test containers were removed; end-to-end Telegram rendering and Stop remain a user staging check.

Codex implementation workflows (2026-09-21)

Local implementation adds workspace-scoped repository/base-branch/maintainer configuration, actor-bound coding approvals, durable issue creation and Actions dispatch, progress/PR reporting, cancellation requests and an admin task list. The supplied repository workflow runs Codex and checks, then publishes a draft PR from a fresh job. No live GitHub/Codex run has been performed; repository setup and a staging run remain required. See setup and recovery.

Validation: bun run check, bun run typecheck, bun run build, and Node runtime contracts passed. bun test passed 211 tests with PostgreSQL integration tests enabled, including 14 coding-specific tests. All 12 browser tests passed, including repository/maintainer configuration, edit cancellation and mobile layout. The repository workflow template passed actionlint 1.7.12. External Codex execution, GitHub publication and paid-provider billing were not exercised.

Local native-streaming deployment (2026-09-21)

Updated the existing Podman app and worker to localhost/deepx-agent:native-stream-20260921 (c44c74b5f22a), also tagged :local. Private requests now start with Telegram's native thinking draft, stream the answer, and persist one final reply. Normal requests, recaps and one-off corrections no longer send queue acknowledgements. This image contains the current workspace source, including the Codex implementation described above.

Validation: all 214 tests passed with PostgreSQL integration enabled, along with lint, strict TypeScript and build checks. All 12 browser tests passed against a disposable database. The Docker-format image passed its isolated Node 24 runtime contract and HTTP/UI smoke tests; the existing Compose configuration validated.

No runs or sends were active at replacement. The protected database backup is backups/deepx-before-native-stream-20260921.dump (mode 0600; archive verified). The previous image remains tagged localhost/deepx-agent:before-native-stream-20260921 (484928d97634). PostgreSQL storage and the running Code Truth service were retained. The migration container exited successfully. Both app and worker are healthy, /healthz reports ok, /readyz reports ready, and Telegram polling reconnected. No test messages or paid model requests were sent; live rendering remains a staging check. Temporary test containers were removed.

Private conversation threads (2026-09-21)

Standalone private requests create user/workspace threads; replies to retained user messages or delivered answers continue them. Thread sources are isolated and refreshed at execution, with ordered runs and separate recovery transcripts. The read-only query_chat_history tool searches owned private threads with bounded excerpts, cursor pagination, retention and citation checks. Legacy records are not inferred into threads. No SQL migration or new dependency is required.

Deterministic unit and isolated PostgreSQL tests cover reply mapping across restart, duplicate ingress, cross-user/workspace denial, ordered execution, history citations and retention. Live Telegram/model acceptance remains a staging check. See design and examples.

Native private Topics (2026-09-22)

Private interactive requests now prefer Telegram's native message_thread_id over reply anchors. Same-topic messages share a persisted workspace/user/chat/bot thread without requiring Reply; different topics have separate default context. Cross-topic and unavailable reply anchors do not switch the current topic. Outside Topics, the existing reply-based fallback remains. Native mappings persist in workspace JSONB without a SQL migration or new dependency; pre-upgrade threads remain searchable but are not automatically merged into native topic history. Help, design rules and manual examples now describe native Topics as the primary conversation interface, with no parallel New/Recent conversation controls.

Validation: bun run check, bun run typecheck, bun run build and all 247 tests passed with disposable PostgreSQL and fake external transports. The full test run used NO_PROXY=localhost,127.0.0.1, TEST_DATABASE_URL and bun test --timeout 30000. The initial focused run hit an existing streaming test's default five-second timeout; that test and the full suite passed with local proxy bypass and the longer limit. Coverage includes native topic continuation, conflicting reply anchors, user/bot/ workspace isolation, concurrent duplicate ingress, restart persistence, ordered execution and failed predecessors, retained history, topic-preserving thinking/ streaming/final delivery, and native Stop isolation. No deployment or live Telegram messages were performed for this change. See examples.

