recall（recall 搜索）

Context Recall

Reconstruct the development story from contextual commit history and present it as a natural briefing.

Argument Detection

Check how recall was invoked:

• No arguments — run the default mode (full branch/session briefing below).
• Bare word (e.g. recall auth) — treat as a scope query. Jump to Scope Query.
• word(word) pattern (e.g. recall rejected(auth)) — treat as an action+scope query. Jump to Action+Scope Query.

Default Mode (no arguments)

Step 1: Detect Branch State

Determine the working state:

CURRENT_BRANCH=$(git branch --show-current)
DEFAULT_BRANCH=$(git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@' || echo "main")

# Find the actual parent branch, not just the repo default.
# Try upstream tracking branch first (fast path).
BASE_BRANCH=$(git rev-parse --abbrev-ref @{upstream} 2>/dev/null | sed 's|^origin/||')

# If no upstream, find the nearest local branch by commit distance.
if [ -z "$BASE_BRANCH" ]; then
    BASE_BRANCH=$(git for-each-ref --format='%(refname:short)' refs/heads/ | while read branch; do
        [ "$branch" = "$CURRENT_BRANCH" ] && continue
        echo "$(git log --oneline "$branch..$CURRENT_BRANCH" 2>/dev/null | wc -l | tr -d ' ') $branch"
    done | sort -n | head -1 | awk '{print $2}')
fi

# Final fallback to default branch.
BASE_BRANCH=${BASE_BRANCH:-$DEFAULT_BRANCH}

UNSTAGED=$(git diff --stat)
STAGED=$(git diff --cached --stat)
BRANCH_COMMITS=$(git log ${BASE_BRANCH}..HEAD --oneline 2>/dev/null | wc -l | tr -d ' ')

DEFAULT_BRANCH identifies the repository's primary branch (for scenario C/D detection). BASE_BRANCH identifies the closest ancestor branch — which may differ when feature branches are created from non-default branches (e.g., a feature branched from develop). All branch-relative queries use BASE_BRANCH.

Step 2: Gather Raw Material

Scenario A — On a feature branch with commits

This is the richest scenario. Gather:

# Contextual action lines from all branch commits
git log ${BASE_BRANCH}..HEAD --format="%H%n%s%n%b%n---COMMIT_END---"

# Unstaged changes (what's in progress right now)
git diff --stat
git diff  # read the actual diff for key changes

# Staged changes
git diff --cached --stat

Scenario B — On a feature branch with no commits yet

# Unstaged and staged changes only
git diff --stat
git diff --cached --stat

# Last few commits on the parent branch for project context
git log ${BASE_BRANCH} -10 --format="%H%n%s%n%b%n---COMMIT_END---"

Scenario C — On the default branch

# Recent commit history with contextual action lines
git log -20 --format="%H%n%s%n%b%n---COMMIT_END---"

Scenario D — On the default branch with uncommitted changes

# Same as C plus uncommitted changes
git log -20 --format="%H%n%s%n%b%n---COMMIT_END---"
git diff --stat
git diff --cached --stat

Step 3: Extract Action Lines

From the gathered commit bodies, extract lines matching:

^(intent|decision|rejected|constraint|learned)\(

Group them by commit (preserve chronological order) and by type (for synthesis).

Step 4: Synthesize Output

Signal density over narrative flow. The output should be compact, scannable, and grounded entirely in what the commits and diffs show. Every line should be actionable information. No fluff, no conversational padding.

For Scenario A (branch with commits):

Output the branch state, then synthesize the contextual action lines into a dense briefing organized by what matters most for continuing work.

Example output:

Branch: feat/google-oauth (4 commits ahead of main, unstaged changes in tests/)

Active intent: Add Google as first social login provider. GitHub and Apple to follow.
Approach: passport.js with /api/auth/callback/:provider convention.
Rejected: auth0-sdk — session model incompatible with redis store.
Constraints:
  - Redis session TTL 24h, tokens must refresh within window
  - Callback routes must follow existing :provider pattern
Learned: passport-google needs explicit offline_access scope for refresh tokens.
In progress: Integration tests for callback handler (unstaged).

When branched from a non-default branch:

Branch: feat/oauth-refresh (2 commits ahead of develop, no uncommitted changes)

Active intent: Implement token refresh for OAuth providers.
Constraint: Must stay compatible with the session store changes on develop.

Priority order:

1. Active intent (what we're building and why)
2. Current approach (decisions made)
3. Rejected approaches (what NOT to re-explore — critical)
4. Constraints (hard boundaries)
5. Learnings (things that save time)
6. In-progress work (unstaged/staged changes)

If intent evolved during the branch (a pivot), show both the original and current intent to make the pivot visible.

For Scenario B (branch with no commits):

Branch: feat/new-feature (0 commits ahead of develop)

No contextual history on this branch yet.
Staged: 2 files (src/auth/provider.ts, src/auth/types.ts)
Unstaged: none

Recent project activity (from develop):
  - Auth: OAuth provider framework merged, Google working
  - Payments: Multi-currency support shipped (EUR, GBP alongside USD)

For Scenario C (default branch, no changes):

Synthesize recent merged work from the last 20 commits. Group by area of activity. Surface any active constraints or learnings that apply broadly.

