forge

Forge Skill

Cross-vendor analog: the capture half of Anthropic Managed Agents' memory + dreaming pair (May 2026). Forge mines transcripts; /dream curates between sessions. Off the API, local, cross-vendor.

Typically runs automatically via SessionEnd hook.

Loop position: capture sub-step of move 7 in the [operating loop](../../docs/architecture/operating-loop.md). Extracts candidate learnings from transcripts; the [promotion ratchet](../../docs/architecture/operating-loop.md#the-promotion-ratchet) decides which ones survive (one-offs die at handoff; repeats promote to .agents/learnings/). Forge is the funnel, not the filter.

Extract knowledge from session transcripts.

How It Works

The SessionEnd hook runs:

ao forge transcript --last-session --queue --quiet

This queues the session for knowledge extraction.

Flags

| Flag | Default | Description | |------|---------|-------------| | --promote | off | Process pending extractions from .agents/knowledge/pending/ and promote to .agents/learnings/. Absorbs the former extract skill. |

Promote Mode

Given /forge --promote:

Promote Step 1: Find Pending Files

ls -lt .agents/knowledge/pending/*.md 2>/dev/null
ls -lt .agents/ao/pending.jsonl 2>/dev/null

If no pending files found, report "No pending extractions" and exit.

Promote Step 2: Process Each Pending File

For each file in .agents/knowledge/pending/:

1. Read the file content
2. Validate it has required fields (# Learning:, Category:, Confidence:)
3. Copy to .agents/learnings/ (preserving filename)
4. Remove the source file from .agents/knowledge/pending/

Promote Step 3: Process Pending Queue

if [ -f .agents/ao/pending.jsonl ] && [ -s .agents/ao/pending.jsonl ]; then
  # Process each queued session
  cat .agents/ao/pending.jsonl
  # After processing, clear the queue
  > .agents/ao/pending.jsonl
fi

Promote Step 4: Report

Promoted N learnings from pending → .agents/learnings/
Queue cleared.

Done. Return immediately after reporting.

Manual Execution

Given /forge [path]:

Step 1: Identify Transcript

With ao CLI:

# Mine recent sessions
ao forge transcript --last-session

# Mine specific transcript
ao forge transcript <path>

Without ao CLI: Look at recent conversation history and extract learnings manually.

Step 2: Extract Knowledge Types

Read skills/forge/references/uncaptured-lesson-patterns.md for signal patterns and the 26 known uncaptured lesson categories.

Look for these patterns in the transcript:

| Type | Signals | Weight | |------|---------|--------| | Decision | "decided to", "chose", "went with" | 0.8 | | Learning | "learned that", "discovered", "realized" | 0.9 | | Failure | "failed because", "broke when", "didn't work" | 1.0 | | Pattern | "always do X", "the trick is", "pattern:" | 0.7 |

Uncaptured Lesson Matching: During transcript scanning, match events against the 26 known uncaptured lesson patterns (see references/uncaptured-lesson-patterns.md). Pre-fill learning templates with matched pattern metadata (category, base confidence, pattern number tag).