Discussion memory and Topic compaction (2026-09-22)

Native Topics now support internal discussion summaries, decisions, todos, source links and overlapping relations through core Pi tools. Model policy continues short follow-ups, answers clear topic changes without confirmation, and retrieves older summaries/original excerpts when needed. Record updates are staged and commit with a visible answer; they do not create approvals, persistent instructions or external actions. Source hashes, retention and topic/actor scope guard reads, updates and replay. The bounded catalog uses literal search and model semantic judgment, without an embedding service or extra classifier call each turn.

Context places frozen summaries and growing history before changing metadata. Stable hashed workspace/thread session keys support Pi/provider cache routing. At 80% of available context after output reserve, bounded summary batches target 40% while keeping the two most recent runs. Summaries use the configured model with no tools/extensions or previews and share run budgets, deadline and cancellation. Separate durable checkpoints recover completed summaries without another model call; unknown charges are not replayed. Original sources remain available under retention, and edits/deletion invalidate derived memory. Provider cache token counts are retained on each usage attempt, including compaction charges.

Validation: bun run check, bun run typecheck, bun run build and all 261 tests passed. The full test run used NO_PROXY=localhost,127.0.0.1, TEST_DATABASE_URL and bun test --timeout 30000 with deterministic providers and disposable PostgreSQL. Tests cover prefix stability, discussion updates/relations/retrieval, permissions and replay, threshold compaction, accounting, invalid summaries, budget/cancellation/ source-removal failures, unknown outcomes, and checkpoint recovery. No dependency, SQL migration, deployment, live Telegram send or paid model call is introduced by local verification. Live summary fidelity, semantic recall and actual cache hit rates remain evaluation work. See design and examples.

Group conversations and natural follow-ups (2026-09-22)

Interactive group requests now keep per-user continuity scoped to workspace, bot, group and native Topic (including topic zero). Replying to a confirmed retained bot answer can join a public discussion. Group history retrieval cannot access private chats or other topics. Discussion tools, frozen summaries, stable cache identities and threshold compaction are reused. Manual recaps and schedules keep group coverage.

A plain received message can become a candidate only after that same user's recent confirmed answer. Local filtering excludes other addressees, commands, stale messages, anonymous identities and unauthorized users. A bounded tool-free Pi classifier must return REPLY before ordinary generation. Classification has its own durable checkpoint and usage purpose, shares budgets and cancellation, and remains silent on uncertainty or failure. Ignored candidates are cleared from conversation history; ordinary consented collection remains independent. No dependency or SQL migration was added.

Local tests cover user/topic/bot/workspace isolation, shared replies, discussion retrieval, prefix stability, compaction, duplicate events, completed-checkpoint recovery, source changes, revocation, budget/turn limits and unknown-charge recovery. Live classifier precision, Telegram visibility and provider cache hit rates remain staging evaluation work. No deployment, BotFather changes or live messages were performed. See design and examples.

Validation: bun run check, bun run typecheck, bun run build and all 278 tests across 36 files passed. The full suite used disposable PostgreSQL, fake Telegram/model transports, NO_PROXY=localhost,127.0.0.1, TEST_DATABASE_URL and bun test --timeout 30000. Additional gate cases verify that changing addressee, cancellation and source removal during classification suppress the answer.

Local Podman deployment (2026-09-22)

Updated the existing deepx-code-telegram-bot stack using compose.yaml plus examples/code-truth.compose.yaml. App and worker now run image ID sha256:e0a006674e7a7c8b99c0f7f19ff7cc7927a79c0bf3086a1c0cdbc96821118071, tagged localhost/deepx-agent:release-20260922-topics and localhost/deepx-agent:local. This includes native Topic memory, group conversation continuity and natural follow-ups. Model configuration retains Chat Completions. PostgreSQL and Code Truth containers were not recreated.

