[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M79FMXG15K3SFKP8FVYASF8Q":3,"skill-same:01M79FMXG15K3SFKP8FVYASF8Q":49},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":29,"repositoryHtmlUrl":30,"repositoryHomepageUrl":31,"ownerHtmlUrl":32,"ownerBlogUrl":33,"ownerDescription":6,"updatedAt":34,"install":35,"resources":48},"01M79FMXG15K3SFKP8FVYASF8Q","review-pr",null,"Review a Penca PR against project conventions and architecture","penca-io\u002Fpenca","penca-io","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F294652540?v=4","01M79FKD3WVDHQXKBT43KMQFPR","Penca","01M79FKD4JJK7894AKV1R7RHTR",8,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Freview-pr\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: review-pr\ndescription: Review a Penca PR against project conventions and architecture\nargument-hint: \"\u003CPR number or URL>\"\nallowed-tools: Bash Glob Grep Read\n---\n\n# Penca PR Review\n\nReview a pull request against Penca's architecture, conventions, and coding standards.\n\n**PR:** \"$ARGUMENTS\"\n\n## Scope: review-only, no code changes\n\nRead the PR, post structured review comments on GitHub, and enqueue follow-up work as kata tasks. Do not edit source files, even when the user describes the desired fix during the review conversation. User direction like \"I think the right fix is X\" or \"we shouldn't do Y\" is meant to inform updated review comments and queued follow-ups — not to authorize you to apply the fix yourself.\n\nThe deliverable of a review is two things in lockstep:\n- **Inline GitHub PR comments** on specific diff lines (small \u002F local findings).\n- **Kata tasks** under the PR's `cha-NNN` label (the same small \u002F local findings *and* any larger behavior-level findings that warrant a follow-up Linear ticket).\n\nFindings never become Linear comments. Large \u002F behavior-level findings become a kata task whose body says \"file follow-up Linear ticket for X\"; the human actually files the Linear ticket at drain time. The skill does not call `save_issue` to create follow-ups autonomously.\n\nIf the user explicitly asks for code changes mid-review, pause and confirm before touching files; edits go on a separate branch as follow-up work, not as commits to the PR being reviewed.\n\n## Read the conventions up front\n\nBefore reviewing, read `docs\u002Fstyle-guide.md` and `docs\u002Fdevelopment-methodology-guide.md` in full — they are the rubric this review applies, loaded **preemptively**, not on demand. Other references (`docs\u002Falgorithms.md`, `docs\u002Fservices\u002F`, `README.md`) stay on demand.\n\n## What NOT to flag\n\nThese look like findings but are settled non-issues. Don't raise them at any severity — and if a previous round flagged one, mark the thread resolved.\n\n- **Proto wire-compat — pre-release.** Penca isn't shipped to clients yet, so there's no installed-base of generated stubs that could mis-decode a re-typed or renumbered field. Tag reuse with a different wire type, field removal without `reserved` markers, and renumbering after a removal are all fine. Don't gate the change, don't add `reserved`, don't suggest renumbering as a follow-up. (This rule retires once Penca ships; the right place to revisit is in this skill body, not via a one-off review finding.)\n- **\"Phantom delete cleanup at compact\"** as a correctness fix. A tombstone in `delete_log` for a `row_uuid` with no upsert anywhere in the row's history is *not* a separate class of bug — the cross-branch propagation case is the same write-write conflict that CHA-5's serializable-write isolation owns. `MutateData`'s contract is unconditional-tombstone-on-row_uuid; whether the source ever held the row is irrelevant to storage. Compact-time phantom cleanup is timing-dependent (only catches phantoms whose compact races merge) and uneven (doesn't address non-phantom WW conflicts). If a PR proposes this, push back and point at CHA-5.\n- **\"Committed rows + missing files = corruption\"** without checking reader-visibility. Before writing \"this creates zombie committed rows \u002F dangling references \u002F corruption,\" verify the read path. The Penca pattern across persist \u002F snapshot \u002F compact is: **parent-or-collective commit is the atomic boundary.** Per-segment commits inside the write loop are operationally visible (internal queries can see them) but user-invisible (reader queries INNER JOIN the parent's committed flag, so partial state is unreachable). Retry \u002F idempotency via deterministic UUIDs + `ON CONFLICT DO NOTHING` re-creates missing files before the final flip, closing the loop. If the design is atomic via last-step commit, the \"orphan\" rows aren't actually visible to readers — that's not corruption.\n- **Deterministic-UUID tie-breaks** on `ORDER BY ... DESC LIMIT 1` over Penca metadata tables. Before suggesting a tie-break column, trace the table's deterministic-UUID derivation (ADR 0016, `crates\u002Fpenca-core\u002Fsrc\u002Fnaming.rs` \u002F the Python `naming.py`). If the `ORDER BY` column is one of the `row_uuid_for_pk` inputs for that table's PK *and* the `WHERE` filters by the other inputs, two rows with the same sort value would derive the same UUID and collide on PK (`ON CONFLICT DO UPDATE` collapses them) — the tie is structurally impossible and the `LIMIT 1` is already deterministic. Only suggest a tie-break when the ordering column is NOT load-bearing in the PK derivation. (Caught on PR #79: `table_snapshot_uuid` already derives from `snapshotted_at_micros`, so a `, table_snapshot_uuid DESC` tie-break would never fire.)\n\n## Step 0: Check CI status\n\n```bash\ngh pr checks $ARGUMENTS\n```\n\nIf any checks are failing, note which ones and whether the failures are related to the PR changes. If Rust compilation fails, flag it — there's no point reviewing code that doesn't compile.\n\n## Step 1: Gather PR context\n\n`gh pr view` without `--json` hits the deprecated Projects Classic API and errors. Always use `--json` to select specific fields.\n\n```bash\ngh pr view $ARGUMENTS --json title,body,headRefName,baseRefName,files,additions,deletions,changedFiles,commits\ngh pr diff $ARGUMENTS\n```\n\nIf the diff is very large (>5000 lines), pipe through `head -5000` and note that the review is partial. Read the changed files in full (not just the diff) to understand context.