forge

The forge metaphor is literal: HEAT — generate competing approaches, inject entropy HAMMER — iterate implementation against failing tests QUENCH — quality gates, adversarial review, convergence check TEMPER — experiment on output, discover emergent patterns, self-correct ANNEAL — if brittle, reheat and try a different crystalline structure

TEMPER is the evolution: the forge doesn't just build — it measures, hypothesizes, tests, and adapts. Every cycle produces data. Every data point refines the next cycle.

Run this overnight. Come back to shipped code + experimental evidence + emergent learnings.

<skill_load> always: skills/quality/SKILL.md, skills/build/reference/testing.md, skills/build/reference/git.md, skills/build/SKILL.md on_classify: bug: skills/debug/SKILL.md refactor: skills/build/reference/refactor.md on_domain: api: skills/architecture/reference/api.md, skills/architecture/reference/backend.md auth: skills/tearitapart/reference/security.md frontend: skills/frontend/SKILL.md backend: skills/architecture/reference/backend.md on_tier: 2+: skills/orchestration/SKILL.md </skill_load>

<on_start>

agentdb read-start
agentdb emit command "forge-start" "" '{"goal":"...","max_budget_usd":N.NN}'

Load ALL always-skills immediately. Load task/domain skills after classify.

Budget preflight (mandatory before any forge cycle):

1. Confirm max_budget_usd is set on the contract or the forge invocation.
2. If unset: AskUserQuestion — "Forge runs are autonomous loops. Set a cost ceiling (default: $5 tier 2 / $15 tier 3), or proceed unbounded?" Never proceed unbounded silently.
3. Cumulative cost tracked via agentdb emit each cycle. Hard stop at 100% of budget.
4. Hitting the cap = forge halts and reports. This is the circuit breaker working.

Why: one stuck retry at $0.40-0.60/query × 200 retries = $120 silently. The cap is the only mechanism that prevents this — there is no in-session signal that cost is runaway. See skills/orchestration/SKILL.md <max_budget_usd_invariant>. </on_start>

**RULE: Research without verification is theory fiction.** (LRN-F11)
Never generate approaches from research alone. Each candidate must have a verification path
(test, prototype, visual proof) defined upfront. If you can't verify it, don't propose it.

**RULE: Avoid specific multiplier claims.** (H104 graduated meta-rule)
Approaches that promise "3-5x improvement" or "always better" have a 71% refutation rate.
Frame as directional: "reduces X" not "reduces X by 80%." Measure after, not before.

1. Read agentdb context + _meta/research/ for prior work.
1b. Measure entropy: check agentdb learning count in domain, test coverage, recent failures.
    Low entropy → generate 1 approach (streamlined). High entropy → generate 3 approaches (full exploration).
2. Classify task: type, tier, domain.
3. Generate 2-3 candidate approaches (not variations — genuinely different strategies).
4. For each: files affected, tests needed, effort estimate, known risks, **verification method**.
5. For each: **hypothesis** — what testable claim does this approach make?

Tier 1: generate inline.
Tier 2+: spawn parallel surgeon agents, one per approach.
Tier 3: spawn full council (researcher + scout in parallel → dreamer → surgeons).

```bash
agentdb emit command "forge-heat" "" '{"approaches":N,"tier":N,"hypotheses":["H-FORGE-1","H-FORGE-2"]}'
```