Release evidence: the unchanged source had passed all 278 unit/integration tests and check/typecheck/build; all 16 browser tests passed before deployment. The image passed the Node 24.21.0 / Pi 0.85.1 runtime contract with networking disabled, and migration plus app/worker startup passed against a separate disposable database. Compose configuration validation passed. The production database had no queued or running assistant work or pending/sending deliveries at the pre-cutover check.

A protected pre-deployment PostgreSQL archive was created and its archive directory validated at backups/deepx-pre-topics-20260922T043308Z.dump (ignored by Git). Migration completed successfully; the schema remains at 10 applied migrations. Previous image c8ef95b29cd6a212750d60096d59198d5f27579895f3ff203a845ba98bbd4c29 is retained as localhost/deepx-agent:rollback-pre-topics-20260922.

Post-deployment app and worker health checks passed; http://127.0.0.1:3000/readyz returned ready, and the new worker logged telegram_polling_connected at 2026-09-22 04:34:06 UTC. Temporary browser/database/smoke containers were removed. No manual Telegram test message or paid model evaluation was sent. Actual follow-up classification quality and cache hit rates remain live acceptance work.

Silent unsupported media deployment (2026-09-22)

Non-text requests no longer receive the text-only pilot warning. Unsupported media, including captions, is silently ignored by the ask/recap handler.

Updated the existing Podman app and worker to localhost/deepx-agent:quiet-media-20260922, also tagged :local, image 3d4f7a5356e571af1d9a60d2c4924b195fcd2e9da63235e5a78bf311dc4f1df1. Lint, strict TypeScript and build passed, along with all 278 tests against disposable PostgreSQL and all 16 browser tests. The Docker-format image passed the isolated Node runtime contract, migration and HTTP/UI startup smoke checks. Compose validated.

No runs or deliveries were active at the pre-cutover check. The protected backup backups/deepx-pre-quiet-media-20260922T0704.dump was created and its archive directory verified. The prior image 82dc91185b8992898417447f20e792dd4a74cbc887d4054cfc7b71e74b3655be is retained as localhost/deepx-agent:rollback-pre-quiet-media-20260922. Migration completed successfully; PostgreSQL and Code Truth were not recreated.

Both replaced containers are healthy, liveness/readiness return ok/ready, and Telegram polling reconnected at 2026-09-22 07:05:11 UTC. The removed warning is absent from the deployed worker bundle. Temporary test containers were removed. No manual Telegram test message or paid model evaluation was sent.

Local Codex Podman backend — 2026-09-22

• Added the optional podman coding backend alongside the existing GitHub Actions default, with panel selection and per-repository setup/check commands pinned to the maintainer approval. Provider settings follow the inspected device's custom Responses-provider configuration. Workspace operators can save, replace and remove API keys in the panel; keys are encrypted and redacted. The supervisor environment variable CODEX_PROVIDER_API_KEY remains an optional fallback.
• Added a trusted supervisor, per-task rootless containers/volumes, bounded execution, temporary model proxy credentials, separate preparation and publication, durable task state, cancellation, restart reconciliation and retention cleanup. Publication rechecks authorization after minting repository-scoped GitHub write credentials.
• Validation: full suite with isolated PostgreSQL 286 passed; final focused coding/runner suite after recovery hardening 24 passed; browser suite 16 passed. Biome, TypeScript and Bun build passed. All three Docker image targets built with Podman; the app served health/setup/assets successfully. The real Podman smoke passed with fake Codex and a local Git fixture. Merged Docker Compose configuration validated. Tests excluded this device's HTTP proxy for loopback requests. No live model call, GitHub publication or deployment was performed.
• Panel credential follow-up: save/replace/remove uses workspace-bound AES-256-GCM, redacted responses, revision invalidation and active-task cancellation. The runner retains an encrypted credential only while preparing/running. Validation: 292 tests passed, 16 browser tests passed, and 11 focused coding integration tests passed after adding key-rotation cancellation coverage. Biome, TypeScript, build, all three image targets, app health/assets smoke, real Podman smoke without an environment API key, and Compose validation passed. No deployment performed.
• Existing Telegram issue-to-PR semantics remain. Codex session files/thread IDs are retained, but continuing a completed task from Telegram is not exposed yet. Live custom-provider execution and GitHub publication remain staging gates. See local runner setup and boundaries.

