progress-review

name: progress-review description: Cross-session learning pattern analysis. Reads multiple session logs and learning state to detect stalls, regressions, goal drift, and arc readiness that no single session review can see. Proposes updates to learning state files. Use when the user asks "how am I doing", "review my progress", "am I improving", or "progress review". Also dispatched automatically by startwork when enough sessions have accumulated.

Progress Review

Analyzes learning patterns across multiple sessions. Detects what single session reviews miss: stalls, regressions, goal drift, arc readiness, and calibration mismatches. Proposes structural updates to learning state files. All changes are human-gated.

Two activation paths. The phase structure is the same for both.

Activation

Path A: Dispatched from startwork (primary)

Startwork dispatches this skill as a background sub-agent when both threshold gates pass (minimum days since last review AND minimum substantive session count — both configurable in ~/.config/weft/config.json). The sub-agent receives:

• The full session-discovery manifest (all sessions, unfiltered)
• All session log frontmatter for the review period
• Full contents of learning/current-state.md, learning/goals.md, learning/arcs.md
• List of scaffold files in learning/scaffolds/ (with dates)
• git log --oneline -20
• Contents of learning/.progress-review-log.md (for deferred findings)

When running as a sub-agent:

1. Skip Gather phase gates (day + session count) — startwork already applied them.
2. Cross-reference the passed manifest against session log frontmatter dates to identify unreviewed sessions. Match by calendar date of session start time.
3. If unreviewed sessions exist: dispatch session-digest as foreground sub-agent. Pass:
   • The filtered manifest (unreviewed sessions only)
   • Full contents of learning/current-state.md
   • Full contents of .claude/skills/session-digest/SKILL.md
   • Instruction: "Skip your Phase 1 discovery. Use this session manifest as your session list: [manifest JSON]. Proceed directly to Phase 2 Extract for these sessions. Return structured diff per your Output Format section." Receive the structured diff. (Do not parse JSONL directly — digest handles extraction.)
4. Proceed to Phase 1 with session log data + digest output.

If session-discovery manifest was not passed (startwork failed to run discovery): proceed with session log frontmatter only. Note reduced coverage in Phase 3 output.

Path B: Direct invocation (secondary)

The user invokes the skill directly. Run the Gather phase first, then proceed through Phases 1-4 inline.

Gather (Path B only)

Read all data sources. Every source is optional — degrade gracefully.

1. Read ~/.config/weft/config.json. Extract progressReviewDays (default 3) and progressReviewSessions (default 5).
2. Read learning/.progress-review-log.md for last review date. If file doesn't exist: first run. Set window start to intake date (creation date of learning/current-state.md via stat). Pre- intake sessions have no learning profile to compare against.
3. Day gate: compute days since last review. If < progressReviewDays: "Last progress review was N days ago — too recent for cross-session patterns. Run again in M days, or proceed anyway?" If they decline, stop. First run (no prior review date): day gate passes automatically.
4. Run session-discovery for the review window:

   bun "$(cat ~/.config/weft/root)/scripts/session-discovery.ts" --since <window-start>
   

   If session-discovery fails (bun not found, script error): warn and fall back to session log count for the session gate (step 5). Set a flag (discoveryFailed = true) so later steps know to skip cross-referencing and digest dispatch.
5. Session gate: filter the discovery manifest for substantive sessions (userMessageCount >= 10). Count them. If count < progressReviewSessions: "Only N substantive sessions since the last review — not enough for cross-session patterns yet." Offer to proceed. If they decline, stop.
   • Fallback (discoveryFailed): count session logs in learning/session-logs/ since the window start. Apply progressReviewSessions threshold against that count.
6. Cross-reference discovery manifest against session logs. Match by calendar date of session start time: a session starting on 2026-03-04 matches 2026-03-04.md if that file exists. Sessions with a match = reviewed. Others = unreviewed. (Skip if discoveryFailed.)
7. Dispatch session-digest as foreground sub-agent for unreviewed sessions. Pass:
   • The filtered discovery manifest (unreviewed sessions only)
   • Full contents of learning/current-state.md
   • Full contents of .claude/skills/session-digest/SKILL.md
   • Instruction: "Skip your Phase 1 discovery. Use this session manifest as your session list: [manifest JSON]. Proceed directly to Phase 2 Extract for these sessions. Return structured diff per your Output Format section." Receive the structured diff. Hold it for Phase 1.
   • If no unreviewed sessions: skip dispatch, proceed normally.
   • If digest fails: warn, continue with session logs only.
   • (Skip entirely if discoveryFailed.)
8. Read YAML frontmatter from each session log in the review window. Read the body of the most recent 3 logs.
9. Read learning/current-state.md, learning/goals.md, learning/arcs.md (each optional — note what's missing).
10. List learning/scaffolds/ files with dates.
11. Run git log --oneline -20 for project work evidence.
12. Read learning/.progress-review-log.md deferred entries (if any).

