feat(commit-push): auto-detect cohesive chunks and propose multi-commit splits

[giwankim]

Adds an auto-detect phase to the commit-push skill that groups changed files by inferred Angular scope (skill subdir → top-level dir → root-files bucket) and proposes a multi-commit split via a 3-option AskUserQuestion when 2+ groups are detected, replacing the prior single-commit-per-invocation behavior. Users can override with --single/-1 or natural-language phrases like "commit as one"; the override is acknowledged in the reply, and the per-commit loop creates NEW commits even after pre-commit hook failures (never --amend) so earlier iterations are preserved. The skill now reports all SHAs and warns when the proposed split exceeds 5 groups. Also adds skills/commit-push/evals/evals.json with 7 scenarios (single-group default, two-group split, both override mechanisms, --select no-regression, cap warning, mid-split hook failure); iteration-1 grading on 4 smoke scenarios shows 100% pass rate vs 73% on the snapshot of the prior skill. Existing --select interactive picker and explicit @file arg flows are preserved with no regression.

---

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• New Features

  • Enhanced commit-push workflow with intelligent multi-commit splitting based on file groupings and auto-detection
  • Added secret file filtering across all paths
  • Interactive preview and adjustment options for proposed commit splits

• Tests

  • Added comprehensive 21-scenario evaluation suite for commit-push functionality

• Chores

  • Updated workspace directory ignore patterns

Walkthrough

Compute candidate files (args / interactive select / changed), apply secret-file exclusions, auto-detect or accept overrides for commit groups, run per-group stage→commit loops with hook-fix retries, perform post-loop status normalization and conditional reconciliation commit, then push once and report all created SHAs.

Changes

Cohort / File(s)	Summary
Commit-push skill doc skills/commit-push/SKILL.md	Expanded from a single-stage flow into a scoped, secret-filtered multi-commit pipeline: candidate precedence (@file, --select/-i, changed), secret-file exclusion, commit-count overrides (--single/-1/phrases), group auto-detection & merging (including skills/<name>, .github→ci, top-level/root inference), test re-attribution, interactive split/adjust loop, per-group git restore --staged :/ → stage → commit iterations with hook-fix retries (no amend), post-loop status normalization with conditional reconciliation commit, and aggregated push/reporting of created SHAs.
Evaluation specs skills/commit-push/evals/evals.json	Adds a 21-scenario end-to-end eval suite validating single vs multi-group flows, AskUserQuestion split proposals and iterative adjust flows, override routing (--single, -1, --select, natural-language), grouping heuristics (impl/test merge, .github→ci, root-file inference, scope-name merging), secret-path exclusion behavior, pre-commit hook failure/rescue semantics (no amend on retry), reconciliation commit conditions, snapshot normalization, and assertion that git push runs exactly once with expected reported SHAs.
Workspace ignore .gitignore	Adds skills/commit-push-workspace/ to ignore list so the local workspace directory for the skill is excluded from version control.

Sequence Diagram(s)

sequenceDiagram
  participant User as User
  participant Skill as Commit-Push Skill
  participant Git as Git (index/worktree)
  participant Hooks as Pre-commit Hooks

  rect rgba(135,206,250,0.5)
    User->>Skill: invoke /commit-push (flags / natural language)
    Skill->>Skill: compute candidate set\n(`@file` / --select / changed)
    Skill->>Skill: apply secret exclusions\nand detect commit groups
  end

  rect rgba(144,238,144,0.5)
    alt multi-group detected
      Skill->>User: AskUserQuestion (propose split/combine/adjust)
      User->>Skill: selection/adjust
      loop per-group
        Skill->>Git: git restore --staged :/ (clear index)
        Skill->>Git: stage group's files
        Skill->>Git: git commit (create new SHA)
        Git->>Hooks: run pre-commit hooks
        alt hook fails and fixes
          Hooks->>Git: modify working tree
          Skill->>Git: stage group's files
          Skill->>Git: git commit (new SHA, do not amend)
        end
      end
    else single-group
      Skill->>Git: stage candidate files
      Skill->>Git: git commit (single SHA)
    end
  end

  rect rgba(255,182,193,0.5)
    Skill->>Git: snapshot pre-loop status
    Skill->>Git: compute post-loop delta
    alt delta within original candidate set
      Skill->>Git: create reconciliation commit
    else
      Skill->>User: warn about out-of-scope changes (no auto-stage)
    end
    Skill->>Git: git push (single push)
    Skill->>User: report created SHAs
  end

Loading

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related PRs

• feat(skills): add interactive file selection to commit-push #16: Modifies interactive file-selection and index-reset/staging behavior that this change integrates into candidate precedence and per-group staging loops.
• feat(skills): add install-skills target and commit-push skill #14: Earlier edits to skills/commit-push/* (skill addition and install targets) that this PR extends with multi-commit workflow and evaluation scenarios.