Codex panel credentials deployment (2026-09-22)

Updated the existing Podman stack with compose.yaml, examples/code-truth.compose.yaml and deploy/codex/compose.yaml. App and worker run localhost/deepx-agent:codex-panel-20260922 (also :local), image 8eb16f6b41ffd6d5c553b3f41437b8408902270929c7bda3d17f27ed8163caee. The new healthy Codex supervisor uses image 0028cf13ee5f; the prebuilt local job image is 39beb8943be7. Private deployment configuration now includes the Codex overlay, a generated runner authentication token and the rootless Podman socket. Provider API keys are configured through the workspace panel; no provider key was copied from the host or configured as an environment fallback. Existing workspace backend selections and plugin enablement were preserved.

Pre-cutover counts showed no queued/running runs, pending/sending deliveries or active coding tasks. A protected database archive was created and validated at backups/deepx-pre-codex-panel-20260922T100107Z.dump; the previous private environment was also backed up. The old app image remains tagged localhost/deepx-agent:rollback-pre-codex-panel-20260922. Migration completed. PostgreSQL and Code Truth containers were not recreated.

The rebuilt app image matches the previously tested image, and its isolated Node runtime contract passed. Compose validation and deployment health checks passed. The app serves the API-key management bundle, the worker reaches the supervisor, and the supervisor confirms rootless Podman access. /readyz returns HTTP 200 with ready; Telegram polling reconnected at 2026-09-22 10:02:32 UTC. No manual Telegram message, live model test or GitHub publication was performed.

Local-only Codex and workspace device sign-in (2026-09-29)

Removed the GitHub Actions coding backend from new settings and dispatch. Existing Actions tasks stop locally with a migration status. The Podman supervisor remains the sole coding execution path. Split local work into preparation, setup, Codex, checks and publication phases; setup and checks receive neither model credentials nor GitHub write tokens. Improved timeout, patch export and cleanup handling.

Added a per-workspace sign-in choice beside the custom provider API key in the Codex panel. ChatGPT device-code login runs in the supervisor, displays the OpenAI verification link and one-time code, and seals the account cache per workspace. Only the Codex implementation container receives a temporary copy; disconnect and workspace deletion remove the stored credential. Device-code tasks require selected private repositories and verify current GitHub visibility again before issue creation and before starting Codex. The custom provider mode and its per-workspace API key remain available.

Biome, TypeScript, Bun build and Compose validation passed. Focused coding and device tests passed, including real Podman smoke for both sign-in modes with a fake Codex executable and local Git fixture. The full suite with loopback proxy disabled had 307 passes and two unrelated native Telegram Stop failures; the additional repository privacy regression test passed afterward. The coding panel browser scenario passed in the full browser run, while a separate run summary assertion failed. At this validation point, no live ChatGPT sign-in, model call, GitHub write or deployment was performed; the existing Podman stack still ran its prior :local images.

Local-only Codex Podman deployment (2026-09-29)

Updated the existing local stack using compose.yaml, examples/code-truth.compose.yaml and deploy/codex/compose.yaml. The app and worker now run localhost/deepx-agent:device-20260929 (also :local), image 2232179827b4d0cd66236d3c0ef87dd54b5ad43aee8f6ba097f8dfe6962ba46a. The Codex supervisor runs localhost/deepx-codex-supervisor:device-20260929 (also :local), image 20ab636b90bd51f847d64c5a8097b23830f0769ca0bf483019cd3539d9219cad. The prebuilt job image is localhost/deepx-codex-job:device-20260929 (also :local), image 22b9810f7af650ceef9ebed5e9b23ca4d834fe80a2d634e9a5ae2b84c8e25551.

