Tq forge

/tq-forge — autonomous skill + agent forge

When to use

A workflow keeps coming up that you'd want to one-shot in the future, or you need a specialized agent (researcher, coder, business-analyst, ops-manager, scraper, reviewer, sales-agent) tuned to your own domain context. /tq-forge takes the intent in plain English, decides which kind to build, writes it to the sandbox, scores it, and prints the next command. Nothing goes live until you run /tq-forge-promote <slug>.

Procedure

Scripts live at ${CLAUDE_PLUGIN_ROOT:-${TQ_FORGE_HOME:-$HOME/.tq-forge}/install}/scripts. State lives at ${TQ_FORGE_HOME:-$HOME/.tq-forge}. Initialize both up front:

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. Check the halt flag. If $TQ_FORGE_HOME/halt.flag exists, the user has paused forge work. Queue the intent and exit cleanly:

   test -f "$TQ_FORGE_HOME/halt.flag" && \
     bash "$S/forge-checkpoint.sh" queue "<intent>" && \
     echo "⏸  Forge paused (halt.flag set). Queued. Remove the flag to resume." && exit 0
   

2. Classify intent. Decide:

   • agent if intent uses persona language ("an agent that…", "a researcher for…", "ops manager that watches X") or names a known kind: researcher, coder, business-analyst, ops-manager, scraper, reviewer, sales-agent.
   • skill if intent describes a procedure ("a command for daily summaries", "scaffold a skill that does X", "automate Y").
   • ambiguous -> AskUserQuestion: "Skill (one-shot procedure) or Agent (persona with system prompt + handoff rules)?"

3. Scan for duplicates. Don't forge something that already exists:

   ls "$CLAUDE_SKILLS_DIR" "$SANDBOX_SKILLS" "$SANDBOX_AGENTS" 2>/dev/null
   grep -liE "<keyword from intent>" "$CLAUDE_SKILLS_DIR"/*/SKILL.md 2>/dev/null
   

   If a near-match exists, print the path and ask: fork, edit, or abort. Never silently overwrite.

4. Pick a slug. Lowercase, hyphenated, <=32 chars. "Notion daily summary" -> notion-daily-summary. Confirm via AskUserQuestion only if ambiguous.

5. Hand off to the right builder.

   • skill -> follow the /tq-forge-skill procedure (writes $SANDBOX_SKILLS/<slug>/SKILL.md).
   • agent -> follow the /tq-forge-agent procedure (writes the 5-file agent under $SANDBOX_AGENTS/<slug>/, injecting $TQ_FORGE_HOME/context.md wherever {{CONTEXT}} appears).

6. Score the output.

   bash "$S/quality-score.sh" "$SANDBOX_SKILLS/<slug>"   # or $SANDBOX_AGENTS
   bash "$S/dry-test.sh" "$SANDBOX_SKILLS/<slug>"
   

   Score must be >=7. If lower, refine the weakest dimension once and re-score. If still <7, flag it:

   bash "$S/forge-checkpoint.sh" review "<slug>" "<weak-dim>" "<score>"
   

7. Log it. Append to $TQ_FORGE_HOME/skill-log.json:

   python3 -c "import json,datetime,pathlib,os;p=pathlib.Path(os.environ['TQ_FORGE_HOME'])/'skill-log.json';d=json.loads(p.read_text() or '[]');d.append({'slug':'<slug>','kind':'<skill|agent>','score':<S>,'at':datetime.datetime.now(datetime.timezone.utc).isoformat(),'intent':'<intent>','status':'sandbox'});p.write_text(json.dumps(d,indent=2))"
   

8. Print summary. slug, kind, sandbox path, score, and the exact next command: /tq-forge-promote <slug>.

Pitfalls

• The sandbox is mandatory. Even a trivial skill goes through promote — that's the only path that writes into $CLAUDE_SKILLS_DIR.
• Don't forge a duplicate. The duplicate scan in step 3 is not optional.
• Slug collisions: if <slug> already exists in the sandbox, append -v2 rather than overwriting.
• Agents need $TQ_FORGE_HOME/context.md filled in. With the placeholder still in place, agents will be generic — tell the user to edit it.

Verification

• ls "$SANDBOX_SKILLS/<slug>" (or $SANDBOX_AGENTS) shows the new files.
• skill-log.json includes the new entry with status: sandbox.
• The printed score is >=7, or it was flagged for manual review with a reason.

Tags

tq-forge meta factory sandbox-only