Recent project activity:

Auth: OAuth provider framework merged. Google working, GitHub and Apple planned.
  - Rejected auth0-sdk (session model incompatible with redis store)
  - Constraint: redis session TTL 24h, tokens must refresh within window

Payments: Multi-currency support shipped (USD, EUR, GBP).
  - Per-transaction currency, not account-level
  - Constraint: Stripe locks currency at PaymentIntent creation
  - Learned: presentment ≠ settlement currency in Stripe

What do you want to work on?

For Scenario D (default branch with uncommitted changes):

Same as Scenario C, with uncommitted changes noted at the top.

When there are no contextual commits at all

If the history contains conventional commits but no contextual action lines, still produce useful output from what exists:

Branch: main (no contextual commits found in recent history)

Recent activity (from commit subjects):
  - auth: 6 commits (OAuth implementation, middleware updates) — 2 days ago
  - payments: 4 commits (currency handling, Stripe integration) — last week
  - tests: 3 commits — last week
  - config: 2 commits

Most recent: feat(auth): implement Google OAuth provider

No explicit intent, decisions, or constraints are captured in commit history.
Using the contextual-commit skill on future commits will surface
the reasoning behind changes that commit subjects alone cannot show.

This is honest about the signal quality while still providing useful orientation. The suggestion to adopt the skill is a natural next step, not a sales pitch.

Guidelines

• Dense over conversational. Every line should carry information. No "Here's what's been happening" or "Let me tell you about."
• Grounded in data. Only report what the action lines, commit subjects, and diffs actually show. Do not infer, speculate, or fill gaps.
• Surface rejections prominently. Rejected approaches are the highest-value signal — they prevent wasted exploration.
• Group by scope when multiple scopes exist. On the default branch with broad history, organize by domain area rather than chronologically.
• End with a prompt. Close with "What do you want to work on?" or similar. Keep it short.
• Scale to the data. If there are 2 contextual commits, the output is 3-4 lines. If there are 20, it's a few grouped paragraphs. Never pad thin data into a long output.

Scope Query (recall <scope>)

Targeted query across full repo history for a given scope. Prefix matching: auth matches auth, auth-tokens, auth-library, etc.

Step 1: Query

git log --all --grep="(${SCOPE}" --format="%H%n%s%n%b%n---COMMIT_END---"

Step 2: Extract

From the gathered commit bodies, extract lines matching:

grep -E "^(intent|decision|rejected|constraint|learned)\(${SCOPE}"

This captures all action lines whose scope starts with the query term.

Step 3: Output

Group by action type, chronological within each group. Show which sub-scopes were found.

Example (recall auth):

Scope: auth (also found: auth-tokens, auth-library)

intent:
  - social login starting with Google, then GitHub and Apple
  - original session approach incompatible with redis cluster setup

decision:
  - passport.js over auth0-sdk for multi-provider flexibility
  - JWT with short expiry + refresh token pattern

rejected:
  - auth0-sdk — locks into their session model, incompatible with redis store
  - auth-sessions — redis cluster doesn't support session stickiness needed by passport sessions

constraint:
  - callback routes must follow /api/auth/callback/:provider pattern per existing convention
  - redis 24h TTL means tokens must refresh within that window

learned:
  - passport-google needs explicit offline_access scope for refresh tokens
  - redis-cluster session affinity requires sticky sessions at load balancer level — too invasive

If no matches are found, say so plainly and suggest checking the scope name.

Action+Scope Query (recall <action>(<scope>))

Query a specific action type for a scope across full repo history.

Step 1: Query

git log --all --grep="${ACTION}(${SCOPE}" --format="%H%n%s%n%b%n---COMMIT_END---"

Step 2: Extract

grep "^${ACTION}(${SCOPE}"

Step 3: Output

Flat chronological list with commit subject for provenance.

Example (recall rejected(auth)):

rejected(auth) across history:

  - auth0-sdk — locks into their session model, incompatible with redis store
    from: feat(auth): implement Google OAuth provider

  - auth-sessions — redis cluster doesn't support session stickiness needed by passport sessions
    from: refactor(auth): switch from session-based to JWT tokens

If no matches are found, say so plainly.

Proactive Usage

Before making a significant decision in any scope area, check rejected and constraint lines for that scope first:

git log --all --grep="rejected(${SCOPE}" --format="%b" | grep "^rejected(${SCOPE}"
git log --all --grep="constraint(${SCOPE}" --format="%b" | grep "^constraint(${SCOPE}"

If a previous commit recorded a rejected line for the approach you're about to propose, surface the rejection reason to the user before re-proposing it. The rejection may still stand, or circumstances may have changed — but the user should make that call with full context, not rediscover the same dead end.

Similarly, check constraint lines before proposing an approach that might violate a known boundary.

This check is lightweight and should become habitual for any scope you're actively working in.