Pre-cutover checks found no active runs, deliveries or Codex tasks. A protected PostgreSQL archive was created and validated at backups/deepx-pre-local-device-20260929T071754Z.dump. Previous app, supervisor and job images are retained under :rollback-pre-device-20260929. Migration completed without changes. PostgreSQL and Code Truth were not recreated.

The three images built successfully. The exact job and supervisor images passed the real Podman smoke with fake Codex for provider-key and device-code tasks; the app image passed the isolated Node runtime contract. Merged Compose configuration validated. After cutover, app, worker and supervisor are healthy; /readyz returns HTTP 200, the served UI bundle includes both sign-in methods, the worker can reach the authenticated device-status endpoint, and Telegram polling reconnected. No live ChatGPT sign-in, model call or GitHub publication was performed.

Verified Telegram GitHub account linking — 2026-10-05

Added private /github connect, /github sync and /github disconnect controls for eligible members. GitHub App OAuth uses PKCE and a one-use ten-minute flow; the same Telegram sender confirms the stable GitHub identity before linking. Encrypted user/refresh credentials stay outside workspace API responses. Permission snapshots sync about every five minutes, intersect user access with workspace-selected repos, and appear in Members & access. Linked users' reads/issues require read access; coding also needs write/admin plus the existing maintainer grant. Missing, stale or failed snapshots deny access; downgrades invalidate pending work. Automatic token rotation is serialized and refresh failures clear grants. Existing unlinked members retain the manually configured policy; mandatory linking is not enabled globally.

Apply migration 012_github_users.sql. Documentation and a runnable Telegram command example were added. Local verification passed 328 tests across 41 files against disposable PostgreSQL and 18 browser scenarios, including member permission display and search. The final focused linking/permission suite passed 11 tests. The build, strict typecheck and checks scoped to these changed files passed. Repository-wide lint still reported issues in concurrently edited Codex collaboration files; these edits were preserved. The final table-layout follow-up also passed three focused browser scenarios. No deployment, account connection, live Telegram message or GitHub write was performed. See GitHub linking and command examples.

Admin site custom domain (2026-10-05)

Deployment operators can configure the admin panel hostname from Site domain. The persisted deployment setting updates admin origin checks and GitHub/Telegram callback generation in the app and worker without restarting. Domain changes are versioned and audited atomically, expire pending GitHub connections and require explicit webhook re-registration. The original environment address remains a recovery origin. DNS/TLS/proxy provisioning and reachability remain external and unverified; the panel provides instructions and callback URLs.

Validation: all 350 unit/integration tests passed with isolated PostgreSQL and NO_PROXY=localhost,127.0.0.1; lint, strict type checking and build passed. The new browser scenario passed on desktop and at 390px width, including invalid-hostname errors, retained drafts, saving and restoring the default address. Domain-specific coverage includes operator-only access, Origin/CSRF, normalization, persistence, revision conflicts, audit deduplication, secure cookies, generated GitHub callbacks (including Telegram linking), explicit webhook registration and removed-origin rejection. Public DNS, certificate issuance and live provider changes were not run.

2026-10-07 card update: Update connected services groups GitHub App URLs into labeled, copyable rows and separates Telegram guidance. Copy controls briefly show a check on success and provide manual-copy guidance if clipboard access fails. Long URLs wrap on narrow screens in both themes. Before settings load, labels stay visible and copy controls remain disabled, including after a failed initial request. Polling needs no webhook update; webhook mode retains its status and explicit registration action. Deterministic browser coverage in tests/browser/connected-services.e2e.ts checks copying, clipboard failure, loading/retry, both Telegram modes, and light/dark layouts at 1280px and 390px.

Codex device sign-in recovery (2026-10-05)

The configuration dialog now shows progress while starting device sign-in, bounds the request to 20 seconds, and unlocks controls with an actionable error if it stalls. Recheck connection checks runner availability and recovers a pending login without closing the dialog or discarding unsaved configuration. Pending status polling uses bounded, sequential requests and ignores results after its owning effect has stopped.

