forge

Forge

Forge runs spec-driven development as four phases. You invoke one phase at a time. Each phase does its work, updates shared state, and recommends (but does not run) the next phase. A fifth command, fix, re-enters the flow to correct course at any layer. Every phase's depth auto-sizes to the change — see Sizing.

/forge specify <name>   Understand the problem + capture requirements   → spec.md
/forge design           Architecture, contracts, verification gates     → design.md
/forge plan             Atomic tasks, dependencies, AC traces            → plan.md
/forge execute          Implement, then run independent verifiers        → working code
/forge fix <change>     Correct course mid-stream, keep the chain aligned → aligned artifacts + code

Dispatch

Read the first argument as the phase verb. Read the matching reference file fully before acting — it is the step-by-step playbook for that phase.

Verb	Read	Produces
`specify`	`references/specify.md` (+ `references/review.md`)	`./.specs/<slug>/spec.md`
`design`	`references/design.md` (+ `references/review.md`)	`./.specs/<slug>/design.md`
`plan`	`references/plan.md`	`./.specs/<slug>/plan.md`
`execute`	`references/execute.md` (+ `references/verification.md`)	code + commits
`fix`	`references/fix.md`	re-aligned artifacts + code

references/structural-standards.md is the single always-applied baseline for what good structure is and how a structural review runs. Three dispatches load it: review.md's under-design lens, execute.md's per-batch structural axis, and fix.md's claim lenses. It is the only structural check set in forge.

If no verb is given, infer the phase from the request and confirm it. Never fail because a verb is missing or unknown. Common mappings:

Request pattern	Phase
"write a PRD", "what should this do", problem with no proposed solution	`specify`
"design this", "how should we build X", architecture question	`design`
"break this into tasks", "plan the work"	`plan`
"implement X", "build this", spec + plan already exist	`execute`
"this detail is wrong", "correct course", "I don't like this design", "found a bug while testing", change to an in-flight spec/design/plan/code	`fix`

specify takes a feature name. If absent, ask for one and derive a kebab-case slug.

Forge does not orchestrate phase transitions — you do. Start at any phase. A phase that finds no earlier artifacts gathers the minimum context it needs; each reference describes its own fallback.

Sizing

Every change is one of three sizes. The size controls how deep each phase goes and which phases are needed at all. See references/sizing.md for the full rubric.

Size	Roughly	Pipeline
quick	one file/function, no new interface/schema/dep	inline spec → execute (skip design + plan)
standard	one component/module, a few files	full spec → light design → phased plan → execute
complex	crosses components/repos, new subsystem, ambiguous	full pipeline, all gates

The first phase to run picks the size and writes it to state.md. Size can only go up. Any later phase may promote it with a logged reason. A downgrade requires the user's confirmation — ratcheting prevents work from quietly skipping rigor a later phase already showed it needed. Each phase ends by listing which downstream phases the current size needs or lets you skip.

Artifacts and state

Artifact root (session CWD)

All ./.specs/* paths in this skill resolve against the session CWD — the directory the agent was invoked in — not the shell's current directory at the moment of the write. Subsequent cds do not move the artifact root.

At the start of every phase, resolve the artifact root once and reuse it. Capture the session CWD as an absolute path (pwd at the very first Bash call of the session, or the harness-provided starting directory) and set SPECS_ROOT="<abs>/.specs". Every read, write, mkdir, and ls inside this skill must go through $SPECS_ROOT/.... Never use a bare .specs/... after that resolution — a bare path would be interpreted against the current shell CWD and drift into a subfolder.

If the resolved root points inside another repo's .specs/ (e.g. the user invoked the skill from a nested package), stop and confirm with the user before writing.

All work for a change lives under ./.specs/<slug>/ (create it if missing):

./.specs/<slug>/spec.md       what & why (requirements, acceptance criteria, scope)
./.specs/<slug>/design.md     how (architecture, contracts, verification gates)
./.specs/<slug>/plan.md       tasks grouped into phases (parallel within a phase, AC-traced)
./.specs/<slug>/state.md      size, decisions log, task status, handoff — the source of truth
./.specs/<slug>/lessons.md    what went wrong here and the rule going forward

Templates for each are in templates/. state.md is the single source of truth. Re-read it at the start of every phase. Do not assume the runtime preserves its own internal state across phases — if it is not in state.md, it did not happen.

Acceptance criteria are written once, in spec.md, with stable IDs (PREFIX-NN). Design, plan, and execute reference those IDs; they never restate the criterion text.

Lessons

Each change keeps its own ./.specs/<slug>/lessons.md with ## Standing Rules (short imperatives, always loaded) and ## Log (tagged, append-only). Load it at the start of every phase, reading only above the  marker. The text below it is parked, and no phase loads it. Append only when something non-obvious was learned: a hack, a gotcha, a corrected wrong assumption, a skipped gate. Triage first — a mechanical observation becomes a proposed check. Routine success writes nothing. See references/lessons.md.

