Problem

serge only injects a single rules file — REVIEW_RULES_PATH (default .ai/review-rules.md, reviewbot/config.py:401) — into the system prompt under the "REVIEW RULES" block (reviewbot/prompts.py:127-128, via _load_review_rules in reviewbot/reviewer.py:1214).

Repos commonly split their conventions across several files and reference them from review-rules.md. For example, diffusers' .ai/review-rules.md tells the reviewer to also read AGENTS.md, models.md, pipelines.md, modular.md, testing.md, etc. Those files exist in the checkout, but nothing loads them — the model is merely told to read_file them on its own initiative, and there's no step that loads them nor any check that they were read. Loading the extra guides is therefore best-effort.

Observed impact

On the Stable Audio 3 PR (diffusers#14119), the review did not flag that StableAudio3InpaintPipeline inherits directly from StableAudio3Pipeline rather than from DiffusionPipeline, which violates gotcha (4) in .ai/pipelines.md. The review clearly had review-rules.md loaded and browse tools working (it even calls StableAudio3Pipeline "the base" of the inpaint pipeline), yet never connected that to the pipelines.md rule — consistent with pipelines.md not having been read.

Likely contributing bug

Links in review-rules.md are relative to .ai/ (e.g. [pipelines.md](pipelines.md)), but the system prompt never tells the model where the rules file lives — the header just says "from the target repo's default branch" (reviewbot/prompts.py:127). A model that obediently calls read_file("pipelines.md") hits the not a file: pipelines.md error (reviewbot/tools.py:630-631), since read_file resolves paths against the repo root, and may or may not recover by guessing the .ai/ prefix. This makes the "read the sub-guides" instruction fragile in exactly the way observed.

Proposed options (rough order of preference)

1. Resolve markdown links in review-rules.md and inline the referenced .ai files into the prompt so the guides are always loaded rather than fetched on model initiative. (Downside: grows context even when not needed — may want to do this only conditionally.)
2. Cheaper fix: add one line to the prompt stating the rules file's path, e.g. "these rules live at .ai/review-rules.md; relative links resolve against .ai/", so the read-it-yourself flow at least gets the paths right.

To confirm what happened on a given run, the run transcript emits every tool call (emit("tool", ...) at reviewbot/reviewer.py:933), so the Serge web UI / pod logs will show whether read_file was called on .ai/pipelines.md.

Requested by Sayak Paul - Slack thread - Agent trace