tq-forge-improve-all

/tq-forge-improve-all — batch improve every sub-7 artifact

When to use

You want a single sweep that finds every forged skill and agent scoring below 7 and fixes them in one session. Good after an initial forge spree, or after /tq-forge-scan surfaces several weak entries. Each artifact gets one improve attempt; if it still fails, it's flagged for manual review and the loop moves on.

Procedure

export TQ_FORGE_HOME="${TQ_FORGE_HOME:-$HOME/.tq-forge}"
S="${CLAUDE_PLUGIN_ROOT:-${TQ_FORGE_HOME:-$HOME/.tq-forge}/install}/scripts"
source "$S/common.sh" && tq_ensure_home

1. Enumerate artifacts below threshold. From the log plus any un-logged sandbox dirs:

   python3 - <<'PY'
   import json, pathlib, os
   home = pathlib.Path(os.environ['TQ_FORGE_HOME'])
   log = json.loads((home/'skill-log.json').read_text() or '[]')
   weak = [e for e in log if isinstance(e.get('score'),(int,float)) and e['score'] < 7]
   weak.sort(key=lambda e: e['score'])
   for e in weak:
       print(f"{e['slug']}\t{e['score']}\t{e.get('status','?')}")
   PY
   

2. Print the improvement queue so the user can abort if it's longer than expected:

   🔧 Improvement queue — <N> artifacts below 7/10
    1. notion-daily-summary    5.8/10  (weak: actionability, specificity)
    2. inventory-snapshot      6.2/10  (weak: completeness)
   

3. For each artifact in order: a. Check the halt flag before each item:

   test -f "$TQ_FORGE_HOME/halt.flag" && echo "⏸  halt.flag set — stopping batch." && exit 0
   

   b. Run the /tq-forge-improve logic inline:

   bash "$S/improve-loop.sh" assess "<slug>"
   

   Read the artifact, rewrite only the weak sections, then:

   bash "$S/improve-loop.sh" verify "<slug>"
   

   c. Print a one-line status: ✅ notion-daily-summary: 5.8 -> 8.1 or ⚠️ inventory-snapshot: 6.2 -> 6.5 — still weak, flagged for review.

4. Print the final summary.

   📊 Improve-all complete
   Improved:   <N>  (now >=7)
   Still weak: <M>  (flagged in forge-queue.json)
   Skipped:    <K>  (halt.flag hit mid-run)
   

5. If any remain in needs_manual_review, print: "Run /tq-forge-queue to see what needs a human touch."

Pitfalls

• An artifact scoring 6.9 on one run can score 7.1 on the next with no edit (scorer variance). Accept >=7.0 as passing; don't re-run needlessly.
• If an agent's tools.json is malformed, skip that agent with a warning rather than aborting the whole batch.
• Promoted artifacts under $CLAUDE_SKILLS_DIR are editable here — the path is already writable. Don't treat "promoted" as read-only.

Verification

• skill-log.json score fields are updated for every processed artifact.
• The printed "Improved + Still weak + Skipped = N" matches the step-2 queue.
• Running the skill again immediately prints "0 artifacts below 7" if all passed.

Tags

tq-forge improve batch quality bulk