session-digest

name: session-digest description: >- Returns a structured learning-state diff from session evidence without quizzing or writing files. Callers (session-review, startwork, or the user's main context) handle presentation, approval, and writes. Use when learning state is stale, before startwork, or when the user wants to update scores without a full session review.

Session Digest

Three phases. All autonomous. Returns a structured diff as output. The calling context (session-review, startwork, or the user's main conversation) handles presentation, approval, and file writes.

Path Resolution

Resolve all harness file paths (learning/, .claude/references/, .claude/consent.json) from the harness root in ~/.claude/CLAUDE.md, not the current working directory. If needed, read ~/.config/weft/root for the absolute path.

Phase 1: Discover

1. Resolve harness root from ~/.config/weft/root.
2. Determine the digest window start. Check in order: a. learning/.last-digest-timestamp — if it exists, use it. Format: single line containing YYYY-MM-DD, matching session-discovery's --since flag. b. Oldest file in learning/session-logs/ (by filename date) — this is when the harness started tracking sessions. c. learning/current-state.md file creation date (via stat -f %SB on macOS or stat -c %W on Linux) — intake wrote this file. d. If none of the above exist, default to 30 days ago. This is a safety bound for misconfigured installs, not a normal case.
3. Run session-discovery with paths-only output:

   bun "$(cat ~/.config/weft/root)/scripts/session-discovery.ts" \
     --since <window-start> --min-user-messages 5 --paths-only
   

   If the command fails: report the failure and exit. If output is empty: report "no undigested sessions" and exit. Each output line is a JSONL file path for Phase 2.
4. Proceed directly to Phase 2. If the caller wants confirmation (e.g., first-run warning), they handle it before dispatching.

Phase 2: Extract

For each session in the manifest, run session-extract:

bun "$(cat ~/.config/weft/root)/scripts/session-extract.ts" <filePath>

If session-extract fails (script not found, bun error, exit non-zero): report the failure and exit. Do not attempt to parse raw JSONL directly — the tool budget required makes synthesis impossible.

Context management gate

Each reader (inline or sub-agent) receives:

• Current learning/current-state.md (so it knows existing concepts/scores)
• Instructions to extract:

concepts_encountered:
  - concept: [name — match existing current-state names when possible]
    evidence: [specific quote, paraphrase, or behavioral description]
    signal_type: new_exposure | struggle | breakthrough | teaching | deepening
    estimated_score: [0-5, or null if uncertain]
    gap_type: [conceptual | procedural | recall, or null]
    confidence: [high | moderate | low]

procedural_observations:
  - [workflow patterns, tool usage, debugging approach — brief]

What to look for in extracted text

What counts as a growth-edge encounter (extract these):

• New concept exposure
• Struggle/debugging that reveals a gap (tool errors are marked with x)
• Breakthrough that demonstrates fluency change
• Teaching/explaining a concept (shows depth)
• Deepening — using a concept in a novel context or combining it

What doesn't count (skip these):

• Routine use of a familiar concept without struggle
• Pure orchestration ("commit push merge")
• Routine tool calls

Phase 3: Synthesize

Load learning/current-state.md. Compare extracted concepts against current state. Build a proposed diff with four sections:

A. Score changes (existing concepts with new evidence)

• Show: concept name, current score -> proposed score, evidence, reasoning
• Only propose changes when evidence is strong enough (high/moderate confidence, clear signal type)

B. New evidence (existing concepts, score unchanged, but new evidence worth recording)

• Append to history with digest:observed tag
• Show: concept name, current score (unchanged), new evidence note

C. New concepts (not in current-state.md)

• Show: proposed name, arc (existing or TBD), proposed score, gap type, evidence
• Flag confidence: low entries as quiz candidates

D. Flags (observations that don't map to score changes)

• Activity that doesn't map to existing arcs (possible new arc)
• Concepts with contradictory signals across sessions
• Concepts that appear heavily used but have no current-state entry

Output Format

Return the structured diff using these four sections. This is the return value — digest does not write files.

## Proposed Updates (N sessions, date range)

### Score Changes
  concept-name: current -> proposed (gap type)
    Evidence: [specific quote, paraphrase, or behavioral description]

### New Evidence (score unchanged)
  concept-name: score (unchanged)
    Evidence: [new observation worth recording]

### New Concepts
  concept-name (arc: arc-name, score: N, gap: type)
    Evidence: [description]
    ! Low confidence — suggest for quiz  (when applicable)

### Flags
  > [observations that don't map to score changes]

YAML entry schema for callers writing to current-state.md:

  - name: concept-name
    score: 3
    gap: procedural        # omit or use -- when score >= 4
    source: digest:observed
    last-updated: YYYY-MM-DD
    history:
      - { date: YYYY-MM-DD, score: 3, note: "brief qualitative note" }

No last-quizzed or times-quizzed fields — digest doesn't quiz.

Callers are responsible for writing approved entries to current-state.md and updating learning/.last-digest-timestamp to the latest digested session's end date.

Standalone Invocation

When the user invokes /session-digest directly (not as a sub-agent):

1. Run Phases 1-3 to produce the structured diff.
2. If this is the first digest (no .last-digest-timestamp), note that this covers all sessions since intake — user may want to narrow the window. Ask before proceeding.
3. Present the diff using the Output Format above.
4. User approves all, approves with modifications, skips items, or rejects.
5. On approval: write approved entries to current-state.md using the YAML schema in Output Format. Update .last-digest-timestamp to the latest digested session date.
6. On rejection: exit without writing.

Scope

v1: current-state.md only. Goals.md and arcs.md stay untouched. If digest sees activity that doesn't map to existing arcs, it flags it in the diff presentation — user decides.

No new arcs. Concepts that don't fit existing arcs get proposed with arc: TBD and a note. User assigns the arc during approval.

Digest writes nothing. Returns structured diff. Caller writes approved changes to current-state.md and updates .last-digest-timestamp. (Standalone invocation handles its own writes per the Standalone Invocation section above.)

No session logs. Session logs are session-review's domain.

Anti-Patterns

• Don't inflate scores from transcript evidence. Bias conservative — reading about someone debugging !== watching them debug.
• Don't create near-duplicate concept names. Always check current-state first and match existing names.
• Don't digest the current session. The digest window ends before the session that's running the digest.

Interoperation