\n\n## Step 2: Check conventional commit compliance\n\nVerify all commits follow `\u003Ctype>(\u003Cscope>): \u003Cdescription>`:\n- Valid types: feat, fix, refactor, perf, test, docs, build, chore\n- Scopes must match `linear\u002Flabels.toml`\n- Description: lowercase, imperative mood, no period, under 72 chars\n- `CHA-XX` footer when a ticket exists\n\n```bash\ngh pr view $ARGUMENTS --json commits --jq '.commits[].messageHeadline'\n```\n\n## Step 3: Rust-specific checks (if Rust files changed)\n\nFor any changes under `crates\u002F`, audit against these checklists:\n\n### 3a: Python reference implementation\nPython is the correctness reference; Rust is the production implementation for perf. When reviewing Rust changes to core algorithms (merge-on-read, branch merge, persist, compact, snapshot, storage clients, metadata client):\n- Read the corresponding Python source file to verify the Rust port mirrors the same logic, data flow, and correctness guarantees.\n- Flag any divergence that changes *behavior* (dropping a query, different merge semantics, altered conflict resolution).\n- Do NOT flag divergences that are deliberate Rust improvements (streaming where Python materializes, zero-copy where Python copies, projection pushdown, concurrent execution).\n- Rust-only code (Flight SQL, DataFusion integration, gRPC microservice plumbing, config, deployment) has no Python counterpart — skip this check for those areas.\n\n### 3b: Query efficiency audit\n- Count SQL queries per method in both Python and Rust versions.\n- Flag if the Rust version issues MORE queries than Python (never acceptable).\n- Note FEWER queries as an improvement.\n\n### 3c: Streaming vs materialization\n- Verify unbounded reads use `fetch_stream` pattern, not `fetch_all`.\n- Verify bounded reads (system metadata, small result sets) use materialized fetch.\n- Check cold storage reads are per-segment (each fits in memory).\n\n### 3d: DbDriver abstraction\n- Verify methods use `&impl DbDriver` not `&dyn DbDriver`.\n- Verify optional driver pattern matches Python (`driver = driver or self._driver`).\n- Verify no sqlx types leak through the DbDriver trait boundary (except via associated types).\n\n### 3e: Error handling\n- Custom error types use `thiserror`.\n- No `.unwrap()` or `.expect()` in library code (OK in tests and main.rs).\n- Errors propagate with `?`, not manual match blocks.\n\nFor a focused second-pair-of-eyes pass on a Rust diff, you can invoke `\u002Fcode-quality-reviewer` — it captures the same checklist plus type-idiom audit.\n\n## Step 4: General code quality\n\nAudit checklist (language-agnostic):\n- No unnecessary `.clone()` calls (check if a borrow would work).\n- No heap allocation where stack works (Vec where array suffices, Box where value works).\n- SQL identifiers use dialect quoting functions, never string formatting.\n- No bind parameters contain user-generated data in system table queries.\n- System-generated UUIDs formatted with `format_sql_text_array` or typed UUID params.\n- Intermediate structs feeding proto responses follow the hot-path rule: in Rust, owned-and-moved into the proto (no `.clone()` on payload fields when ownership transfer would do); in Python, no materialized list of intermediate dataclasses just to build a parallel list of protos (go row → proto, or yield protos via a generator for streaming RPCs).\n\n## Step 5: Architecture checks\n\nLightweight architecture checks (greppable patterns):\n- Validation logic lives in server\u002Fservicer layer, not in library\u002Fstorage clients.\n- No intermediate variable aliasing (inline `settings.x.y` at call sites).\n- No method aliasing (inline the full method call, don't assign to short names).\n- Imports at top of file, not inside functions.\n- **Purge-state consumers read the stored watermark, not a derivation.** Code that needs \"the Purge watermark for table T\" (tx_log GC, segment GC, cleanup-pass cutoffs) reads `table_purge_metadata.purged_at_micros(T)` directly — flag any `MAX(persisted_at_micros) ... WHERE now - committed_at > grace` expression standing in for it, even when functionally close. \"Use `purged_at`, not `persisted_at`\" is categorical.\n- **`stream_merged` row filters qualify columns with `l.`.** When a `stream_merged` \u002F `resolve_table_metadata` caller passes a `row_filter`, columns must be `l.`-qualified (e.g. `l.row_uuid = '...'`). All three SQL paths (hot, cold, snapshot) expose the same `l` alias; an unqualified column is ambiguous on the join and any other prefix breaks exactly one path. Canonical pattern: `crates\u002Fpenca-storage-meta\u002Fsrc\u002Ftable.rs::resolve_table_metadata`.\n- **Cold DataFusion reads derive their session from the driver template, never `SessionContext::new()`.** A cold read that builds a `SessionContext` must derive it via `penca_dl::derive_cold_session(&template)` from the driver's `session_template` (`Arc\u003CSessionState>`, held by `QueryManager`). The template carries the shared function registry plus the analyzer\u002Foptimizer rules; a fresh `SessionContext::new()` silently diverges from every other cold read. Derive once per read call and pass `&SessionContext` down into the penca-merge helper (mirrors `output.rs`'s `session: &SessionContext` param). Flag any bare `SessionContext::new()` on a cold path — this is a recurring review finding, most recently `penca_merge::cold_audit_batches` on the `audit_data` path.\n- **penca-api names DataFusion types only through penca-dl re-exports.** penca-api does not depend on the `datafusion` crate — penca-merge owns DataFusion — so `use datafusion::prelude::SessionContext` in penca-api is an `E0433 unlinked crate` error, not a style nit. Use `penca_dl::{SessionState, SessionContext}` \u002F `penca_dl::derive_cold_session`; if a needed type isn't re-exported, add the `pub use` to `crates\u002Fpenca-dl\u002Fsrc\u002Flib.rs` rather than taking a direct dependency.\n\nFor cross-bounded-context architecture concerns — bounded-context boundaries, microservice ownership, ADR alignment, CQRS-shaped splits — invoke `\u002Fsoftware-architect`. It applies a Fowler-principled checklist (single bounded context per service, CQS vs CQRS, TolerantReader, StranglerFig, DesignStaminaHypothesis) against the diff and returns structured findings.