Poem

🐇 I hopped through diffs and grouped each tiny patch,
I staged and nibbled, fixed hooks, then made another batch.
One reconcile, one push, all SHAs in a row,
I dug neat burrows of commits — tidy and aglow.
Hooray — the meadow's clean; now off to nibble more!

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately describes the main feature addition: auto-detection of cohesive file chunks and proposal of multi-commit splits, which is the primary change in this pull request.
Description check	✅ Passed	The description is directly related to the changeset, detailing the auto-detect phase, multi-commit split proposal, override mechanisms, per-commit loop behavior, reporting, and references the evaluation scenarios added.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch giwankim/commit-push-split

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[giwankim]

@codex review

[giwankim]

@coderabbitai — addressed all 3 nitpicks from your CHANGES_REQUESTED review on fa71f1d in c964ea9:

Lines 36–38 — Override phrases too permissive: Dropped bare "as one" and "one commit" from step 2c's match list. The list now requires unambiguous multi-word phrases — "single commit", "as one commit", "as a single commit", "commit this as one", "all together" — which a user has to express as a deliberate directive. Common prose like "commit this as one last sanity check" (contains "as one" substring) or "this should be one commit per scope" (contains "one commit" substring) no longer triggers single-commit mode by accident. The tightened prose explicitly cites those false-positive examples so future readers understand the rationale.

Line 81 — git reset HEAD → git restore --staged: Switched step 5.1 to git restore --staged . per Git 2.23+ guidance, with the legacy git reset HEAD noted parenthetically for older Git. Agreed that the modern form is clearer and avoids the git reset --hard confusion.

Lines 42–46 — Cross-rule scope name collisions: Added a "Cross-rule scope merging" paragraph to step 3 stating that groups are keyed by resolved scope name, not by the rule that produced them. So docs from docs/README.md (rule 3.3) and docs from a root-level README.md (rule 3.4) merge into one docs group / one docs(scope) commit. Same applies to ci collisions across rules 3.2 and 3.4. Eval 17 (new) asserts this end-to-end with the docs/README.md + root README.md case from your example.

LanguageTool nits (skipped, both false positives): the .github/ references are filesystem paths (intentional lowercase, per Git/GitHub Actions convention), not platform-name prose; and the If 1 group → ... If 2+ groups → ... parallel structure is intentional for readability of the routing rules.

Threads/branch: giwankim/commit-push-split at c964ea9. Commit and eval references inline above.

[coderabbitai]

Tip

For best results, initiate chat on the files or code changes.

🧩 Analysis chain

Let me verify the changes in the current state of the file.

---

🏁 Script executed:

#!/bin/bash
# Read the SKILL.md file to verify the three fixes
cat skills/commit-push/SKILL.md

Length of output: 12965

---

@giwankim All three fixes are confirmed in the current SKILL.md:

1. Override phrases (step 2c) ✅ — The phrase list is now tightened to only unambiguous multi-word directives ("single commit", "as one commit", "as a single commit", "commit this as one", "all together"), and the prose explicitly calls out the bare "as one" / "one commit" false-positive examples with explanatory rationale. This is a solid improvement.

2. git restore --staged . (step 5.1) ✅ — Switched to the Git 2.23+ form with the legacy git reset HEAD preserved parenthetically for older Git. The phrasing cleanly distinguishes it from git reset --hard.

3. Cross-rule scope merging (step 3) ✅ — The new paragraph correctly specifies that groups are keyed by resolved scope name, not by producing rule, with concrete docs and ci examples covering both collision cases. Eval 17 providing end-to-end coverage of this is a nice touch.

The LanguageTool skips are reasonable — .github/ is clearly a path literal and the If 1 group → ... If 2+ groups → parallel structure reads well as a routing table.

LGTM — happy to approve once you're ready.

[giwankim]

@CodeRabbit review

[coderabbitai]

✅ Actions performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[coderabbitai]

♻️ Duplicate comments (1)

skills/commit-push/SKILL.md (1)

29-32: ⚠️ Potential issue | 🔴 Critical

Secret filter bypass claim contradicts the absolute safety invariant.

Line 32 states users can "re-include them with --select or explicit args if the exclusion was a false positive." This directly contradicts:

• Line 29's claim that the filter applies "on every path, including when --single or explicit @file args are used"
• The retrieved learning documenting that the filter "cannot be bypassed by any flag" and is "enforced by evals 9 and 18"
• The past review thread where this exact issue was supposedly fixed in commit ae511e9

Based on learnings, the correct behavior is: the filter is an absolute safety invariant with no bypass mechanism. Users hitting false positives have three workarounds: (a) rename the file, (b) adjust filter rules in the skill, or (c) commit manually outside the skill. Re-invoking with --select or explicit args will NOT re-include a filtered file.

