Recap

/recap — Session Orientation & Awareness

Goal: Orient yourself fast. Rich context by default. Mid-session awareness with --now.

Usage

/recap           # Rich: retro summary, handoff, tracks, git
/recap --quick   # Minimal: git + focus only, no file reads
/recap --now     # Mid-session: timeline + jumps from AI memory
/recap --now deep # Mid-session: + handoff + tracks + connections

DEFAULT MODE (Rich)

Run the rich script, then add suggestions:

bun ~/.claude/skills/recap/recap-rich.ts

Script reads retro summaries, handoff content, tracks, git state. Then LLM adds:

• What's next? (2-3 options based on context)

Step 1.5: Detect INCUBATED_BY (#229)

The recap-rich.ts script auto-detects .claude/INCUBATED_BY breadcrumbs. If present, shows:

## ⚠️ INCUBATED REPO
oracle: mawui-oracle
date: 2026-04-13
source: https://github.com/...

This tells the oracle: "You are in a repo tracked by another oracle. Check the breadcrumb for context."

Step 2: Git context

git status --short
git log --oneline -1

Check what's appropriate from git status:

• Uncommitted changes? → show them, suggest commit or stash
• On a branch (not main)? → git log main..HEAD --oneline to see branch work
• Branch ahead of remote? → suggest push or PR
• Clean on main? → just show last commit, move on

Only read what matters — don't dump 10 commits if status is clean.

Step 2.5: Skill shelf staleness

Skills are copied into ~/.claude/skills/ (fork model), so fixes never propagate on their own — an oracle can run a superseded skill indefinitely with nothing to signal it. This surfaces it during orientation, where you'd notice.

bun "$(dirname "$0")/skills-staleness.ts" 2>/dev/null || true

Silent when the shelf is current. Prints only when you are behind, the install is 30+ days old, the manifest can't be reconciled against disk, or --verbose. Never blocks — any failure exits 0 and prints nothing, because orientation must not break on a diagnostic.

If it reports BEHIND, offer the update command it prints; don't run it unprompted — an install can change the behaviour of skills mid-session.

Origin: written by neo (laris-co/neo-oracle) 2026-08-16 after a /learn contamination bug was diagnosed in a shelf version that had been superseded for two months. Adopted into the shelf so every oracle gets it.

Step 3: Read latest ψ/ brain files

Sort all ψ/ files by modification time, read the most recent:

find ψ/ -name '*.md' -not -name 'CLAUDE.md' -not -name 'README.md' -not -name '.gitkeep' 2>/dev/null | xargs ls -t 2>/dev/null | head -5

Read those top 5 files. This recovers the same context /compact restores — handoffs, retros, learnings, drafts, whatever was touched last.

Step 4: Dig last session

ORACLE_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
ENCODED_PWD=$(echo "$ORACLE_ROOT" | sed 's|^/|-|; s|[/.]|-|g')
PROJECT_BASE=$(ls -d "$HOME/.claude/projects/${ENCODED_PWD}" 2>/dev/null | head -1)
export PROJECT_DIRS="$PROJECT_BASE"

# Strip -wt* suffix to find parent project dir
PARENT_ENCODED=$(echo "$ENCODED_PWD" | sed 's/-wt-[^/]*$//')
if [ "$PARENT_ENCODED" != "$ENCODED_PWD" ]; then
  PARENT_BASE=$(ls -d "$HOME/.claude/projects/${PARENT_ENCODED}" 2>/dev/null | head -1)
  [ -n "$PARENT_BASE" ] && export PROJECT_DIRS="$PROJECT_DIRS:$PARENT_BASE"
fi

# nullglob-safe worktree scan (both parent and self)
for base in "$PROJECT_BASE" "$PARENT_BASE"; do
  [ -z "$base" ] && continue
  for wt in "$base"-wt-*(N); do  # (N) = zsh nullglob qualifier
    [ -d "$wt" ] && export PROJECT_DIRS="$PROJECT_DIRS:$wt"
  done
done

python3 ~/.claude/skills/dig/scripts/dig.py 1

Include in recap:

📡 Last session: HH:MM–HH:MM (Xm, N msgs) — [topic]

Need more? /dig 5 or /dig --timeline.

Total: 1 bash call + LLM analysis

QUICK MODE (/recap --quick)

Minimal, no content reads:

bun ~/.claude/skills/recap/recap.ts

Script outputs git status + focus state (~0.1s). Then LLM adds:

• What's next? (2-3 options based on git state)

"What's next?" Rules

Hard Rules

1. ONE bash call — never multiple parallel calls (adds latency)
2. No subagents — everything in main agent
3. Ask, don't suggest — "What next?" not "You should..."
4. Verify pending before reporting — see "Verify Before Reporting" section below. This is NON-NEGOTIABLE.
5. Print absolute paths — when referencing vault files, render the resolved $ROOT/ψ/... path (starts with /). Bare ψ/... is not clickable. See CONVENTIONS.md.

