ai_review: Add deep review mode

[jiegan0107]

Use --deep-review with --repo-path to run prompt-backed AI review against a Linux kernel codebase.

[aditya-qti]

Thank you for your contribution!

Can you please remove memory/ from this PR? It makes it very difficult to review. We can consider it a separate effort.

[jiegan0107]

Thank you for your contribution!

Can you please remove memory/ from this PR? It makes it very difficult to review. We can consider it a separate effort.

removed memory and updated to the latest version.

[aditya-qti]

Please take a look at this initial AI review:

Findings

[BUG] Deep-review artifacts are not visible inside the agent container

patchwise/patch_review/ai_review/deep_review.py:382 writes tmp/patch_1_rules.md, tmp/patch_1_diff.txt, tmp/review_runtime_config.json, and related artifacts into the host kernel tree during setup(), after the Docker overlay/container has already been prepared by PatchReview.__init__. The agent is then instructed to read those files via container tools, but they can be absent from /home/patchwise/kernel/tmp/....

I reproduced this with a stubbed DeepReview: docker_manager.read_file("/home/patchwise/kernel/tmp/patch_1_rules.md") returned False.

Suggested fix: generate these artifacts before mounting the overlay, write them through docker_manager into the container tree, or store them in a mounted per-run sandbox.

[BUG] The new prompt bundle fails its own self-check

patchwise/patch_review/ai_review/prompts/scripts/self_check.py:404 runs memory_lint.py, and patchwise/patch_review/ai_review/prompts/scripts/memory_lint.py:131 rejects flat memory files when split layout exists. PR #99 adds refs/memory/{commit-message,dt-bindings,patch-scope,subsystem-specific}.md alongside refs/memory/{active,draft,deprecated}/..., so python3 scripts/self_check.py fails with:

Memory lint failed:
- .../refs/memory: split layout is active; remove legacy flat topic files: commit-message.md, dt-bindings.md, patch-scope.md, subsystem-specific.md

Suggested fix: remove the four legacy flat files or adjust the layout/lint contract.

[CONCERN] DeepReview dirties the user’s kernel checkout

patchwise/patch_review/ai_review/deep_review.py:382 uses <repo>/tmp and never cleans it, leaving generated review artifacts in the target kernel repo. Even if the container-visibility issue is fixed by copying files, this still pollutes git status and risks collisions between concurrent runs.

Suggested fix: use SANDBOX_PATH, output-dir, or a per-run ignored temp directory with cleanup.

[CONCERN] Group selectors may pull in DeepReview unexpectedly

patchwise/patch_review/decorators.py:58 registers DeepReview as an LLM review, so llm_reviews and all_reviews paths in patchwise/patch_review/__init__.py:314 include it. The early CLI validation in patchwise/main.py:73 only checks --deep-review or explicit DeepReview. Non-CLI/shared entry points using those selectors may suddenly require a Linux kernel tree and run the heavier deep-review flow.

Suggested fix: keep DeepReview out of broad groups unless explicitly requested, or ensure every selector path applies the same Linux-tree validation and documents the behavior change.

[jiegan0107]

fixed mentioned issues. Please have a review again.

The pushed skill switched to a new architecture with better performance and less token costs