[khezen]

What

codespy review-path <repo-url> [--path P] [--ref BRANCH] reviews existing code with no PR. It sparse-clones a path, presents every reviewable file as newly added, identifies all scopes under it, and runs the full pipeline. There is no PR to comment on, so the report is delivered as a new PR/MR opened by codespy:

• branch codespy/review/<path-slug>, file .codespy/reviews/<path-slug>.md (full report)
• the report in the PR description (lowest-severity issues are cut first to stay under the size limit, with a link to the file); every issue links to its lines at the reviewed commit
• re-runs update the open report PR (new commit, new description, one comment); a stale branch with no open PR is reset to base
• --no-open-pr is a dry run (stdout only)

Cost caps review.full_path.max_files (300, --max-files) and max_bytes (2 MB) count files left after all skip rules, and stop the run before any LLM call. Symlinks in the reviewed repo are never followed. make review-path REPO_URL=... [REVIEW_PATH=] [REF=] [MAX_FILES=] [OPEN_PR=false] runs it in Docker.

Implementation planner (review.plan, opt-in)

A new agent turns the review's findings into an ordered implementation plan, for both PR reviews and full-path reviews. Disabled by default (review.plan.enabled, REVIEW_PLAN_ENABLED, Action input plan-enabled).

• Runs after the audit and gets its quality assessment and recommendation.
• One RLM run per scope with issues, concurrently (review.max_concurrent_scopes), with filesystem, ripgrep and tree-sitter tools rooted at the scope to find the call sites and tests a fix touches. Each issue goes to the scope with the longest matching subroot (path-segment aware); issues outside every scope are planned against the repo root.
• Plans every issue left after confidence filtering and rejected-feedback suppression; skipped with no LLM call when there are none.
• Read-only memory: uses the run-level recall and recall_memory, writes no episode.
• Soft-fail: a failing scope is left out; a planner failure posts the review without a plan.
• Steps list title, files, change, verification and the issues they address.
• Rendered as a collapsible 🛠️ Implementation Plan section after the recommendation:
  • PR/MR review body: capped at 20,000 chars (memories get what is left);
  • review-path report PR description: same cap, with a link to the full plan in the report file;
  • report file, stdout and MCP markdown: uncapped; JSON: implementation_plan field.
  • Several scopes are nested in their own collapsible sections; truncation drops trailing steps and keeps tags balanced.

Fixes that affect every review

• Sparse checkout selected the whole repo. The pattern list started with a bare /*, which in non-cone mode matches every path, and the list was re-sorted, so ! lines landed before /* and had no effect. Patterns are now ordered cone-style (/*, !/*/, /A/, !/A/*/ …) and applied with git read-tree -mu HEAD, which also shrinks old cached clones. New patterns are written before checkout so treeless clones only fetch the new set.
• Missing Deno → empty APPROVE. Every RLM agent soft-failed to "no findings" without the Deno sandbox. Reviews now check Deno 2.x first and stop with install guidance.

PullRequest.url keeps its existing values for PR/local reviews (it's the memory topic id); only path mode overrides it.

Not included

• MCP review_path tool: not requested; it would skip the report PR and isn't safe under concurrent MCP calls.
• Action (action.yml) support for review-path (the planner's Action inputs are included).

Testing

• tests/test_full_path_review.py (62 tests): real-git sparse checkout (bare /* reproduces the bug; PR mode, path mode, Xcode bundles, narrowing a full clone), URL parsing/rejection, synthetic patches and skip rules, symlinks, caps after skips, scope resolution in path mode, report rendering/truncation, GitHub/GitLab upsert with mocks (create / update open PR / reset stale branch / cleanup only own branch), Deno preflight.
• tests/test_implementation_plan.py (28 tests): config default and env, issue→scope grouping and path handling, gating (no LLM/MCP call when disabled or no issues), per-scope failure isolation, planner inputs (scope-relative issues, audit verdict) and repo-relative output paths, pipeline soft-fail, rendering (flat/nested, truncation within budget with balanced tags and overflow link), both reporters under the 60,000-char body limit with a huge plan, many issues and memories.
• Full suite: 816 passed; the 23 failures and 3 errors are the same set that fails on the base commit (memory/cerebral, doc_extractor, memories, statistics, an outdated test_not_a_signature). Lint and mypy: no new errors beyond the dspy.Signature/Module subclass errors every agent has.
• Manual: Docker image builds with review-path and Deno 2.9.5. A live clone of src/codespy/tools/git of this repo materialised only that path plus ancestor files. A first live run on peaks-com/mono opened a report PR but found nothing because Deno was missing locally, which led to the preflight fix. A full LLM run that opens a PR has not been verified end-to-end yet, and the planner has not been run against a real model.