Proceed to Phase 1 with whatever data is available.

Phase 1: Analyze

Single-pass analysis across four lenses. Returns findings only — zero file writes. If session-digest was dispatched (Gather step 7 or Path A step 3), its structured diff supplements session log data for all four lenses.

Deferred finding check

Before analyzing new patterns, check learning/.progress-review-log.md for the most recent entry's deferred: items. For each:

• If the pattern still exists or worsened: promote to high priority
• If the pattern resolved: note as resolved (no action needed)

Concept lens

Detects patterns in concept scores across sessions:

• Stalls: quizzed ≥ 3 times, score stuck ≤ 2. The learner is putting in reps but the approach isn't working.
• Regressions: score dropped ≥ 2 from a previously-solid concept, or dropped below 3 after being ≥ 4. Something that was working isn't anymore.
• Score velocity: improving / flat / declining per concept across sessions in the review window. Flat at a low score is a stall signal. Flat at a high score is fine.
• Scaffold calibration mismatches: lesson-scaffold classified a concept as "solid" but session-review scored it ≤ 2, or scaffold said "prerequisite gap" but learner scored ≥ 4. Indicates the learner model was wrong at scaffold time.

Data: learning/current-state.md, session log frontmatter (concepts, scores), learning/scaffolds/ files, digest structured diff (for unreviewed sessions — digest:observed weighted below session-review:quiz).

If current-state.md is missing: concept lens limited to session log scores only.

Arc lens

Detects patterns in developmental arc progression:

• Arc readiness: repeated procedural success across sessions suggests the learner is ready to move from reps to abstraction. Look for score ≥ 3 on multiple arc-related concepts + evidence of successful application in session logs or git.
• Unblocking sequences: goal requires arc X, which is blocked by arc Y or concept Z. Surface the dependency chain.
• Compounding breakdown: arc touched in many sessions but scores and arc state not advancing. Effort is scattered, not focused.
• Workflow friction: repeated blockers or never-completed "remaining work" items across session logs.

Data: learning/arcs.md, session log frontmatter (arcs), session log bodies, git log, digest concept-arc mappings (for unreviewed sessions).

If arcs.md is missing: arc lens skipped.

Goal lens

Detects patterns in goal alignment:

• Project-goal mapping gaps: active project work (from session logs and git) not connected to any stated goal.
• Unconnected effort: sessions with no goal or arc mapping in frontmatter. Work is happening but not tracked.
• Goal drift: stated goal framing doesn't match observed work trajectory. The learner may be pursuing something different than what they wrote down.
• Play-state vs grind-state: ratio of new concepts to repeated concepts across sessions. High repetition with low novelty suggests grind; high novelty suggests exploration. Neither is bad, but the pattern should be visible.

Data: learning/goals.md, learning/arcs.md, session log frontmatter (project, concepts, arcs), digest flags (activity not mapping to existing arcs feeds directly into "unconnected effort" detection).

If goals.md is missing: goal lens skipped.

Learner model lens

Detects patterns in learner behavior that should update the CLAUDE.md predictive sections. Reads enrichment principles from .claude/references/claude-md-template.md and the user's current ~/.claude/CLAUDE.md (weft section).

• Recurring learning mechanics (2+ sessions): the learner consistently acquires understanding through a specific pattern not yet captured in "How {name} learns."
• Recurring unblocking patterns (2+ sessions): the same intervention type works repeatedly — not yet in "How {name} gets unblocked."
• Error shape patterns: mistakes cluster in a recognizable shape across sessions (e.g., "correct structure, wrong boundary" or "right concept, wrong execution order").
• Strength evidence accumulation: a prior-domain capability surfaces in multiple sessions — warrants a "Strengths" entry or upgrade.
• Model contradictions (3+ sessions): an existing CLAUDE.md entry doesn't match observed behavior. Higher threshold because contradictions modify existing entries.

