RepoDesk
Browse documentation

REPODESK DOCUMENTATION

Original Markdown from stonega/repodesk. This versioned snapshot includes proposals and research, not just shipped features. See the implementation evidence for current scope and limits.

View source at 521ccc0

Telegram image and attachment input

Implemented locally, 2026-10-06. Live Telegram/provider behavior still needs the manual staging check.

Supported input

  • Telegram photos, including messages with no caption. The largest photo variant is selected. Images sent as documents are supported too.
  • JPEG, PNG, GIF and WebP images, identified from bytes. Images reach Pi as base64 image content blocks, paired with their source ID in the text prompt.
  • UTF-8 text/code files, including Markdown, JSON, CSV and ordinary source files.
  • Unencrypted PDFs with selectable text, extracted with pinned unpdf 1.8.1 (its serverless PDF.js distribution). No PDF scripts, rendering, remote font retrieval, OCR, audio transcription, archive extraction or Office-file parsing.

Limits are application policies: 5 MiB per image/PDF, 256 KiB per text/code file, 32,000 extracted text characters per file, and 40 PDF pages. A request uses at most four files and 10 MiB of downloaded data from its unsummarized conversation context. Current input has priority. Older inaccessible/unsupported or over-limit references are marked unavailable; a bad earlier file does not block a new text request. Files are never silently truncated. Scanned PDFs, oversized files, unsupported formats and download failures get readable feedback. Send scanned pages as images. Album messages are independent Telegram updates; album aggregation is not implemented.

Routing and authorization

Private images/files start a request with their caption, or a brief default request when no caption exists. Native Topics and reply anchors retain their existing conversation behavior. In linked groups, use a caption command/mention or reply to the bot. Caption entities use Telegram's UTF-16 offsets. Unaddressed media does not open a follow-up window; group collection consent may retain its metadata.

Authenticated, deduplicated ingress retains caption and file metadata on the tenant's source record. It performs no file download. The worker checks current run policy, source retention, bot identity and bot credential generation before/after fetching and decoding. Files are downloaded only from validated Bot API relative paths; redirects, arbitrary URLs and traversal are rejected. Metadata and streamed bytes both enforce limits, even without Content-Length. Downloads share run cancellation and have a 15-second timeout. File URLs/tokens never enter prompts or routine logs.

Known model IDs retain Pi catalog input capabilities. Text-only models get a clear image-support error before provider dispatch. Custom OpenAI-compatible endpoints receive image blocks using Chat Completions syntax and must support vision; a provider rejection remains a provider failure. Per-call reservations use text bytes plus a conservative allowance of 32,768 input tokens per image, rather than counting base64 as text tokens. Actual usage settles through the existing accounting path. Existing dollar budgets, context limits and cancellation still apply.

Pi downloads remain in memory; no attachment directory or public file endpoint is created. User transcript checkpoints retain supplied image blocks and extracted text for safe resumption, under the existing tenant retention/deletion policy. Source metadata changes invalidate derived memory hashes. No SQL migration or new Telegram update subscription is needed; restart updated API and worker together.

Codex task media

Direct Codex tasks also receive authorized source images as app-server image input blocks, with source IDs and image order identified in reference text. Text/code files and selectable PDF text use the same guarded decoder and media limits as Pi. The worker resolves original sources for both initial requirements and follow-ups; it checks current task access, source retention and bot credential generation before and after downloads and immediately before runner dispatch. Application attempt records retain source references without copying image bytes.

Photo-only and captioned answers use the same task reply/Topic bindings and task selection as text. A photo-only answer retains its empty caption. During intake, its preceding authenticated text supplies the schema-pinned intent evidence; images remain reference data and cannot grant new actions. A standalone image follow-up does not inherit this answer exception. Legacy reviewed issue tasks still accept text issue requirements.

The private runner record retains bounded image content for repairs, authentication pauses and restart recovery. New isolated attempts reload authorized retained sources, including when the old Codex session is unavailable. Runner expiry and task erasure scrub this media with the other task content. Images are not added to the repository checkout or patch. The runner accepts bounded data URLs and uses larger bounded HTTP and JSONL buffers to accommodate base64 media. Update app, worker, supervisor and task image together; no SQL migration is required. Existing cancelled tasks are not restarted.

Local verification

Unit tests exercise real local PDF text extraction, byte-signature image handling, caption commands/mentions, streaming size caps, rejected paths, cancellation, checkpoint restoration, vision reservations and text-only model rejection. Isolated PostgreSQL integration tests cover deduplicated photo-only input, captioned code files, Topics, group routing, denied actors, queued/during-download revocation and recovery after unsupported files. Tests use fake Telegram/model transports. Codex tests cover original screenshot handoff, captioned image documents, photo-only answers, edits, selection between tasks, duplicates, guarded downloads, runner restart/erasure and image payloads larger than the old JSONL limit. The pinned real Codex protocol proof checks image delivery at a local fake Responses endpoint; container smoke checks image forwarding through continuation and repair.