🔒 Required fix to match the documented invariant

-   - List any excluded files explicitly in your reply so the user can re-include them with `--select` or explicit args if the exclusion was a false positive.
+   - List any excluded files explicitly in your reply so the user knows what was omitted. If the exclusion was a false positive (e.g., a genuine test fixture whose filename matches the secret pattern), the user's workarounds are (a) rename the file so it no longer matches, (b) adjust the filter rules above if the mis-match is structural, or (c) commit the file manually outside this skill (`git add` + `git commit` + `git push`). Re-invoking `/commit-push` with `--select` or explicit `@file` args will NOT re-include the filtered file — the filter is deliberately absolute as a safety invariant, and no flag overrides it.

Based on learnings: In skills/commit-push/SKILL.md, the secret filter (step 2b) is an absolute safety invariant that cannot be bypassed by any flag, with evals 9 and 18 enforcing this behavior.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@skills/commit-push/SKILL.md` around lines 29 - 32, Update the "b) Secret
filter" section to remove the claim that filtered files can be re-included with
`--select` or explicit args (the phrase "so the user can re-include them with
`--select` or explicit args" must be deleted) and instead state that the secret
filter is an absolute safety invariant (applies on every path and cannot be
bypassed by any flag), reference that this behavior is enforced by evals 9 and
18, and list the three supported workarounds for false positives: rename the
file, adjust the filter rules in the skill, or commit manually outside the
skill; ensure this change is applied to the "b) Secret filter" paragraph and any
nearby text that contradicts the invariant.

🧹 Nitpick comments (1)

skills/commit-push/SKILL.md (1)

1-4: Consider shortening the description field for better readability.

The description on line 3 is over 400 characters and packs multiple concepts (auto-detect, overrides, file scoping, secret filtering, triggers) into a single YAML field. While comprehensive, this density may harm readability when the skill is listed in tooling or parsed by systems with line-length limits.

📝 Suggested refactor to split concerns

Move detailed behavior to the Workflow section and keep the description field focused on the primary use case:

-description: Commit staged/unstaged changes and push to origin with an Angular-style commit message. By default, auto-detects when changes span multiple cohesive scopes and proposes splitting them into multiple commits; user can override with `--single`/`-1` or natural-language phrases like "as one commit" or "single commit". File scoping (explicit `@file` args, `--select` interactive picker) and secret-file filtering apply on every path, including the override. Use when the user says "commit and push", "commit this", "/commit-push", or after completing work that should be committed. Accepts optional file arguments to scope the commit, or interactive selection to pick files from the working tree. Also use when the user wants to choose, select, or pick which changed files to include in a commit.
+description: Commit changed files and push to origin with Angular-style messages. Auto-detects cohesive scopes and proposes multi-commit splits; supports overrides, file scoping, and interactive selection. Use for "commit and push", "commit this", or when choosing files to commit.

Then add a "Key Features" subsection under Workflow that enumerates:

• Auto-detect vs. single-commit modes
• File scoping mechanisms (explicit args, --select, default)
• Secret filtering (always enforced)
• Multi-commit proposal with preview

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@skills/commit-push/SKILL.md` around lines 1 - 4, The description for the
skill "commit-push" is too long and dense; shorten the top-level YAML
description field to a single-sentence summary of the primary use case and move
detailed behavior into a new "Workflow" section (or a "Key Features" subsection)
inside the skill body; specifically, trim the current long description in
SKILL.md under name: commit-push to a concise phrase (e.g., "Commit
staged/unstaged changes with an Angular-style message and push to origin"), then
create bullets in Workflow that enumerate Auto-detect vs --single behavior, file
scoping (explicit `@file` args and --select), secret-file filtering, and
multi-commit proposal/preview so all details are removed from the description
field.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Duplicate comments:
In `@skills/commit-push/SKILL.md`:
- Around line 29-32: Update the "b) Secret filter" section to remove the claim
that filtered files can be re-included with `--select` or explicit args (the
phrase "so the user can re-include them with `--select` or explicit args" must
be deleted) and instead state that the secret filter is an absolute safety
invariant (applies on every path and cannot be bypassed by any flag), reference
that this behavior is enforced by evals 9 and 18, and list the three supported
workarounds for false positives: rename the file, adjust the filter rules in the
skill, or commit manually outside the skill; ensure this change is applied to
the "b) Secret filter" paragraph and any nearby text that contradicts the
invariant.

