smithy

Debugging: reproduce → read-only root-cause analysis → the user approves the fix → regression test first (jigsmith) → ONE review. Triggers: 'anneal', 'debug', 'why is this broken', a failing test or run with an unknown cause.

Research → spec: explore the code, turn every assumption into a question or a recommendation, write spec.md. Building a feature starts here. Triggers: 'assay', 'research this', 'write a spec', 'new feature'.

Spec → plan: ≤8 tasks, each with a verify check, a persona pass, parallel batch markers with proof, and one self-contained brief per task. Works from a verbal request too (writes a mini spec). Triggers: 'blueprint', 'plan this', 'break into tasks'.

Design review and polish: screenshot the live local UI, judge it against DESIGN.md (or stated rules), score it, and send the fixes the user approves through strike — with before/after screenshots as proof. Triggers: 'burnish', 'polish the UI', 'design review'.

View or change smithy settings for ALL projects (global) or THIS project — model and effort per role, TDD settings, gates, review panel, where memory lives. Tests that a model works before saving it. Triggers: 'calibrate', 'smithy config', 'change the model for review'.

Write project test personas from the system's real user roles (code evidence + a short interview). They power per-persona QA in wield and the end-user view in guild. Triggers: 'commission', 'define the users', 'who uses this system'.

Build the plan task by task with forger or jigsmith agents (each self-checks and makes one commit), then ONE review of the whole job. Parallel batches in worktrees. Works without a plan for a single task. Triggers: 'forge', 'implement the plan', 'build this'.

Production-readiness panel: several persona reviewers run in parallel (masters judge the craft, patrons judge the experience) → one PRODUCTION_READY or NOT_READY verdict. Works on a job or on any diff. Triggers: 'guild', 'ready to ship?', 'review panel'.

Session handoff where every claim cites evidence, so the next session resumes with no re-discovery. Covers the active job, open lanes and a half-done TDD task. Triggers: 'handover', 'handoff', 'save session', 'wrap up'.

Performance: measure a baseline first, at least 3 runs and the median, find hot spots with profiler proof, give recommendations only (never edits code). Works inside a job or on its own. Triggers: 'hone', 'why is it slow', 'benchmark'.

Code review with two verdicts (does it meet the spec? is the code good?) — every finding has proof, a severity reason and a confidence. Reviews a branch, a commit range or uncommitted changes. Triggers: 'inspect', 'review this change', 'code review'.

TDD: write failing tests first, then the code, then ONE clean commit per task — the test-first order is proven by snapshots, not extra commits. Triggers: 'jig', 'TDD this', 'test-first', bug fixes with a repro.

Design creation: a direction grounded in the product's own world, shown as HTML previews, then tokens, states, motion and voice written to the project DESIGN.md. Triggers: 'pattern', 'design system', 'make it look good'.

Stress / load test a running service against limits the user sets (never invented), on a local or user-approved target only. Works inside a job or on its own. Triggers: 'proof', 'stress test', 'load test'.

Unit tests for the changed code, per the stack playbook: write the missing ones, run them, flag flaky ones. Works inside a job or on its own. Triggers: 'ring-test', 'unit test this', 'write unit tests'.

Full pipeline orchestrator (research → plan → build + review → panel → test) with approval gates and ledger resume. Triggers: 'run smithy', 'build end to end', 'full pipeline', resume smithy work.

Fix lane for small KNOWN changes and for review/QA findings: mini plan → one yes → bug items test-first (jigsmith), other items by the forger → targeted tests → ONE review → one report. No spec needed. Triggers: 'strike', 'quick fix', 'fix these items', 'fix these findings', 'small change'.

Full test pass: runs ring-test, wield, proof and hone (skipping, with the reason, any suite that cannot run) and gives one READY or NOT READY verdict. Works inside a job or on its own. Triggers: 'temper', 'test everything', 'full test pass'.

Skill router: which smithy skill to use when, what each one needs to run, the priority rules, and the thoughts that mean STOP. A short digest is injected each session; invoke this for the full router.

Functional QA done the way a user would: run real flows, screenshots required for UI, a 0-100 health score, per-persona flows, severity tiers. Works inside a job or on its own. Triggers: 'wield', 'QA this', 'does it work'.