\n\n### Commit-message-vs-diff mechanism audit\n\nCommit message's mechanism claim matches the diff's actual code path. If the message names a canonical mechanism (e.g. `stream_merged`, `apply_changes`, `compact_persist_segments`), verify the new code path actually invokes it — not a parallel resolver, helper, or workaround that merely mentions the symbol. This is the canonical responsibility of `\u002Freview-pr`.\n\n## Step 6: Report\n\nOrganize findings into these sections. For each finding, include the file path and line number.\n\n### Critical (must fix before merge)\nIssues that affect correctness, security, or violate architecture boundaries.\n\n### Important (should fix)\nConvention violations, missing patterns, query efficiency regressions.\n\n### Suggestions (nice to have)\nStyle improvements, performance micro-optimizations, documentation.\n\n### Strengths\nWhat the PR does well. Acknowledge good decisions.\n\n## Step 7: Post review to GitHub\n\nPost the review using the GitHub API with inline line-level comments on specific diff lines.\n\n### 7a: Get the latest commit SHA\n\n```bash\ngh pr view $ARGUMENTS --json commits --jq '.commits[-1].oid'\n```\n\n### 7b: Build and submit the review\n\nThe `event` field controls the review type:\n- `\"REQUEST_CHANGES\"` — critical or important issues found (comments must be resolved before merge)\n- `\"COMMENT\"` — suggestions only, no blocking\n- `\"APPROVE\"` — no critical or important issues\n\n**Reviewing a PR this account authored → `event` MUST be `\"COMMENT\"`.** GitHub rejects `APPROVE` and `REQUEST_CHANGES` on your own pull request (\"Can not approve\u002Frequest-changes your own pull request\"), and the `\u002Fdo-issue` `orch:spawn-review` path always hits this: the review subagent shares the author's `gh` credentials, so reviewer and author are the same account regardless of the subagent being fresh-context. Convey severity in the review body and inline comments instead — \"**Critical**: …\" in the comment text, not in the review verdict. The finding mechanics are unaffected: inline comments, `cha-NNN review-pr` kata tasks, and round-2 thread resolution via the GraphQL `resolveReviewThread` mutation all work normally. Only reach for `APPROVE`\u002F`REQUEST_CHANGES` when reviewing someone else's PR.\n\n```bash\ngh api repos\u002F{owner}\u002F{repo}\u002Fpulls\u002F{number}\u002Freviews \\\n  --method POST \\\n  --field commit_id=\"\u003CSHA>\" \\\n  --field event=\"COMMENT\" \\\n  --field body=\"## Penca PR Review Summary\n\n\u003Chigh-level summary of findings>\" \\\n  --field comments='[\n    {\n      \"path\": \"crates\u002Fpenca-storage-cold\u002Fsrc\u002Flib.rs\",\n      \"line\": 42,\n      \"body\": \"**Important**: This helper function may be unnecessary indirection...\"\n    }\n  ]'\n```\n\nFor multi-line comments, use `start_line` and `line` to highlight a range.\n\n### 7c: Comment placement rules\n\n- **Critical and Important** findings: always post as inline review comments on the specific lines. These create resolve-before-merge threads.\n- **Suggestions**: post as inline comments if they target a specific line, otherwise include in the review body summary.\n- **Strengths**: include in the review body summary only.\n\n### 7d: Determine the correct line numbers\n\nThe `line` field refers to the line number in the **new version** of the file (the right side of the diff). Use the diff output from Step 1 to map findings to the correct line numbers. Lines prefixed with `+` in the diff are new lines — use the line number shown in the `@@` hunk header plus the offset within the hunk.\n\n### 7e: Fallback\n\nIf the review has no findings that target specific lines, use `gh pr review` instead:\n\n```bash\ngh pr review $ARGUMENTS --comment --body \"$(cat \u003C\u003C'EOF'\n## Penca PR Review\n\n\u003Cfull review here>\nEOF\n)\"\n```\n\n## Step 8: Enqueue findings as kata tasks\n\nAfter the GitHub review is posted, mirror each finding into the PR's `cha-NNN` kata queue so the drain loop in `\u002Fdo-issue` Step 7 picks them up.\n\n### 8a: Extract `cha-NNN` from the PR\n\nThe PR's `cha-NNN` is the slug used for both the inline-GitHub-comment finding and any follow-up kata tasks. Sources, in order of preference:\n\n```bash\n# Head ref shape: nhobin219\u002Fcha-NN-...\ngh pr view $ARGUMENTS --json headRefName --jq '.headRefName' | grep -oE 'cha-[0-9]+'\n\n# Fallback: PR body's \"Closes CHA-NN\" footer\ngh pr view $ARGUMENTS --json body --jq '.body' | grep -oE 'CHA-[0-9]+' | head -1 | tr 'A-Z' 'a-z'\n```\n\nIf neither shape yields a `cha-NNN`, halt and ask the user — kata tasks must be scoped to a ticket for the drain loop to pick them up.\n\n### 8b: Per-finding kata task\n\nFor each Critical \u002F Important \u002F Suggestion finding, emit one kata task. The task body is the same prose as the inline GitHub comment (so the human reading the kata task has all the context); the `idempotency-key` is content-stable so re-reviews on the same PR don't duplicate. **Findings are auto-approved** (`--label approved`) so the `\u002Fdo-issue` drain picks them up without a human gate — the initial development plan is the only thing that still passes through `plan-draft`:\n\n```bash\nkey=\"pr${PR_NUMBER}:${FINDING_FILE}:${FINDING_LINE}:$(printf '%s' \"$FINDING_TITLE\" | sha256sum | cut -c1-12)\"\ncase \"$SEVERITY\" in\n  critical|important) prio=1 ;;\n  *)                  prio=2 ;;\nesac\nout=$(kata create \\\n  --label \"$CHA\" \\\n  --label review-pr \\\n  --label approved \\\n  --priority \"$prio\" \\\n  --idempotency-key \"$key\" \\\n  --body \"$FINDING_BODY\" \\\n  --json \\\n  -- \"$FINDING_TITLE\")\nnew_ref=$(printf '%s' \"$out\" | jq -r '.qualified_id \u002F\u002F empty')\n\n# Dynamic blocker extension: every still-open \u002Fdo-issue orchestration task\n# under this CHA gains the new finding as a blocker, so PR open \u002F post-open\n# review wait on the finding before unblocking. `kata edit --blocked-by` is\n# additive + idempotent. No-op when \u002Freview-pr is invoked outside \u002Fdo-issue\n# (no orch:* tasks exist).