---

Nitpick comments:
In `@skills/commit-push/SKILL.md`:
- Around line 1-4: The description for the skill "commit-push" is too long and
dense; shorten the top-level YAML description field to a single-sentence summary
of the primary use case and move detailed behavior into a new "Workflow" section
(or a "Key Features" subsection) inside the skill body; specifically, trim the
current long description in SKILL.md under name: commit-push to a concise phrase
(e.g., "Commit staged/unstaged changes with an Angular-style message and push to
origin"), then create bullets in Workflow that enumerate Auto-detect vs --single
behavior, file scoping (explicit `@file` args and --select), secret-file
filtering, and multi-commit proposal/preview so all details are removed from the
description field.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: b0e9ebec-f14c-4dbc-a05b-c8e1666d7b1e

📥 Commits

Reviewing files that changed from the base of the PR and between fdab19e and bfe8122.

📒 Files selected for processing (1)

• skills/commit-push/SKILL.md

📜 Review details

🧰 Additional context used

🧠 Learnings (5)

📓 Common learnings

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 26
File: skills/commit-push/SKILL.md:29-33
Timestamp: 2026-04-19T12:13:30.522Z
Learning: In `skills/commit-push/SKILL.md`, the secret filter (step 2b) is an absolute safety invariant: it cannot be bypassed by any flag (`--single`, `-1`, `--select`, or explicit `file` args). Users who hit a false positive have three workarounds: (a) rename the file so it no longer matches the pattern, (b) adjust the filter rules in the skill, or (c) commit the file manually outside the skill (`git add` + `git commit` + `git push`). Re-invoking `/commit-push` with `--select` or explicit `file` args will NOT re-include a filtered file. This is enforced by evals 9 and 18 in `skills/commit-push/evals/evals.json`.

📚 Learning: 2026-04-19T12:13:30.522Z

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 26
File: skills/commit-push/SKILL.md:29-33
Timestamp: 2026-04-19T12:13:30.522Z
Learning: In `skills/commit-push/SKILL.md`, the secret filter (step 2b) is an absolute safety invariant: it cannot be bypassed by any flag (`--single`, `-1`, `--select`, or explicit `file` args). Users who hit a false positive have three workarounds: (a) rename the file so it no longer matches the pattern, (b) adjust the filter rules in the skill, or (c) commit the file manually outside the skill (`git add` + `git commit` + `git push`). Re-invoking `/commit-push` with `--select` or explicit `file` args will NOT re-include a filtered file. This is enforced by evals 9 and 18 in `skills/commit-push/evals/evals.json`.

Applied to files:

• skills/commit-push/SKILL.md

📚 Learning: 2026-02-23T12:48:28.418Z

Learnt from: CR
Repo: giwankim/my-claude-code PR: 0
File: hooks/claude-notify/AGENTS.md:0-0
Timestamp: 2026-02-23T12:48:28.418Z
Learning: Apply instructions in `AGENTS.md` only within `hooks/claude-notify/` directory, not repository-wide

Applied to files:

• skills/commit-push/SKILL.md

📚 Learning: 2026-02-23T12:48:28.418Z

Learnt from: CR
Repo: giwankim/my-claude-code PR: 0
File: hooks/claude-notify/AGENTS.md:0-0
Timestamp: 2026-02-23T12:48:28.418Z
Learning: Applies to hooks/claude-notify/Tests/required-cases.txt : List every test scenario in `Tests/required-cases.txt` (Case-ID gate)

Applied to files:

• skills/commit-push/SKILL.md

📚 Learning: 2026-02-23T12:48:28.418Z

Learnt from: CR
Repo: giwankim/my-claude-code PR: 0
File: hooks/claude-notify/AGENTS.md:0-0
Timestamp: 2026-02-23T12:48:28.418Z
Learning: Applies to hooks/claude-notify/README.md : Keep `README.md` synchronized with behavior, environment variables, and test changes in hooks/claude-notify

Applied to files:

• skills/commit-push/SKILL.md

🪛 LanguageTool

skills/commit-push/SKILL.md

[uncategorized] ~44-~44: The official name of this software platform is spelled with a capital “H”.
Context: ... 2. CI/infra directory: files under .github/... group as scope ci. This rule fir...

(GITHUB)

---

[uncategorized] ~44-~44: The official name of this software platform is spelled with a capital “H”.
Context: ...neric top-level-directory rule below so .github doesn't end up as a literal directory ...

(GITHUB)

---

[uncategorized] ~48-~48: The official name of this software platform is spelled with a capital “H”.
Context: ...s(scope)commit). Same forci(e.g.,.github/workflows/test.yml` from rule 3.2 plus ...

(GITHUB)

---

[style] ~58-~58: Three successive sentences begin with the same word. Consider rewording the sentence or use a thesaurus to find a synonym.
Context: ...1 group → silently jump to step 5. - If 2+ groups → continue to step 4. 4. **P...