Step 3: Write Candidates

Write to: .agents/forge/YYYY-MM-DD-forge.md

# Forged: YYYY-MM-DD

## Decisions
- [D1] <decision made>
  - Source: <where in conversation>
  - Confidence: <0.0-1.0>

## Learnings
- [L1] <what was learned>
  - Source: <where in conversation>
  - Confidence: <0.0-1.0>

## Failures
- [F1] <what failed and why>
  - Source: <where in conversation>
  - Confidence: <0.0-1.0>

## Patterns
- [P1] <reusable pattern>
  - Source: <where in conversation>
  - Confidence: <0.0-1.0>

Step 4: Index for Search

if command -v ao &>/dev/null; then
  ao forge markdown .agents/forge/YYYY-MM-DD-forge.md 2>/dev/null
else
  # Without ao CLI: auto-promote high-confidence candidates to learnings
  mkdir -p .agents/learnings .agents/ao
  for f in .agents/forge/YYYY-MM-DD-*.md; do
    [ -f "$f" ] || continue
    # Extract confidence (numeric or categorical)
    CONF=$(grep -i "confidence:" "$f" | head -1 | awk '{print $NF}')
    # Normalize categorical to numeric: high=0.9, medium=0.6, low=0.3
    case "$CONF" in
      high) CONF_NUM=0.9 ;; medium) CONF_NUM=0.6 ;; low) CONF_NUM=0.3 ;; *) CONF_NUM=$CONF ;;
    esac
    # Auto-promote if confidence >= 0.7, prepending required frontmatter
    if (( $(echo "$CONF_NUM >= 0.7" | bc -l) )); then
      { printf -- '---\ntype: learning\nsource: forge\ndate: %s\nmaturity: provisional\nutility: 0.5\n---\n' "$(date +%Y-%m-%d)"; cat "$f"; } > .agents/learnings/"$(basename "$f")"
      TITLE=$(head -1 "$f" | sed 's/^# //')
      echo "{\"file\": \".agents/learnings/$(basename $f)\", \"title\": \"$TITLE\", \"keywords\": [], \"timestamp\": \"$(date -Iseconds)\"}" >> .agents/ao/search-index.jsonl
      echo "Auto-promoted (confidence $CONF): $(basename $f)"
    fi
  done
  echo "Forge indexing complete (ao CLI not available — high-confidence candidates auto-promoted)"
fi

Step 5: Update Capture Tracking

After extracting learnings that match uncaptured lesson patterns (Step 2), record which patterns were captured. This state lives in .agents/forge/capture-tracking.json (a runtime artifact, never in skills/).

mkdir -p .agents/forge

1. Read .agents/forge/capture-tracking.json if it exists, otherwise start with {}
2. For each matched pattern, add or update an entry keyed by pattern number:

``json { "3": {"captured": true, "date": "2026-03-30", "learning_path": ".agents/learnings/tooling/use-bin-cp.md"}, "7": {"captured": true, "date": "2026-03-29", "learning_path": ".agents/learnings/operations/worktree-commit.md"} } ``

1. Write the updated JSON back to .agents/forge/capture-tracking.json

Pattern numbers correspond to the numbered headings in references/uncaptured-lesson-patterns.md (1-30, 26 total patterns).

Step 6: Report Results

Tell the user:

• Number of items extracted by type
• Location of forge output
• Candidates ready for promotion to learnings
• Capture progress: "X/26 uncaptured lesson patterns captured" (read from .agents/forge/capture-tracking.json)

The Quality Pool

Forged candidates enter at Tier 0 (.agents/forge/), then promote to Tier 1 (.agents/learnings/) via human review, 2+ citations, or auto-promote when confidence >= 0.7 (ao-free fallback).

Key Rules

• Runs automatically - usually via hook
• Extract, don't interpret - capture what was said
• Score by confidence - not all extractions are equal
• Queue for review - candidates need validation

Examples

See [references/examples.md](references/examples.md) for the SessionEnd hook invocation walkthrough and manual transcript-mining walkthrough.

Troubleshooting

| Problem | Cause | Solution | |---------|-------|----------| | No extractions found | Transcript lacks knowledge signals or ao CLI unavailable | Check transcript contains decisions/learnings; verify ao CLI installed | | Low confidence scores | Weak signals or vague conversation | Focus sessions on concrete decisions and explicit learnings | | forge --queue fails | CLI not available or permission error | Manually append to .agents/ao/pending.jsonl with session metadata | | Duplicate forge outputs | Same session forged multiple times | Check forge filenames before writing; ao CLI handles dedup automatically |

Reference Documents

• [references/forge.feature](references/forge.feature) — Executable spec: mine transcripts → queued candidates, funnel-not-filter, --promote, consumes-transcripts-not-skills (soc-qk4b)
• [references/uncaptured-lesson-patterns.md](references/uncaptured-lesson-patterns.md) — signal patterns and 26 known uncaptured lesson categories for transcript mining
• [references/feedback-compiler-drafts.md](references/feedback-compiler-drafts.md) — auto-drafted learning workflow (F