\nif [ -n \"$new_ref\" ]; then\n  for orch in orch:run-cleanup orch:open-pr orch:spawn-review; do\n    orch_ref=$(kata list --label \"$CHA\" --label \"$orch\" --status open --json 2>\u002Fdev\u002Fnull \\\n      | jq -r '.issues[0].qualified_id \u002F\u002F empty')\n    [ -n \"$orch_ref\" ] && kata edit \"$orch_ref\" --blocked-by \"$new_ref\" >\u002Fdev\u002Fnull\n  done\nfi\n```\n\nIdempotency note: `kata create --idempotency-key` errors on fingerprint mismatch (same key, different content). If a re-review adjusts the wording of a finding, the kata create call will fail; that is intentional — the re-reviewer should either re-key (`pr\u003CN>:\u003Cfile>:\u003Cline>:\u003Cnew-hash>`) or update the existing kata task in place via `kata edit`. Don't paper over the conflict.\n\n### 8c: Large \u002F behavior-level findings\n\nIf a finding is too large for inline-comment-resolution (e.g. \"the stream_merged path needs a structural rewrite, not a code-level fix in this PR\"), the kata task body becomes:\n\n> **File follow-up Linear ticket for:** `\u003Cone-sentence summary of the problem>`\n>\n> Context: `\u003Cparagraph or bullets pulled from the inline review comment>`\n\nThe skill does **not** call `save_issue` to create the follow-up Linear ticket. That's the human's call at drain time — they read the kata task, decide if it's worth a Linear ticket, and file it themselves. Mechanism non-goal: no autonomous follow-up ticket creation from the review skill.\n\n### 8d: Strengths\n\n`Strengths` findings (positive callouts in the review body summary) do NOT get kata tasks — they're informational and have no action item.\n\n## Re-reviews\n\nWhen asked to re-review a PR, only review what changed since the last review.\n\n### R1: Get the interdiff\n\n```bash\ngh pr view $ARGUMENTS --json commits --jq '.commits[-1].oid'\ngit diff \u003Clast_reviewed_sha>..\u003Clatest_sha> -- crates\u002F\n```\n\nSave to a temp file if large: `git diff ... > \u002Ftmp\u002FprN_interdiff.txt`.\n\n### R1.5: Fetch author replies before drafting findings\n\nA previous-round finding may have been *addressed by argument* in the comment thread rather than fixed in code. Re-flagging it without reading the reply wastes the author's time and undermines the review.\n\n```bash\ngh api repos\u002F{owner}\u002F{repo}\u002Fpulls\u002F{number}\u002Fcomments\ngh api repos\u002F{owner}\u002F{repo}\u002Fissues\u002F{number}\u002Fcomments\n```\n\nFor each unresolved thread carrying forward, read the author's reply. Treat a substantive reply as either:\n- **Resolution-by-argument** — drop the finding (and resolve the thread per R3 once the argument holds up).\n- **Guidance on what counter-argument to make** — sharpen the next round's finding so it engages with the reply rather than repeating the original.\n\nNever carry a round-1 finding into round 2 without acknowledging the round-1 reply.\n\n### R2: Review only new changes\n\nFocus on:\n- Whether previously flagged issues were addressed.\n- Any new issues introduced by the fix commits.\n- Don't re-flag things that haven't changed.\n\n### R3: Resolve addressed review threads\n\nAfter confirming a previously flagged issue has been fixed, resolve its review thread. Both `Bash(gh api repos\u002F*)` and `Bash(gh api graphql:*)` need to be on the project's `Bash` permission allow list — `.claude\u002Fsettings.local.json` carries them.\n\n```bash\ngh api graphql -f query='\nquery {\n  repository(owner: \"{owner}\", name: \"{repo}\") {\n    pullRequest(number: {number}) {\n      reviewThreads(first: 100) {\n        nodes {\n          id\n          isResolved\n          comments(first: 1) { nodes { body } }\n        }\n      }\n    }\n  }\n}' --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isResolved == false) | {id, comment: .comments.nodes[0].body[:100]}'\n```\n\n```bash\ngh api graphql -f query='\nmutation {\n  resolveReviewThread(input: {threadId: \"\u003CTHREAD_ID>\"}) {\n    thread { id isResolved }\n  }\n}'\n```\n\nResolve threads one at a time after verifying the fix in the interdiff. Do NOT resolve threads whose issues are still present.\n\n### R4: Post the re-review\n\nUse the same Step 7 process. Title the review body `## Penca PR Review — Re-review` (or `Re-review #N` for subsequent rounds). The body should:\n- List each previously flagged issue with ✓ (addressed) or ✗ (still open).\n- Report any new findings from the interdiff.\n- Include strengths if the fix commits are well done.\n","# Penca PR Review\n\nReview a pull request against Penca's architecture, conventions, and coding standards.\n\n**PR:** \"$ARGUMENTS\"\n\n## Scope: review-only, no code changes\n\nRead the PR, post structured review comments on GitHub, and enqueue follow-up work as kata tasks. Do not edit source files, even when the user describes the desired fix during the review conversation. User direction like \"I think the right fix is X\" or \"we shouldn't do Y\" is meant to inform updated review comments and queued follow-ups — not to authorize you to apply the fix yourself.\n\nThe deliverable of a review is two things in lockstep:\n- **Inline GitHub PR comments** on specific diff lines (small \u002F local findings).\n- **Kata tasks** under the PR's `cha-NNN` label (the same small \u002F local findings *and* any larger behavior-level findings that warrant a follow-up Linear ticket).\n\nFindings never become Linear comments. Large \u002F behavior-level findings become a kata task whose body says \"file follow-up Linear ticket for X\"; the human actually files the Linear ticket at drain time. The skill does not call `save_issue` to create follow-ups autonomously.\n\nIf the user explicitly asks for code changes mid-review, pause and confirm before touching files; edits go on a separate branch as follow-up work, not as commits to the PR being reviewed.\n\n## Read the conventions up front\n\nBefore reviewing, read `docs\u002Fstyle-guide.md` and `docs\u002Fdevelopment-methodology-guide.md` in full — they are the rubric this review applies, loaded **preemptively**, not on demand. Other references (`docs\u002Falgorithms.md`, `docs\u002Fservices\u002F`, `README.md`) stay on demand.