(ENGLISH_WORD_REPEAT_BEGINNING_RULE)

🪛 markdownlint-cli2 (0.22.0)

skills/commit-push/SKILL.md

[warning] 96-96: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 97-97: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 98-98: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 99-99: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 99-99: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 102-102: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 103-103: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 104-104: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 109-109: Spaces inside code span elements

(MD038, no-space-in-code)

[coderabbitai]

Tip

For best results, initiate chat on the files or code changes.

🧩 Analysis chain

@giwankim Both points acknowledged — you're right on both counts.

1. Lines 29–32 — Stale duplicate: Confirmed. My review 4136073328 was scoped to fdab19e..bfe8122 and re-flagged something already resolved in ae511e9. The invariant is correctly documented in the current tip and the learning was already persisted at 2026-04-19T12:13:30.522Z. Resolving that thread now.

2. Lines 1–4 — Dense description: Point well taken. If the frontmatter description is the primary signal the skill-matcher reads for invocation routing, every phrase in there earns its place by covering a distinct trigger pattern. Trimming for human readability would be the wrong optimization target. Deferral noted — I'll add this to long-term memory and resolve the thread.

Let me resolve both threads:

---

🏁 Script executed:

#!/bin/bash
# Find review threads from review 4136073328 on PR 26 to resolve
gh api graphql -f query='
{
  repository(owner: "giwankim", name: "my-claude-code") {
    pullRequest(number: 26) {
      reviewThreads(first: 50) {
        nodes {
          id
          isResolved
          comments(first: 1) {
            nodes {
              author { login }
              body
              url
              databaseId
            }
          }
        }
      }
    }
  }
}' | jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isResolved == false) | select(.comments.nodes[0].author.login == "coderabbitai") | {id: .id, url: .comments.nodes[0].url, body: (.comments.nodes[0].body | .[0:120])}'

Length of output: 261

---

@giwankim Good news — both of those threads (lines 29–32 and lines 1–4) are already resolved in the current state; they don't appear in the open thread set.

The only remaining unresolved thread from me is a separate one: discussion_r3107600181 at line 92, flagging "Hook failure handling lacks retry limits and fallback behavior". That's a distinct concern from the two you addressed above — let me know how you'd like to handle it (address, defer, or dismiss if you disagree).

---

✏️ Learnings added

Learnt from: giwankim
URL: https://github.com/giwankim/my-claude-code/pull/26

Timestamp: 2026-04-19T22:28:43.115Z
Learning: In `skills/commit-push/SKILL.md`, the YAML frontmatter `description` field is intentionally dense (~400 chars) because it is the primary signal read by skill-matcher models when deciding whether to invoke `/commit-push` for a user utterance. Each phrase covers a distinct invocation pattern (e.g., "auto-detect and multi-commit splits", "--single/-1", "as one commit", "file args", "--select", "secret-file filtering"). Trimming for human readability would reduce trigger recall. This density is a deliberate design choice, not a documentation defect.

---

Learnt from: giwankim
URL: https://github.com/giwankim/my-claude-code/pull/26

Timestamp: 2026-04-19T22:28:43.115Z
Learning: In `skills/commit-push/SKILL.md`, the YAML frontmatter `description` field is intentionally dense (~400 chars) because it is the primary signal read by skill-matcher models when deciding whether to invoke `/commit-push` for a user utterance. Each phrase covers a distinct invocation pattern (e.g., "auto-detect and multi-commit splits", "--single/-1", "as one commit", "file args", "--select", "secret-file filtering"). Trimming for human readability would reduce trigger recall. This density is a deliberate design choice, not a documentation defect.

Note: Learnings are effective only in the context of similar code segments. To apply general coding guidelines, please configure review instructions. You can manage existing learnings in the app.

---

🧠 Learnings used

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 26
File: skills/commit-push/SKILL.md:29-33
Timestamp: 2026-04-19T12:13:30.522Z
Learning: In `skills/commit-push/SKILL.md`, the secret filter (step 2b) is an absolute safety invariant: it cannot be bypassed by any flag (`--single`, `-1`, `--select`, or explicit `file` args). Users who hit a false positive have three workarounds: (a) rename the file so it no longer matches the pattern, (b) adjust the filter rules in the skill, or (c) commit the file manually outside the skill (`git add` + `git commit` + `git push`). Re-invoking `/commit-push` with `--select` or explicit `file` args will NOT re-include a filtered file. This is enforced by evals 9 and 18 in `skills/commit-push/evals/evals.json`.

[giwankim]

@CodeRabbit pause

[coderabbitai]

✅ Actions performed

Reviews paused.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[coderabbitai]

Actionable comments posted: 4

♻️ Duplicate comments (1)

skills/commit-push/evals/evals.json (1)

113-123: ⚠️ Potential issue | 🟠 Major

Remove the secret-file “override” expectation.