Verify Before Reporting (MANDATORY)

Handoffs, retros, and memory files are point-in-time claims, not live state. Between the previous session ending and this one starting, work may have been done, PRs may have merged, files may have been copied. Echoing a stale pending list as if it were current is a lie by omission — the human ends up chasing items that are already done.

The rule

Before outputting any "Pending" table or "Next action" suggestion, you MUST verify each claimed pending item against current reality:

What to do with each verified item

• Already done → drop from pending, note in "Actually done since handoff"
• Still pending → keep, show in table
• Partially done → split into remaining sub-items
• Can't verify (offline/ambiguous) → mark ⚠️ unverified in the table, do not assert state

The correction pattern

If the handoff pending list and reality diverge (>1 item stale), show the correction explicitly so the human sees the drift:

| Item | Handoff said | Reality |
|------|--------------|---------|
| Copy cache/ to maw-ui | pending | DONE (Apr 20 04:16) |
| PR #4 merge | open | MERGED |

Why this is non-negotiable

• Handoffs are written before work stops, but work often continues between sessions (other Oracles, scheduled tasks, user actions).
• Memory files age — an older memory claiming "feature X is broken" may be stale if a fix shipped.
• Humans trust recap output as ground truth. An unverified echo breaks that trust fast.

Patterns over intentions — the code is the truth, the handoff is an intention. Always verify.

NOW MODE (/recap --now)

Mid-session awareness from AI memory — no file reading needed. Use when user asks "where are we", "now", "status", "what are we doing".

AI reconstructs session timeline from conversation memory:

## This Session

| Time | Duration | Topic | Jump |
|------|----------|-------|------|
| HH:MM | ~Xm | First topic | - |
| HH:MM | ~Xm | Second topic | spark |
| HH:MM | ongoing | **Now**: Current | complete |

**Noticed**:
- [Pattern - energy/mode]
- [Jump pattern: sparks vs escapes vs completions]

**Status**:
- Energy: [level]
- Loose ends: [unfinished]
- Parked: [topics we'll return to]

**My Read**: [1-2 sentences]

---
**Next?**

Jump Types

Healthy session: Mostly sparks and completes Warning sign: Too many escapes = avoidance pattern

NOW DEEP MODE (/recap --now deep)

Same as --now but adds bigger picture context.

Step 1: Gather (parallel)

1. Current session from AI memory
2. Read latest handoff: ls -t ψ/inbox/handoff/*.md | head -1
3. Git status: git status --short
4. Tracks: cat ψ/inbox/tracks/INDEX.md 2>/dev/null

Step 1.5: VERIFY pending from handoff

Before outputting, run verification checks against each pending item (see "Verify Before Reporting" above). Batch checks in parallel:

• gh pr list --state all for PR claims
• ls path/to/file for "copy X" claims
• grep for "apply pattern" claims

If any diverge from the handoff, show the correction table.

Step 2: Output

Everything from --now, plus:

### Bigger Picture

**Came from**: [Last session/handoff summary - 1 line]
**Working on**: [Current thread/goal]
**Thread**: [Larger pattern this connects to]

### Pending

| Priority | Item | Source |
|----------|------|--------|
| Now | [Current task] | This session |
| Soon | [Next up] | Tracks/discussion |
| Later | [Backlog] | GitHub/tracks |

### Connections

**Pattern**: [What pattern emerged]
**Learning**: [Key insight from session]
**Oracle**: [Related past pattern, if any]

**My Read**: [2-3 sentences - deeper reflection]

**Next action?**

Session Context

The recap scripts (recap.ts and recap-rich.ts) auto-detect and display the current session:

📡 Session: 74c32f34 | arra-oracle-skills-cli | 2h 15m

Detection: scans ~/.claude/projects/[encoded-pwd]/*.jsonl for the most recent session file, extracts short ID and elapsed time from first timestamp.

If session detection fails, skip silently — it's informational only.

Demographics Context

If CLAUDE.md contains demographics from /awaken wizard v2, include in recap output:

**Oracle**: [name] ([pronouns]) | **Human**: [name] ([pronouns]) | **Language**: [pref]

Add this as one line after the timestamp in any mode. If demographics not present, skip silently.

Look for fields in CLAUDE.md: Human Pronouns, Oracle Pronouns, Language, Team, Experience.

Philosophy: Detect reality. Surface blockers. Offer direction. "Not just the clock. The map."

Version: 8.0 (Merged where-we-are into --now mode) Updated: 2026-02-10