\n\n## What NOT to flag\n\nThese look like findings but are settled non-issues. Don't raise them at any severity — and if a previous round flagged one, mark the thread resolved.\n\n- **Proto wire-compat — pre-release.** Penca isn't shipped to clients yet, so there's no installed-base of generated stubs that could mis-decode a re-typed or renumbered field. Tag reuse with a different wire type, field removal without `reserved` markers, and renumbering after a removal are all fine. Don't gate the change, don't add `reserved`, don't suggest renumbering as a follow-up. (This rule retires once Penca ships; the right place to revisit is in this skill body, not via a one-off review finding.)\n- **\"Phantom delete cleanup at compact\"** as a correctness fix. A tombstone in `delete_log` for a `row_uuid` with no upsert anywhere in the row's history is *not* a separate class of bug — the cross-branch propagation case is the same write-write conflict that CHA-5's serializable-write isolation owns. `MutateData`'s contract is unconditional-tombstone-on-row_uuid; whether the source ever held the row is irrelevant to storage. Compact-time phantom cleanup is timing-dependent (only catches phantoms whose compact races merge) and uneven (doesn't address non-phantom WW conflicts). If a PR proposes this, push back and point at CHA-5.\n- **\"Committed rows + missing files = corruption\"** without checking reader-visibility. Before writing \"this creates zombie committed rows \u002F dangling references \u002F corruption,\" verify the read path. The Penca pattern across persist \u002F snapshot \u002F compact is: **parent-or-collective commit is the atomic boundary.** Per-segment commits inside the write loop are operationally visible (internal queries can see them) but user-invisible (reader queries INNER JOIN the parent's committed flag, so partial state is unreachable). Retry \u002F idempotency via deterministic UUIDs + `ON CONFLICT DO NOTHING` re-creates missing files before the final flip, closing the loop. If the design is atomic via last-step commit, the \"orphan\" rows aren't actually visible to readers — that's not corruption.\n- **Deterministic-UUID tie-breaks** on `ORDER BY ... DESC LIMIT 1` over Penca metadata tables. Before suggesting a tie-break column, trace the table's deterministic-UUID derivation (ADR 0016, `crates\u002Fpenca-core\u002Fsrc\u002Fnaming.rs` \u002F the Python `naming.py`). If the `ORDER BY` column is one of the `row_uuid_for_pk` inputs for that table's PK *and* the `WHERE` filters by the other inputs, two rows with the same sort value would derive the same UUID and collide on PK (`ON CONFLICT DO UPDATE` collapses them) — the tie is structurally impossible and the `LIMIT 1` is already deterministic. Only suggest a tie-break when the ordering column is NOT load-bearing in the PK derivation. (Caught on PR #79: `table_snapshot_uuid` already derives from `snapshotted_at_micros`, so a `, table_snapshot_uuid DESC` tie-break would never fire.)\n\n## Step 0: Check CI status\n\n```bash\ngh pr checks $ARGUMENTS\n```\n\nIf any checks are failing, note which ones and whether the failures are related to the PR changes. If Rust compilation fails, flag it — there's no point reviewing code that doesn't compile.\n\n## Step 1: Gather PR context\n\n`gh pr view` without `--json` hits the deprecated Projects Classic API and errors. Always use `--json` to select specific fields.\n\n```bash\ngh pr view $ARGUMENTS --json title,body,headRefName,baseRefName,files,additions,deletions,changedFiles,commits\ngh pr diff $ARGUMENTS\n```\n\nIf the diff is very large (>5000 lines), pipe through `head -5000` and note that the review is partial. Read the changed files in full (not just the diff) to understand context.\n\n## Step 2: Check conventional commit compliance\n\nVerify all commits follow `\u003Ctype>(\u003Cscope>): \u003Cdescription>`:\n- Valid types: feat, fix, refactor, perf, test, docs, build, chore\n- Scopes must match `linear\u002Flabels.toml`\n- Description: lowercase, imperative mood, no period, under 72 chars\n- `CHA-XX` footer when a ticket exists\n\n```bash\ngh pr view $ARGUMENTS --json commits --jq '.commits[].messageHeadline'\n```\n\n## Step 3: Rust-specific checks (if Rust files changed)\n\nFor any changes under `crates\u002F`, audit against these checklists:\n\n### 3a: Python reference implementation\nPython is the correctness reference; Rust is the production implementation for perf. When reviewing Rust changes to core algorithms (merge-on-read, branch merge, persist, compact, snapshot, storage clients, metadata client):\n- Read the corresponding Python source file to verify the Rust port mirrors the same logic, data flow, and correctness guarantees.\n- Flag any divergence that changes *behavior* (dropping a query, different merge semantics, altered conflict resolution).\n- Do NOT flag divergences that are deliberate Rust improvements (streaming where Python materializes, zero-copy where Python copies, projection pushdown, concurrent execution).\n- Rust-only code (Flight SQL, DataFusion integration, gRPC microservice plumbing, config, deployment) has no Python counterpart — skip this check for those areas.\n\n### 3b: Query efficiency audit\n- Count SQL queries per method in both Python and Rust versions.\n- Flag if the Rust version issues MORE queries than Python (never acceptable).\n- Note FEWER queries as an improvement.\n\n### 3c: Streaming vs materialization\n- Verify unbounded reads use `fetch_stream` pattern, not `fetch_all`.\n- Verify bounded reads (system metadata, small result sets) use materialized fetch.\n- Check cold storage reads are per-segment (each fits in memory).\n\n### 3d: DbDriver abstraction\n- Verify methods use `&impl DbDriver` not `&dyn DbDriver`.\n- Verify optional driver pattern matches Python (`driver = driver or self._driver`).\n- Verify no sqlx types leak through the DbDriver trait boundary (except via associated types).\n\n### 3e: Error handling\n- Custom error types use `thiserror`.\n- No `.unwrap()` or `.expect()` in library code (OK in tests and main.rs).\n- Errors propagate with `?`, not manual match blocks.\n\nFor a focused second-pair-of-eyes pass on a Rust diff, you can invoke `\u002Fcode-quality-reviewer` — it captures the same checklist plus type-idiom audit.