Line 122 contradicts the skill’s absolute filter rule: excluded files should be reported with the documented workarounds, not described as overrideable. Based on learnings, re-invoking with --select or explicit file args will NOT re-include a filtered file.

Suggested eval wording

-        "The skill's reply lists which files were excluded so the user can override if needed",
+        "The skill's reply lists which files were excluded and explains that no flag re-includes them; false positives require renaming the file, adjusting the filter rules, or committing manually outside the skill",

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@skills/commit-push/evals/evals.json` around lines 113 - 123, Update the eval
for id 9 to remove the incorrect "override" expectation: delete the expectation
that says "The skill's reply lists which files were excluded so the user can
override if needed" and instead ensure the remaining expectations reflect that
excluded files are reported with documented workarounds (e.g., re-invoke with
--select or pass explicit file args) rather than being re-includable; edit the
expectations array in the eval entry (id 9) to remove the contradictory line.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@skills/commit-push/evals/evals.json`:
- Around line 236-237: Update the "expected_output" JSON for Scenario 18 to
remove the incorrect assertion that the secret-filter excluded .env; instead
state that .env started clean so it was not in the changed-file candidate set,
the per-group loop produced the usual skills/foo commit, and after the loop step
6 detects .env as post-loop residue classified as outside the candidate set so
the warn-branch fires — the tool warns the user that .env was touched by a
repo-wide hook, does NOT auto-stage or auto-commit it, and leaves the decision
to the user. Reference the "prompt" and "expected_output" keys in this JSON
object when making the edit.
- Around line 195-203: The scenario's pre-commit hook currently mutates "ALL
tracked .md files", which can make reconciliation non-deterministic if other
tracked markdown exists; update the eval entry (the JSON object containing the
"prompt" string and associated "files"/"expectations") so the hook only targets
the candidate files or explicitly guarantee the scratch repo contains no other
tracked .md files. Concretely, modify the "prompt" text to either (a) change
"ALL tracked .md files" to a precise list like "skills/spring-init/SKILL.md and
hooks/notes.txt" or (b) add a setup line stating "the repo has no other tracked
.md files", or instead add those two files to the eval's "files" fixture to
ensure determinism; keep the "expectations" unchanged except to ensure they
reflect the deterministic reconciliation behavior.

In `@skills/commit-push/SKILL.md`:
- Around line 46-48: The cross-rule `ci` example conflicts with the root-file
scope mapping in rule 3.4 (the list that maps root files to
`build`/`docs`/`chore`), so update the spec so the example is consistent: either
extend rule 3.4 to include root CI config files (e.g., map `.codecov.yml` and
other top-level CI configs to the `ci` scope) or remove the `ci` example from
the cross-rule scope merging paragraph; edit the text where rule 3.4 enumerates
root-file mappings and/or the cross-rule example sentence referencing `ci` to
reflect the chosen fix (look for the phrases "Root files (no directory prefix)"
and "Same for `ci`" to locate the lines to change).
- Line 10: The secret-filtering logic must expand untracked directory entries
reported by git status --porcelain into individual file paths before applying
the secret filter and before staging; update the candidate-file determination so
when an entry like "?? dir/" is encountered you run git ls-files --others
--exclude-standard -- <dir>/ to enumerate its constituent files, apply the
existing step 2b secret filter to each returned path, and then stage only the
explicit file paths in step 5 (do not run git add on the directory pathspec);
ensure this change is applied to the code that builds the candidate set and to
the staging routine so every commitable path (not just directory names) is
individually checked.

---

Duplicate comments:
In `@skills/commit-push/evals/evals.json`:
- Around line 113-123: Update the eval for id 9 to remove the incorrect
"override" expectation: delete the expectation that says "The skill's reply
lists which files were excluded so the user can override if needed" and instead
ensure the remaining expectations reflect that excluded files are reported with
documented workarounds (e.g., re-invoke with --select or pass explicit file
args) rather than being re-includable; edit the expectations array in the eval
entry (id 9) to remove the contradictory line.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 3871653e-6199-460b-ba5d-045b2f5f9a83

📥 Commits

Reviewing files that changed from the base of the PR and between ae511e9 and 7c9e15b.