Lint, strict typecheck and build passed. All 352 unit/integration tests passed against disposable PostgreSQL. Four deterministic browser scenarios passed: unavailable-runner recovery with draft preservation, code display and polling to connected, failed-login retry, and timeout followed by recovery of a pending code. The live dev.stonegate.me configuration dialog was inspected and reported the runner unavailable; no live account login, model call or deployment was performed. Restoring that deployment's runner and applying this UI change remain outstanding.

Codex repository access through the connected GitHub App (2026-10-06)

Both Codex model authentication methods now support selected public and private repositories. The connected workspace GitHub App supplies fresh installation tokens scoped to the repository and operation, including private checkout access. ChatGPT account authentication no longer adds a repository-visibility restriction. Maintainer grants, connected-repository checks, account sign-in, reviewed approvals, explicit Direct execution policy and publication fencing remain enforced.

Removed the private-only checks during proposal, reviewed issue creation, runner start and continuous-task authentication recovery. Updated panel guidance and operator examples. Historical private-only behavior above describes older builds.

Validation: lint, strict typecheck and build passed. All 400 deterministic tests passed against disposable PostgreSQL with loopback traffic excluded from the host proxy. Coverage includes public/private reviewed and direct starts, scoped GitHub credentials, denied GitHub access, duplicate dispatch, maintainer boundaries and account recovery after repository visibility changes. No production configuration, account connection, GitHub write, Telegram send or deployment was performed.

Automatic repository metadata refresh (2026-10-06)

Manage GitHub and Codex repository lists refresh while visible and on return from GitHub. Five-second polling uses Metadata-read installation tokens, matches saved repositories by numeric ID, updates names/visibility and matching Code Truth URLs, and removes IDs no longer accessible to the App. New IDs still require workspace authorization/selection. Expanded lists, selected IDs and unsaved editor values remain intact; concurrent settings edits retain their version-conflict protection. Failures retain saved metadata with a notice, and delayed responses cannot undo disconnect, reconnect or revoked ownership. Webhooks remain disabled; this is near realtime polling while the panel is open.

Validation: bun run check, bun run typecheck, bun test (411 tests against disposable PostgreSQL) and bun run build passed. All 26 focused browser scenarios passed across repository refresh/search, Codex sign-in and loading states, including rename/link updates, retained drafts and stale-save rejection. No dependency or migration was added. No live GitHub call, Telegram send or deployment was performed. See behavior and limits and the manual staging check.

Compact run history and pagination (2026-10-08)

The runs API now returns explicit summary fields and pagination metadata, with a default page size of 25 and a validated maximum of 100. The panel fetches 25 items per page and retrieves full messages, attempts and delivery records only through the existing detail endpoint. Page transitions discard previous-page rows; loading, retry, empty and out-of-range states are explicit. Detail/back navigation and browser history preserve the list offset. Workspace read and mutation permissions are unchanged. See the API contract and runnable requests.

2026-10-10: Pagination now follows the run list and its loading/empty states. Desktop (1280px) and mobile (390px) screenshots were inspected for populated, loading, empty and error states. All 10 Runs browser scenarios, 661 deterministic tests with disposable PostgreSQL, lint, strict TypeScript and the build passed.

Compatibility: list items no longer contain full run records, and the default page size changes from 100 to 25. Clients needing details must use the detail endpoint; clients needing 100 summaries can specify limit=100. The workspace is still loaded from its JSON database record; this reduces response size and summary construction work, not database reads. Offset pages can shift when new runs arrive or retention removes older runs; they are not a snapshot.

Verification: bash scripts/verify-skills.sh quality runs dependency preparation, quality/type checks, deterministic PostgreSQL tests, the build and Node runtime contract. bash scripts/verify-skills.sh runs-browser prepares its own database and runs the runs/admin browser suites, including desktop/mobile pagination, on-demand details, loading/error recovery and preserved navigation.