\n\n## Step 4: General code quality\n\nAudit checklist (language-agnostic):\n- No unnecessary `.clone()` calls (check if a borrow would work).\n- No heap allocation where stack works (Vec where array suffices, Box where value works).\n- SQL identifiers use dialect quoting functions, never string formatting.\n- No bind parameters contain user-generated data in system table queries.\n- System-generated UUIDs formatted with `format_sql_text_array` or typed UUID params.\n- Intermediate structs feeding proto responses follow the hot-path rule: in Rust, owned-and-moved into the proto (no `.clone()` on payload fields when ownership transfer would do); in Python, no materialized list of intermediate dataclasses just to build a parallel list of protos (go row → proto, or yield protos via a generator for streaming RPCs).\n\n## Step 5: Architecture checks\n\nLightweight architecture checks (greppable patterns):\n- Validation logic lives in server\u002Fservicer layer, not in library\u002Fstorage clients.\n- No intermediate variable aliasing (inline `settings.x.y` at call sites).\n- No method aliasing (inline the full method call, don't assign to short names).\n- Imports at top of file, not inside functions.\n- **Purge-state consumers read the stored watermark, not a derivation.** Code that needs \"the Purge watermark for table T\" (tx_log GC, segment GC, cleanup-pass cutoffs) reads `table_purge_metadata.purged_at_micros(T)` directly — flag any `MAX(persisted_at_micros) ... WHERE now - committed_at > grace` expression standing in for it, even when functionally close. \"Use `purged_at`, not `persisted_at`\" is categorical.\n- **`stream_merged` row filters qualify columns with `l.`.** When a `stream_merged` \u002F `resolve_table_metadata` caller passes a `row_filter`, columns must be `l.`-qualified (e.g. `l.row_uuid = '...'`). All three SQL paths (hot, cold, snapshot) expose the same `l` alias; an unqualified column is ambiguous on the join and any other prefix breaks exactly one path. Canonical pattern: `crates\u002Fpenca-storage-meta\u002Fsrc\u002Ftable.rs::resolve_table_metadata`.\n- **Cold DataFusion reads derive their session from the driver template, never `SessionContext::new()`.** A cold read that builds a `SessionContext` must derive it via `penca_dl::derive_cold_session(&template)` from the driver's `session_template` (`Arc\u003CSessionState>`, held by `QueryManager`). The template carries the shared function registry plus the analyzer\u002Foptimizer rules; a fresh `SessionContext::new()` silently diverges from every other cold read. Derive once per read call and pass `&SessionContext` down into the penca-merge helper (mirrors `output.rs`'s `session: &SessionContext` param). Flag any bare `SessionContext::new()` on a cold path — this is a recurring review finding, most recently `penca_merge::cold_audit_batches` on the `audit_data` path.\n- **penca-api names DataFusion types only through penca-dl re-exports.** penca-api does not depend on the `datafusion` crate — penca-merge owns DataFusion — so `use datafusion::prelude::SessionContext` in penca-api is an `E0433 unlinked crate` error, not a style nit. Use `penca_dl::{SessionState, SessionContext}` \u002F `penca_dl::derive_cold_session`; if a needed type isn't re-exported, add the `pub use` to `crates\u002Fpenca-dl\u002Fsrc\u002Flib.rs` rather than taking a direct dependency.\n\nFor cross-bounded-context architecture concerns — bounded-context boundaries, microservice ownership, ADR alignment, CQRS-shaped splits — invoke `\u002Fsoftware-architect`. It applies a Fowler-principled checklist (single bounded context per service, CQS vs CQRS, TolerantReader, StranglerFig, DesignStaminaHypothesis) against the diff and returns structured findings.\n\n### Commit-message-vs-diff mechanism audit\n\nCommit message's mechanism claim matches the diff's actual code path. If the message names a canonical mechanism (e.g. `stream_merged`, `apply_changes`, `compact_persist_segments`), verify the new code path actually invokes it — not a parallel resolver, helper, or workaround that merely mentions the symbol. This is the canonical responsibility of `\u002Freview-pr`.\n\n## Step 6: Report\n\nOrganize findings into these sections. For each finding, include the file path and line number.\n\n### Critical (must fix before merge)\nIssues that affect correctness, security, or violate architecture boundaries.\n\n### Important (should fix)\nConvention violations, missing patterns, query efficiency regressions.\n\n### Suggestions (nice to have)\nStyle improvements, performance micro-optimizations, documentation.\n\n### Strengths\nWhat the PR does well. Acknowledge good decisions.\n\n## Step 7: Post review to GitHub\n\nPost the review using the GitHub API with inline line-level comments on specific diff lines.\n\n### 7a: Get the latest commit SHA\n\n```bash\ngh pr view $ARGUMENTS --json commits --jq '.commits[-1].oid'\n```\n\n### 7b: Build and submit the review\n\nThe `event` field controls the review type:\n- `\"REQUEST_CHANGES\"` — critical or important issues found (comments must be resolved before merge)\n- `\"COMMENT\"` — suggestions only, no blocking\n- `\"APPROVE\"` — no critical or important issues\n\n**Reviewing a PR this account authored → `event` MUST be `\"COMMENT\"`.** GitHub rejects `APPROVE` and `REQUEST_CHANGES` on your own pull request (\"Can not approve\u002Frequest-changes your own pull request\"), and the `\u002Fdo-issue` `orch:spawn-review` path always hits this: the review subagent shares the author's `gh` credentials, so reviewer and author are the same account regardless of the subagent being fresh-context. Convey severity in the review body and inline comments instead — \"**Critical**: …\" in the comment text, not in the review verdict. The finding mechanics are unaffected: inline comments, `cha-NNN review-pr` kata tasks, and round-2 thread resolution via the GraphQL `resolveReviewThread` mutation all work normally. Only reach for `APPROVE`\u002F`REQUEST_CHANGES` when reviewing someone else's PR.