📒 Files selected for processing (3)

• .gitignore
• skills/commit-push/SKILL.md
• skills/commit-push/evals/evals.json

✅ Files skipped from review due to trivial changes (1)

• .gitignore

📜 Review details

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)

• GitHub Check: integration

🧰 Additional context used

🧠 Learnings (8)

📓 Common learnings

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 26
File: skills/commit-push/SKILL.md:82-93
Timestamp: 2026-04-19T22:41:52.404Z
Learning: In `skills/commit-push/SKILL.md` (PR `#26`), the hook-failure policy in Step 5.4 was tightened in commit 5a27de3: (a) fixer is determined from hook stderr + file types with concrete examples (prettier, ruff/black, gofmt, markdownlint-cli, shfmt); checks with no auto-fix go directly to abort; (b) bounded retry of ≤2 fixer-driven retries (≤3 total git commit attempts per iteration — two retries because eslint --fix can need multiple passes for cascading autofixes); (c) on exhausted retries OR no determinable fixer, the group is marked FAILED and the loop aborts — do NOT attempt remaining groups, do NOT reconcile, do NOT push. Abort was chosen over skip to preserve the atomic-batch invariant (consistent with the secret filter's design philosophy). Steps 6, 7, and 8 all condition on the abort flag. Eval 21 covers the always-fail no-fixer path; eval 7 covers retry-succeeds.

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 0
File: :0-0
Timestamp: 2026-04-19T22:28:43.115Z
Learning: In `skills/commit-push/SKILL.md`, the YAML frontmatter `description` field is intentionally dense (~400 chars) because it is the primary signal read by skill-matcher models when deciding whether to invoke `/commit-push` for a user utterance. Each phrase covers a distinct invocation pattern (e.g., "auto-detect and multi-commit splits", "--single/-1", "as one commit", "file args", "--select", "secret-file filtering"). Trimming for human readability would reduce trigger recall. This density is a deliberate design choice, not a documentation defect.

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 26
File: skills/commit-push/SKILL.md:29-33
Timestamp: 2026-04-19T12:13:30.522Z
Learning: In `skills/commit-push/SKILL.md`, the secret filter (step 2b) is an absolute safety invariant: it cannot be bypassed by any flag (`--single`, `-1`, `--select`, or explicit `file` args). Users who hit a false positive have three workarounds: (a) rename the file so it no longer matches the pattern, (b) adjust the filter rules in the skill, or (c) commit the file manually outside the skill (`git add` + `git commit` + `git push`). Re-invoking `/commit-push` with `--select` or explicit `file` args will NOT re-include a filtered file. This is enforced by evals 9 and 18 in `skills/commit-push/evals/evals.json`.

📚 Learning: 2026-04-19T22:41:52.404Z

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 26
File: skills/commit-push/SKILL.md:82-93
Timestamp: 2026-04-19T22:41:52.404Z
Learning: In `skills/commit-push/SKILL.md` (PR `#26`), the hook-failure policy in Step 5.4 was tightened in commit 5a27de3: (a) fixer is determined from hook stderr + file types with concrete examples (prettier, ruff/black, gofmt, markdownlint-cli, shfmt); checks with no auto-fix go directly to abort; (b) bounded retry of ≤2 fixer-driven retries (≤3 total git commit attempts per iteration — two retries because eslint --fix can need multiple passes for cascading autofixes); (c) on exhausted retries OR no determinable fixer, the group is marked FAILED and the loop aborts — do NOT attempt remaining groups, do NOT reconcile, do NOT push. Abort was chosen over skip to preserve the atomic-batch invariant (consistent with the secret filter's design philosophy). Steps 6, 7, and 8 all condition on the abort flag. Eval 21 covers the always-fail no-fixer path; eval 7 covers retry-succeeds.

Applied to files:

• skills/commit-push/evals/evals.json
• skills/commit-push/SKILL.md

📚 Learning: 2026-04-19T12:13:30.522Z

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 26
File: skills/commit-push/SKILL.md:29-33
Timestamp: 2026-04-19T12:13:30.522Z
Learning: In `skills/commit-push/SKILL.md`, the secret filter (step 2b) is an absolute safety invariant: it cannot be bypassed by any flag (`--single`, `-1`, `--select`, or explicit `file` args). Users who hit a false positive have three workarounds: (a) rename the file so it no longer matches the pattern, (b) adjust the filter rules in the skill, or (c) commit the file manually outside the skill (`git add` + `git commit` + `git push`). Re-invoking `/commit-push` with `--select` or explicit `file` args will NOT re-include a filtered file. This is enforced by evals 9 and 18 in `skills/commit-push/evals/evals.json`.

Applied to files:

• skills/commit-push/evals/evals.json
• skills/commit-push/SKILL.md

📚 Learning: 2026-04-19T22:28:43.115Z

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 0
File: :0-0
Timestamp: 2026-04-19T22:28:43.115Z
Learning: In `skills/commit-push/SKILL.md`, the YAML frontmatter `description` field is intentionally dense (~400 chars) because it is the primary signal read by skill-matcher models when deciding whether to invoke `/commit-push` for a user utterance. Each phrase covers a distinct invocation pattern (e.g., "auto-detect and multi-commit splits", "--single/-1", "as one commit", "file args", "--select", "secret-file filtering"). Trimming for human readability would reduce trigger recall. This density is a deliberate design choice, not a documentation defect.

