Forge Implement

Forge Implement

Execute approved tasks.md packets; do not plan them. The shared workflow contract and workflow_state.can_enter_stage own approval, freshness, retry, blocking, resume, and check semantics. This Skill owns only implementation-stage orchestration and its evidence.

Gate before every implementation mutation

Start read-only. Before a source-code or implementation-artifact write, obtain the current artifact contents and state, then require all of the following:

• the exact current spec.md revision/hash has its required valid approval;
• the exact current plan.md revision/hash has valid technical/Plan approval;
• tasks.md records correspondence to that exact approved Plan revision/hash;
• any adapter implementation approver policy is supplied to the shared gate decision and is satisfied; and
• neither approval nor a materially used input is stale.

Call the shared implementation gate decision; do not recreate its rules. It rejects missing, stale, unauthorized, self, full-workflow, implement-intent, and post-hoc approvals. “Implement this using Agentic SDLC” and any other full-workflow request are continuation intent, never artifact approval. An adapter may narrow acceptable approvers; it cannot open or weaken a gate.

If any condition is absent, mismatched, unproven, or denied, transition to blocked-at-gate, remain read-only for source and implementation artifacts, and report the exact missing or invalid evidence. Do not begin a packet, choose a workaround, or seek approval after mutation. A recorded prior mutation with later approval is GATE_VIOLATION, not retroactive success.

When every condition holds, transition to implementation-active. Capture the pre-existing working-tree baseline before the first packet: paths, staged or unstaged state, untracked paths, and relevant hashes. Those paths are external scope. Never edit, stage, clean, claim, commit, or deliver them without separate authorization.

Establish execution controls once

Select or confirm one configured commit mode before packet writes. If no mode is configured or authorized, ask for that choice while preserving the gate and working tree; do not silently choose. Apply commit modes exactly. A commit checkpoint is not delivery authorization: never automatically push, squash, open a pull request, or merge.

For each recorded check use exactly PASS, FAIL, or UNMEASURED through the shared check record. UNMEASURED includes a reason and is never passing evidence. Record the baseline separately from failures caused by this work.

Execute closed packets

1. Use the dependency graph and verified state to select a dependency-ready, unblocked packet. Read only its Read / Reference Context, exact write scope, frozen decisions, and verification. Expand context only when concrete evidence contradicts a stated packet fact; record why and what expanded.
2. Keep every edit inside the packet scope and preserve its Must Not Change protections. Implement the frozen solution without research, architecture selection, user interviews, or replanning.
3. Run the packet's narrow verification and only the concern-specific guidance in quality routing. Record every run and result. Do not invent a pass for an unrun or unavailable check.
4. Mark a packet verified only when its required evidence is PASS; apply its selected authorized checkpoint mode. Failed, blocked, or UNMEASURED required work is excluded from every commit.

If evidence shows that the Plan is contradictory or cannot meet a binding requirement, stop that packet and every transitive dependant using the shared blocking semantics. Record the contradiction and route the affected frontier to Planning, Specification, or Clarify as appropriate; do not replan in place. Continue only safe independent packets whose gates, dependencies, and scopes remain valid.

If a required change lies outside the packet, stop the affected packet before that write and request explicit scope authorization. Treat a material change as an approved-Plan/SPEC revision and gate concern when it changes their approved meaning. Do not silently broaden the diff. Apply failure recovery for failures, retries, unavailable checks, and resume.

Integrate once, review once, report compactly

After every packet is verified or terminally failed/blocked, run the approved integrated deterministic verification over the verified scoped result. Record each result exactly; separate pre-existing failures. Then perform exactly one final semantic review against the approved Specification, Plan, packets, scope, and shared contracts. This is distinct from per-packet deterministic verification and mechanical scope/evidence checking.

If that one review finds defects, create a bounded remediation list containing only its findings. Remediate only that list, rerun affected checks and integration when the finding affects integration, and do not run another semantic review. Do not claim complete if required evidence remains FAIL or UNMEASURED.

Write compact evidence.md: artifact revisions/hashes and gate decision; adapter policy result; dirty baseline; selected commit mode and checkpoints; packet status, scoped paths, checks, retries, blocks, and scope decisions; integrated results; semantic-review count/result; remediation; pre-existing failures; UNMEASURED reasons; and delivery actions explicitly not performed. It is a handoff, not a replay of reasoning.