\n\n```bash\ngh api repos\u002F{owner}\u002F{repo}\u002Fpulls\u002F{number}\u002Freviews \\\n  --method POST \\\n  --field commit_id=\"\u003CSHA>\" \\\n  --field event=\"COMMENT\" \\\n  --field body=\"## Penca PR Review Summary\n\n\u003Chigh-level summary of findings>\" \\\n  --field comments='[\n    {\n      \"path\": \"crates\u002Fpenca-storage-cold\u002Fsrc\u002Flib.rs\",\n      \"line\": 42,\n      \"body\": \"**Important**: This helper function may be unnecessary indirection...\"\n    }\n  ]'\n```\n\nFor multi-line comments, use `start_line` and `line` to highlight a range.\n\n### 7c: Comment placement rules\n\n- **Critical and Important** findings: always post as inline review comments on the specific lines. These create resolve-before-merge threads.\n- **Suggestions**: post as inline comments if they target a specific line, otherwise include in the review body summary.\n- **Strengths**: include in the review body summary only.\n\n### 7d: Determine the correct line numbers\n\nThe `line` field refers to the line number in the **new version** of the file (the right side of the diff). Use the diff output from Step 1 to map findings to the correct line numbers. Lines prefixed with `+` in the diff are new lines — use the line number shown in the `@@` hunk header plus the offset within the hunk.\n\n### 7e: Fallback\n\nIf the review has no findings that target specific lines, use `gh pr review` instead:\n\n```bash\ngh pr review $ARGUMENTS --comment --body \"$(cat \u003C\u003C'EOF'\n## Penca PR Review\n\n\u003Cfull review here>\nEOF\n)\"\n```\n\n## Step 8: Enqueue findings as kata tasks\n\nAfter the GitHub review is posted, mirror each finding into the PR's `cha-NNN` kata queue so the drain loop in `\u002Fdo-issue` Step 7 picks them up.\n\n### 8a: Extract `cha-NNN` from the PR\n\nThe PR's `cha-NNN` is the slug used for both the inline-GitHub-comment finding and any follow-up kata tasks. Sources, in order of preference:\n\n```bash\n# Head ref shape: nhobin219\u002Fcha-NN-...\ngh pr view $ARGUMENTS --json headRefName --jq '.headRefName' | grep -oE 'cha-[0-9]+'\n\n# Fallback: PR body's \"Closes CHA-NN\" footer\ngh pr view $ARGUMENTS --json body --jq '.body' | grep -oE 'CHA-[0-9]+' | head -1 | tr 'A-Z' 'a-z'\n```\n\nIf neither shape yields a `cha-NNN`, halt and ask the user — kata tasks must be scoped to a ticket for the drain loop to pick them up.\n\n### 8b: Per-finding kata task\n\nFor each Critical \u002F Important \u002F Suggestion finding, emit one kata task. The task body is the same prose as the inline GitHub comment (so the human reading the kata task has all the context); the `idempotency-key` is content-stable so re-reviews on the same PR don't duplicate. **Findings are auto-approved** (`--label approved`) so the `\u002Fdo-issue` drain picks them up without a human gate — the initial development plan is the only thing that still passes through `plan-draft`:\n\n```bash\nkey=\"pr${PR_NUMBER}:${FINDING_FILE}:${FINDING_LINE}:$(printf '%s' \"$FINDING_TITLE\" | sha256sum | cut -c1-12)\"\ncase \"$SEVERITY\" in\n  critical|important) prio=1 ;;\n  *)                  prio=2 ;;\nesac\nout=$(kata create \\\n  --label \"$CHA\" \\\n  --label review-pr \\\n  --label approved \\\n  --priority \"$prio\" \\\n  --idempotency-key \"$key\" \\\n  --body \"$FINDING_BODY\" \\\n  --json \\\n  -- \"$FINDING_TITLE\")\nnew_ref=$(printf '%s' \"$out\" | jq -r '.qualified_id \u002F\u002F empty')\n\n# Dynamic blocker extension: every still-open \u002Fdo-issue orchestration task\n# under this CHA gains the new finding as a blocker, so PR open \u002F post-open\n# review wait on the finding before unblocking. `kata edit --blocked-by` is\n# additive + idempotent. No-op when \u002Freview-pr is invoked outside \u002Fdo-issue\n# (no orch:* tasks exist).\nif [ -n \"$new_ref\" ]; then\n  for orch in orch:run-cleanup orch:open-pr orch:spawn-review; do\n    orch_ref=$(kata list --label \"$CHA\" --label \"$orch\" --status open --json 2>\u002Fdev\u002Fnull \\\n      | jq -r '.issues[0].qualified_id \u002F\u002F empty')\n    [ -n \"$orch_ref\" ] && kata edit \"$orch_ref\" --blocked-by \"$new_ref\" >\u002Fdev\u002Fnull\n  done\nfi\n```\n\nIdempotency note: `kata create --idempotency-key` errors on fingerprint mismatch (same key, different content). If a re-review adjusts the wording of a finding, the kata create call will fail; that is intentional — the re-reviewer should either re-key (`pr\u003CN>:\u003Cfile>:\u003Cline>:\u003Cnew-hash>`) or update the existing kata task in place via `kata edit`. Don't paper over the conflict.\n\n### 8c: Large \u002F behavior-level findings\n\nIf a finding is too large for inline-comment-resolution (e.g. \"the stream_merged path needs a structural rewrite, not a code-level fix in this PR\"), the kata task body becomes:\n\n> **File follow-up Linear ticket for:** `\u003Cone-sentence summary of the problem>`\n>\n> Context: `\u003Cparagraph or bullets pulled from the inline review comment>`\n\nThe skill does **not** call `save_issue` to create the follow-up Linear ticket. That's the human's call at drain time — they read the kata task, decide if it's worth a Linear ticket, and file it themselves. Mechanism non-goal: no autonomous follow-up ticket creation from the review skill.\n\n### 8d: Strengths\n\n`Strengths` findings (positive callouts in the review body summary) do NOT get kata tasks — they're informational and have no action item.\n\n## Re-reviews\n\nWhen asked to re-review a PR, only review what changed since the last review.\n\n### R1: Get the interdiff\n\n```bash\ngh pr view $ARGUMENTS --json commits --jq '.commits[-1].oid'\ngit diff \u003Clast_reviewed_sha>..\u003Clatest_sha> -- crates\u002F\n```\n\nSave to a temp file if large: `git diff ... > \u002Ftmp\u002FprN_interdiff.txt`.\n\n### R1.5: Fetch author replies before drafting findings\n\nA previous-round finding may have been *addressed by argument* in the comment thread rather than fixed in code. Re-flagging it without reading the reply wastes the author's time and undermines the review.\n\n```bash\ngh api repos\u002F{owner}\u002F{repo}\u002Fpulls\u002F{number}\u002Fcomments\ngh api repos\u002F{owner}\u002F{repo}\u002Fissues\u002F{number}\u002Fcomments\n```\n\nFor each unresolved thread carrying forward, read the author's reply. Treat a substantive reply as either:\n- **Resolution-by-argument** — drop the finding (and resolve the thread per R3 once the argument holds up).