Applied to files:

• skills/commit-push/evals/evals.json
• skills/commit-push/SKILL.md

📚 Learning: 2026-02-23T12:48:28.418Z

Learnt from: CR
Repo: giwankim/my-claude-code PR: 0
File: hooks/claude-notify/AGENTS.md:0-0
Timestamp: 2026-02-23T12:48:28.418Z
Learning: Applies to hooks/claude-notify/README.md : Keep `README.md` synchronized with behavior, environment variables, and test changes in hooks/claude-notify

Applied to files:

• skills/commit-push/evals/evals.json
• skills/commit-push/SKILL.md

📚 Learning: 2026-03-20T08:18:17.492Z

Learnt from: giwankim
Repo: giwankim/my-claude-code PR: 17
File: skills/spring-init/SKILL.md:74-89
Timestamp: 2026-03-20T08:18:17.492Z
Learning: For the `skills/spring-init/SKILL.md` skill in the `giwankim/my-claude-code` repo: Spring Boot version options listed in Step 3 (`4.0.4`, `3.5.12`, `4.0.5-SNAPSHOT`, `4.1.0-M3`, `4.1.0-SNAPSHOT`) are intentionally hardcoded as a convenience defaults prompt. Always verify Spring Boot milestone version availability against the live Spring Initializr API or the Spring Boot GitHub wiki release notes rather than relying on search results, which can be stale. The skill fetches the live dependency catalog at runtime via `spring init --list`, which is the primary freshness mechanism.

Applied to files:

• skills/commit-push/evals/evals.json

📚 Learning: 2026-02-23T12:48:28.418Z

Learnt from: CR
Repo: giwankim/my-claude-code PR: 0
File: hooks/claude-notify/AGENTS.md:0-0
Timestamp: 2026-02-23T12:48:28.418Z
Learning: Apply instructions in `AGENTS.md` only within `hooks/claude-notify/` directory, not repository-wide

Applied to files:

• skills/commit-push/evals/evals.json
• skills/commit-push/SKILL.md

📚 Learning: 2026-02-23T12:48:28.418Z

Learnt from: CR
Repo: giwankim/my-claude-code PR: 0
File: hooks/claude-notify/AGENTS.md:0-0
Timestamp: 2026-02-23T12:48:28.418Z
Learning: Applies to hooks/claude-notify/Tests/required-cases.txt : List every test scenario in `Tests/required-cases.txt` (Case-ID gate)

Applied to files:

• skills/commit-push/evals/evals.json
• skills/commit-push/SKILL.md

🪛 LanguageTool

skills/commit-push/SKILL.md

[uncategorized] ~44-~44: The official name of this software platform is spelled with a capital “H”.
Context: ... 2. CI/infra directory: files under .github/... group as scope ci. This rule fir...

(GITHUB)

---

[uncategorized] ~44-~44: The official name of this software platform is spelled with a capital “H”.
Context: ...neric top-level-directory rule below so .github doesn't end up as a literal directory ...

(GITHUB)

---

[uncategorized] ~48-~48: The official name of this software platform is spelled with a capital “H”.
Context: ...olved scope name). Same for ci (e.g., .github/workflows/test.yml from rule 3.2 plus ...

(GITHUB)

---

[style] ~58-~58: Three successive sentences begin with the same word. Consider rewording the sentence or use a thesaurus to find a synonym.
Context: ...1 group → silently jump to step 5. - If 2+ groups → continue to step 4. 4. **P...

(ENGLISH_WORD_REPEAT_BEGINNING_RULE)

---

[uncategorized] ~93-~93: Did you mean the formatting language “Markdown” (= proper noun)?
Context: ...wfor Go,markdownlint-cli --fixfor markdown,shfmt -w` for shell. If the hook is a...

(MARKDOWN_NNP)

---

[style] ~95-~95: Three successive sentences begin with the same word. Consider rewording the sentence or use a thesaurus to find a synonym.
Context: ...ps. Do NOT run step 6 (reconciliation). Do NOT run step 7 (git push). Earlier su...

(ENGLISH_WORD_REPEAT_BEGINNING_RULE)

🪛 markdownlint-cli2 (0.22.0)

skills/commit-push/SKILL.md

[warning] 99-99: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 100-100: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 101-101: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 102-102: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 103-103: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 103-103: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 106-106: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 107-107: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 108-108: Spaces inside code span elements

(MD038, no-space-in-code)

---

[warning] 113-113: Spaces inside code span elements

(MD038, no-space-in-code)

[giwankim]

@CodeRabbit pause

[coderabbitai]

✅ Actions performed

Reviews paused.