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/03_image_description.md

@@ -0,0 +1,28 @@
+# Task 03 — Image description: the vision model writes the only embedded text
+
+**Phase:** `122_image_documents` · **Source:** `TODO.md:6` — "the only embedded part of an image will be the summary generated by the model."
+
+## Objective
+`describe_image` generates the image's description with the CHAT model (vision), the description becomes BOTH `Document.content` and `Document.summary` (so the chunk pipeline embeds exactly that text — and only that text), and a failed description fails soft (skip + count + log, sync continues).
+
+## Work
+1. `app/rag/summarizer.py` — new `async def describe_image(llm, data: bytes, mime: str, settings=None) -> str | None` (the summarizer module owns model-text generation; follow the existing summary-call conventions — client, model, timeout, the `summary_max_chars` cap):
+   - ONE chat-model call (`settings.llm_chat_model` — LOCKED A3; the lite summary model is not assumed vision-capable) with messages `[{role: "user", content: [{type: "text", text: <DESCRIBE_PROMPT>}, {type: "image_url", image_url: {url: f"data:{mime};base64,{b64}"}}]}]` — the multimodal content-list shape the OpenAI-compatible API expects.
+   - `DESCRIBE_PROMPT` (a module constant, pinned by a unit test): a faithful, retrieval-oriented description — what is depicted, any visible text/labels/titles, diagram/table structure, salient details; 2–4 sentences of substance (the description is the ONLY retrievable text of the doc, so it must carry the image's meaning).
+   - Return the stripped text capped at `settings.summary_max_chars` (the phase-30 cap — the description IS the summary); return `None` on any client error, empty response, or non-2xx (the caller fails soft). No retries beyond the SDK's own — a description failure must not stall a sync.
+2. `app/rag/importer.py` — wire the task-02 seam: the image branch's `content`/`summary` come from `describe_image` —
+   - description `None` → **skip the doc entirely** (no row, no `image_dir` copy kept — delete the copy if it was made, or make the copy AFTER a successful description so a failure never leaves an orphan), `summary.images_failed += 1`, `logger.warning("import: image description failed source=%s path=%s", source, rel)` — the fail-soft skip (LOCKED A3).
+   - success → `content = description`, then the existing `_store_summary` path (L582) runs with the description as the summary (the `is_summary` position −1 chunk mirrors it — phase-30 behavior, unchanged), and the normal content chunks embed the description (for a short description that is typically ONE content chunk + the summary chunk — the chunker's existing behavior, no special case).
+   - the phase-118 backfill branch (unchanged image doc, `summary is None`) calls the SAME path — a description failure there keeps the doc as-is and logs (no row mutation).
+3. `app/rag/llm.py` — no new client: `describe_image` reuses the existing `llm.chat`-equivalent client the summarizer already uses for text summaries (verify the exact client method name in `app/rag/summarizer.py` and match it — the multimodal payload is a plain `list[dict]` message, so no client change is needed; IF the existing client hard-codes text-only `content: str` typing, extend its signature to accept `content: str | list` — pyright-clean).
+4. ASSUMPTION (A3 re-stated): the CHAT model describes; if the owner's chat model lacks vision, `describe_image` returns `None` (the SDK errors) and every image doc is skipped + logged — honest, visible failure (the `images_failed` counter in the sync log is the signal).
+
+## Testing & Quality
+- Unit: `tests/unit/test_image_documents.py` (task 06) — with a MOCK client: the prompt shape (text part + `image_url` data-URL part, correct model), the cap is applied, whitespace stripped; `None` on mock error / empty string / client exception; the importer's skip path (no row, `images_failed == 1`, warning logged, no orphan copy) and the success path (content == summary == description, `is_summary` chunk present, embedding called with the description text — the ONLY text embedded).
+- Integration: the mock-vision `import_sources` end-to-end (task 06).
+- Coverage: **>90%** on the touched modules.
+
+## Completion Criteria
+- [ ] A fixture PNG through the mock vision client yields a doc whose `content` == `summary` == the description, with its embedding(s) derived from that text only.
+- [ ] A failing mock client skips the doc, bumps `images_failed`, logs, and the sync completes with the other docs indexed.
+- [ ] `uv run pytest` green; `uv run ruff check . && uv run pyright` clean.