\n- **Guidance on what counter-argument to make** — sharpen the next round's finding so it engages with the reply rather than repeating the original.\n\nNever carry a round-1 finding into round 2 without acknowledging the round-1 reply.\n\n### R2: Review only new changes\n\nFocus on:\n- Whether previously flagged issues were addressed.\n- Any new issues introduced by the fix commits.\n- Don't re-flag things that haven't changed.\n\n### R3: Resolve addressed review threads\n\nAfter confirming a previously flagged issue has been fixed, resolve its review thread. Both `Bash(gh api repos\u002F*)` and `Bash(gh api graphql:*)` need to be on the project's `Bash` permission allow list — `.claude\u002Fsettings.local.json` carries them.\n\n```bash\ngh api graphql -f query='\nquery {\n  repository(owner: \"{owner}\", name: \"{repo}\") {\n    pullRequest(number: {number}) {\n      reviewThreads(first: 100) {\n        nodes {\n          id\n          isResolved\n          comments(first: 1) { nodes { body } }\n        }\n      }\n    }\n  }\n}' --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isResolved == false) | {id, comment: .comments.nodes[0].body[:100]}'\n```\n\n```bash\ngh api graphql -f query='\nmutation {\n  resolveReviewThread(input: {threadId: \"\u003CTHREAD_ID>\"}) {\n    thread { id isResolved }\n  }\n}'\n```\n\nResolve threads one at a time after verifying the fix in the interdiff. Do NOT resolve threads whose issues are still present.\n\n### R4: Post the re-review\n\nUse the same Step 7 process. Title the review body `## Penca PR Review — Re-review` (or `Re-review #N` for subsequent rounds). The body should:\n- List each previously flagged issue with ✓ (addressed) or ✗ (still open).\n- Report any new findings from the interdiff.\n- Include strengths if the fix commits are well done.\n",{"name":5,"description":7,"allowed-tools":24,"argument-hint":25},"Bash Glob Grep Read","\u003CPR number or URL>","main","5857ef2d015e3d9011c5e73b01660da6043e0f0e",".claude\u002Fskills\u002Freview-pr",0,"https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca","https:\u002F\u002Fpenca.io\u002F","https:\u002F\u002Fgithub.com\u002Fpenca-io","https:\u002F\u002Fpenca.io","2026-08-01 08:29:37",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add penca-io\u002Fpenca --skill review-pr",{"tool":41,"command":42},"bunx","bunx skills add penca-io\u002Fpenca --skill review-pr",{"tool":44,"command":45},"pnpm","pnpm dlx skills add penca-io\u002Fpenca --skill review-pr","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Freview-pr\u002FSKILL.md",true,[],[50,57,64,71,78,85],{"skillNo":51,"name":52,"nameZh":6,"description":53,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":19},"01M79FNA3MSFZJJEY8K32KJ7VG","tracing-instrument","Add structured `tracing`-crate instrumentation (spans + events + fields) to Penca Rust code — one combined plan emitted as kata tasks under `cha-NNN plan-draft` and gated once by `\u002Fplan-reviewer`, then one branch carrying every site as its own commit drained from the approved kata queue, gated once by `just check` at PR open. Scope is the `tracing` crate only (logs-as-events + spans); metrics and OpenTelemetry are out of scope. Use for observability tickets, ad-hoc instrumentation passes on a crate, or as the `\u002Fdo-issue` `orch:run-cleanup` cleanup-context pass.","https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Ftracing-instrument\u002FSKILL.md",[],[],{"skillNo":58,"name":59,"nameZh":6,"description":60,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":19},"01M79FN8FWKKYX3PG3MPHFBMSP","test","Run the test suite with pytest","https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Ftest\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":19},"01M79FN4G6C2KG7D737QDQCJFM","software-architect","High-level architecture analysis for Penca design questions — bounded contexts, microservice boundaries, cross-service interfaces, storage tier separations (hot vs cold), CQRS-shaped splits, design-stamina trade-offs, and incremental-replacement strategies. Use when a Linear ticket needs design framing, when an existing plan has a structural concern, or when a proposed change crosses bounded-context boundaries. Advisory — does not edit code.","https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fsoftware-architect\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M79FN1Y86PSSG0A6A4QZSVKW","save-memory","Save or update a memory entry. Always targets the **current repo root's** `.claude\u002Fmemory\u002F`, resolved via `git rev-parse --show-toplevel`; never the runtime `~\u002F.claude\u002Fprojects\u002F\u003Cslug>\u002Fmemory\u002F` path.","https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fsave-memory\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M79FMQVW8TEK002D5SY7N22Z","plan-reviewer","Audit a Penca `\u002Fdo-issue` plan — a set of kata tasks under `cha-NNN plan-draft` — for the mechanism-bound rule + dependency-graph sanity. Use after `\u002Fdo-issue` Step 2 emits the task set and before the user flips `plan-draft` → `approved`. Checks that every task names symbols\u002Ffiles to change AND symbols the new code path must invoke, has mechanism non-goals where relevant, avoids parallel-path workarounds, and that every implementation task is `--blocked-by` at least one red-test task with no cycles in the graph.","https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fplan-reviewer\u002FSKILL.md",[],[],{"skillNo":86,"name":87,"nameZh":6,"description":88,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":19},"01M79FMJW8GZX5SK0GBFTXTJFK","perf-engineer","Algorithmic and profile-driven performance analysis on Penca Rust changes — query-count audits, materialize-vs-stream regressions, allocation patterns under benchmark, hot-path code review backed by samply or cargo bench output. Use post-implementation for profiling, or during planning to flag perf risks. Routing test — if it can be spotted from reading the diff alone, it belongs to code-quality-reviewer; if it needs measurement to know it matters, use this skill.","https:\u002F\u002Fgithub.com\u002Fpenca-io\u002Fpenca\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fperf-engineer\u002FSKILL.md",[],[]]