If predictive sections are missing from the CLAUDE.md (predates the learner model template): propose creating them when cross-session evidence supports entries. Organic migration, same as session-review.

Threshold: 2+ sessions for new entries (higher than session-review's single-session threshold). Contradictions require 3+ sessions.

Data: ~/.claude/CLAUDE.md (weft section), session log bodies and frontmatter, .claude/references/claude-md-template.md (enrichment principles), digest procedural observations (for unreviewed sessions).

If CLAUDE.md weft section is missing: learner model lens skipped.

Phase 2: Synthesize

Group findings into themes. 3-5 themes maximum. For each theme:

• What's happening — plain language, no jargon
• Evidence — specific session dates, scores, log references
• Why it matters — what this means for the learner's trajectory
• Suggested action — one of:
  • state-update — score adjustment, gap reclassification
  • arc-update — readiness transition, dependency reorder
  • goal-update — drift correction, reframing
  • model-update — CLAUDE.md learner model entry (add, upgrade, or correct)
  • process-suggestion — workflow change, focus shift
• Draft file changes where applicable: show current value → proposed value, with one-line rationale

"No findings" is valid. If no patterns meet diagnostic criteria: "Reviewed N sessions. No cross-session patterns detected. Learning state looks consistent." Log the review and stop.

State-write priority rule

Session-review scores from direct observation and quiz supersede progress-review's pattern-inferred adjustments.

When proposing a score change that conflicts with a session-review score from the current review window:

• Note the conflict explicitly
• Explain why cross-session pattern evidence suggests otherwise
• Default to the session-review score unless ≥ 3 data points across sessions support the pattern

Tag all proposed updates as progress-review:pattern per .claude/references/scoring-rubric.md.

Phase 3: Present

When dispatched from startwork: return findings as structured text. Startwork handles presentation to the user.

When invoked standalone: present findings directly.

Format (both paths):

• Coverage line: "Analyzed N sessions (M reviewed via session-review, K digested from transcripts)." If discovery or digest failed, note the gap: "N sessions found in logs only — transcript analysis was unavailable."
• Review summary with themes in priority order
• Deferred findings from prior review: resolved or escalated
• Proposed changes grouped by file, each showing current → proposed
• Ask the user to approve, select specific changes, adjust, defer, or reject

Plain language throughout. The review should be readable in under 3 minutes. Show insights, not raw data.

Phase 4: Write

Apply approved changes only.

State file updates

Write approved changes to learning/current-state.md, learning/arcs.md, learning/goals.md, and/or ~/.claude/CLAUDE.md (weft section, for model-update actions). Tag every score update with progress-review:pattern. CLAUDE.md changes follow the same human-gated approval flow as all other state file updates.

Digest timestamp

If session-digest was dispatched during Gather (step 7) or Path A (step 3), update learning/.last-digest-timestamp to the latest digested session's end date after all state file writes complete. This prevents future digest runs (startwork 3c, session-review, standalone /session-digest) from re-processing sessions already covered by this review.

Review log

Append an entry to learning/.progress-review-log.md:

---
date: YYYY-MM-DD
sessions-reviewed:
  - YYYY-MM-DD
  - YYYY-MM-DD
  - YYYY-MM-DD
themes:
  - name: theme name
    pattern-type: stall | regression | readiness | drift | friction
    action: state-update | arc-update | goal-update | process-suggestion
    applied: true | false | deferred
changes-applied:
  - file: learning/current-state.md
    description: brief description of change
deferred:
  - description of deferred finding
---

The log's most recent date sets the window start for the next review.

If the user approved no changes and deferred nothing, still log the review (with empty changes-applied and deferred lists) so the window advances.

Graceful Degradation

Never error on missing data. Degrade each analytical lens silently. Only surface a message when the degradation meaningfully changes what the skill can offer.

Boundaries

• Don't compete with session-review — session-review owns single-session scoring. Progress-review owns cross-session patterns.
• Don't ignore deferrals — if a deferred finding persists or worsened, escalate its priority.
• Don't block startwork — if analysis fails as a background sub-agent, startwork proceeds unaffected.

Consent Gate

External publishing requires both .claude/consent.json and learning/relationships.md with a signal_repo. If either is absent, skip silently.