Orchestration (canonical — references should cite, not restate)

Use subagents for parallel or heavy work. They run in Claude Code and OpenCode. They keep the main agent's context small and give independent perspectives (execute's verifiers depend on this).
Workflow is an optional speed-up for dispatching many subagents at once, where the runtime supports it. Plain sequential subagent calls always work as a backup. Never require Workflow.
Match model tier and reasoning effort to the task — don't pay frontier rates for mechanical work. See Model & effort selection below. This is the primary cost lever.
One level of delegation. Subagents do not spawn subagents.
Subagents are stateless. Put everything they need in the prompt (file paths, section refs, constraints). Outputs over ~100 lines go to a file; return the path, not the content.
Prefer AskUserQuestion when available, with a recommended default per question. Fall back to a single plain-text question otherwise. Never bundle unrelated questions.

Model & effort selection (canonical — references cite, not restate)

Every subagent dispatch picks two dials. Set them explicitly per task; never let the whole fan-out default to the most expensive model. This is agent-agnostic — map the tiers to whatever your runtime exposes.

Tier — the model's capability class:

Tier	Use for	Claude Code	OpenCode
economy	mechanical, well-scoped, low-ambiguity work with a clear pass/fail or a precedent to mirror	`haiku`	cheapest capable model configured
standard	ordinary implementation and verification needing moderate reasoning	`sonnet`	the default agent model
frontier	ambiguous synthesis, cross-cutting design, adversarial judgment, hard trade-offs	`opus`	strongest model configured

Pass the tier via the runtime's per-subagent model control (model on the Agent tool or Workflow agent(), model in a subagent definition's frontmatter, or the OpenCode agent's model field). If a runtime exposes no per-subagent model control, skip this dial and rely on effort alone — never block on it.

Effort — reasoning/thinking budget, where the runtime supports a reasoning-effort or thinking-budget setting: low for mechanical tasks, medium for ordinary work, high only for genuinely hard reasoning (adversarial refutation, ambiguous design). Where the runtime has no effort dial, fold the intent into tier choice.

Defaults by work type (start here, adjust for the specific task):

Work	Tier	Effort
Read-only codebase scouting (specify)	economy	low
Mirror-a-precedent implementer: a `reuses` pointer and no new file, exported type, interface, or class	economy	low
Mechanical `[P]` implementer task with a clear task gate, creating no new file or exported type	economy	low–medium
Ordinary implementer task, most single-verifier runs	standard	medium
Structure-bearing implementer task — creates a new file, or adds an exported type, interface, or class	standard (never economy)	medium
Design/architecture synthesis, plan decomposition for a complex change	frontier	high
Adversarial or cross-cutting verification (the AC-trace / integration verifier, the under-design lens)	standard→frontier	high

Economy tier needs both conditions — a reuses pointer and no new file, exported type, interface, or class. A precedent says what shape to copy, not where a responsibility belongs or which value deserves a type, and those are decided once, when the file or type is created.

Size interacts with tier. A quick change should almost never dispatch a frontier subagent; a complex change earns frontier for its design and adversarial gates but still routes mechanical tasks to economy. When unsure between two tiers, pick the lower and let a failed gate promote it — the same ratchet as sizing.

Universal rules (apply to every phase)

State hygiene

Re-read state.md at the start of every phase. Write size, decisions, and status changes there as they happen, not at the end.

Write surgically. Edit only the target section — replace the content between its ## header and the next ## or end-of-file. Never regenerate the whole file or reorder sections; that silently clobbers append-only or update-in-place sections you did not mean to touch. Section write modes:

Section	Write mode
`## Decisions`	append-only (never edit existing rows)
`## Tasks`	update-in-place (status, evidence)
`## Validation delta`	clear-on-resolve
`## Verification evidence`	append-on-PASS
`## Handoff`	overwrite
`Door` header field (`state.md`)	update-in-place (one-way only; a downgrade needs the user)
`## Conventions observed` (`spec.md`)	overwrite-in-place (regenerated whole on a specify re-run)
`## Environment candidates` (`lessons.md`)	append-only, never loaded

A section with no row here is clobbered by the next surgical write, so adding one means adding its row in the same change.

Honesty

Never claim a task or gate passed without running it. Done means you saw it pass, not that you believe it would.
Record honestly. Open questions, skipped gates, and assumptions stay visible. Never cover a gap with an invented answer.

Hand-off

Each phase finishes, reports, and recommends the next verb — it never runs the next phase. The one documented exception is fix: for a contained correction it re-aligns the artifacts and runs the code delta end-to-end in the same invocation (see references/fix.md); a larger correction still only recommends the phase chain.
/create-rfc and /create-adr are separate skills. Use /create-rfc when a significant decision needs stakeholder alignment; use /create-adr when an architectural choice deserves a standalone record. Reach for either at any point in the flow.