chore(agent): phase roadmap from TODO.md — 4 phases (120–123)

Protocol B append: failed-turn retry (L3–4), git source tokens (L5),
image documents (L6 ingest), chat image questions (L6 chat side).
TODO.md items now live in .agents/phases/todo/ and the file is cleared.

LLM-Generated: true

.agents/phases/todo/122_image_documents/05_rag_display.md

@@ -0,0 +1,24 @@
+# Task 05 — RAG display: image docs in the chat sources + the agent read marker
+
+**Phase:** `122_image_documents` · **Source:** `TODO.md:6` — "Images retrieved by the RAG should be shown in the chat nicely."
+
+## Objective
+When a retrieved/agent-read document is an image, the chat shows it: the sources block renders a compact inline image with its summary as caption/alt, and the agent's `read` tool tells the model it is reading a generated image description.
+
+## Work
+1. `app/rag/retriever.py` / `app/api/chat.py` — the sources frames the chat bubble renders (the SSE `sources`/related-doc frames and the agent-sourced doc list): add an OPTIONAL `image_url` field to the per-doc ref shape — populated (the `/api/documents/{id}/image` path) iff the doc row has `is_image`, absent otherwise (the omission rule — text-doc frames stay byte-identical). The retriever already has the `Document` row; the agent's doc refs (the read-tool results / source list) do too — set it at the frame-build sites (grep for the source-ref construction in both modules; one shared helper `source_ref_with_image(doc, …)` keeps the two sites in lockstep).
+2. `frontend/assets/app.js` — the chat's sources block renderer: when a source ref carries `image_url`, render a compact inline `<img>` (max-height ~96px, `object-fit: contain`, the theme's surface, `alt` + visible caption = the doc summary — the "shown nicely" requirement) in place of / beside the existing doc chip text (keep the title + the existing chip affordance — the image is additive, not a replacement). A failed image load collapses to the plain chip (never a broken-image icon).
+3. `app/rag/agent.py` — the `read` tool's result for an image doc: prefix the description with the marker line `Image document — the text below is a description generated from the image:` (a module constant) so the model reasons about what it is reading; non-image docs' results are byte-identical.
+4. ASSUMPTION: the chat QUESTION side (users submitting images) is phase 123 — this task only covers RETRIEVED images in the answer's sources.
+5. ASSUMPTION: the sources-frame `image_url` is the only new frame field — no doc-id leak beyond what the frame already carries (the path encodes the doc id, same as the content endpoint).
+
+## Testing & Quality
+- Integration: `tests/integration/test_chat_api.py` (extend, task 06) — a mocked grounded answer that includes an image doc in its sources → the SSE frame carries `image_url` for that ref only; a text-only grounding has NO `image_url` key anywhere (byte check).
+- Unit: `tests/unit/test_image_documents.py` (task 06) — the frame-helper (present/absent), the agent marker (image vs non-image result), house-style source assertions: the sources renderer reads `image_url`, sets `alt`, and falls back on image error.
+- E2E: `test_image_documents.py` scenario (task 06) — ask a question the mock LLM grounds on the fixture image doc → the chat sources block shows the inline image with its caption.
+- Coverage: **>90%** on the touched modules.
+
+## Completion Criteria
+- [ ] A chat answer grounded on an image doc shows the inline image + caption in its sources block; text-doc answers render byte-identically to before.
+- [ ] The agent `read` result for an image doc carries the marker; the model sees the description, not raw bytes.
+- [ ] `uv run pytest` green; `uv run ruff check . && uv run pyright` clean.