Dedicated run detail pages (2026-10-06)

Compact run cards link to /admin/runs/:runId?workspace=:workspaceId instead of opening a modal. The page retains messages, coverage, pinned versions, the model attempt timeline, delivery evidence and existing authorized recovery/cancel/retry controls. Back to Runs preserves the workspace and list offset. Direct visits and refreshes fetch a single run independently of list pagination through GET /api/admin/workspaces/:id/runs/:run; missing runs and failed loads show an explicit error with navigation and retry controls. Access remains scoped to the workspace, and unlinked deployment administrators retain read-only run access. See the runnable requests and page URL.

Validation: bun run check, bun run typecheck, bun test (415 tests against disposable PostgreSQL) and bun run build passed, with the host HTTP proxy disabled for local test fixtures. Four browser scenarios passed for desktop/mobile layouts, keyboard navigation, direct visits, refresh, back navigation, pagination, load recovery and switching to a missing run. Desktop and mobile screenshots were reviewed. No dependency or migration was added.

Bot access follows active membership (2026-10-07)

All active workspace members can use the bot. The Members & access page no longer shows access modes, a whitelist column/toggle or policy preview/apply controls. Roles, repository grants, workspace pause and member deactivation still apply to requests, callbacks, queued work, schedules and delivery. Access-request approval and owner verification enroll active members without a second access list.

Migration 015_remove_workspace_whitelist.sql removes saved workspace policies and preserves their concurrency revision as memberVersion. Active members previously excluded by a whitelist now have bot access; listed non-members remain non-members, and existing cancelled work/suspended schedules retain their state. The member API returns a membership revision with searchable, paginated profiles; retired access-policy API routes are unavailable. Old browser links still redirect.

Deterministic and PostgreSQL checks cover active/inactive/non-member and role boundaries, revocation of runs/approvals/schedules/delivery, stale edits, last-admin protection, access-request replay, repeated migration and retired APIs. All 17 admin browser tests passed, including membership deactivation/reactivation across reload, removed controls, Telegram profiles, desktop/mobile layouts and setup enrollment. Lint, TypeScript and production build passed. See membership and access.

Telegram repository picker (2026-10-08)

/repos now opens native inline buttons for up to ten active workspace-selected repositories. The requester's retained mentions and recent selections rank first; remaining repositories use saved GitHub push/update timestamps with stable name ordering when activity is unavailable. /repos <name> searches beyond the first ten. The current selection is labeled and checked, with a Clear selection action. Archived/disabled repositories, denied upstream grants and private group choices are excluded. GitHub timestamps refresh through existing admin repository refresh; timestamp-only updates preserve the connection revision and existing grants.

Selection is persisted per workspace, actor, bot and chat/topic, expires with retention, and supplies default repository context to subsequent interactive runs. Original requests and explicit targets are preserved; selection grants no write or coding authority. Callback controls bind to a confirmed delivery, actor, bot, destination and current connection, expire after fifteen minutes and are consumed after selection. New menus revoke older controls and queued replies. Access and current saved delivery state are rechecked immediately before sending. Workspace deletion purges selection state.

The active worker publishes Telegram's default/group command menus and a private menu that includes /github for account linking alongside /repos, Ask, Status, Cancel and Help. /github reuses the existing verified private linking flow; /github connect, sync and disconnect remain supported. Registration runs outside polling and maintenance, is once per worker/credential, honors retry backoff and stops old-client writes after credential changes. Scoped menu visibility does not replace request authorization. Explicit language/chat-specific Telegram overrides can take precedence over these menus.

Validation: 610 deterministic tests passed against disposable PostgreSQL with host proxies disabled. Focused tests cover ranking/name boundaries, ten-row limits, search and empty states, selection/clearing, actor/topic/bot isolation, duplicate updates, copied/stale/revoked callbacks, pending delivery revocation and retention. Lint, strict typecheck, build and the Node runtime contract passed. No dependency or migration was added. No live Telegram send, GitHub call or deployment was performed; native Telegram desktop/mobile rendering remains a staging gate. See the runnable picker journey.