Skip to content

Latest commit

 

History

History
186 lines (135 loc) · 8.6 KB

File metadata and controls

186 lines (135 loc) · 8.6 KB

Reference

This page is a compact guide to the terms and decisions used by implementing-staged-plans. The canonical skill instructions and their linked procedures remain authoritative when a user guide and the implemented workflow differ.

Invocation

The command depends on how the skill was installed:

Installation Invoke with
Codex plugin or standalone skill $implementing-staged-plans
Claude Code plugin /implementation-plugin:implementing-staged-plans
Claude Code standalone skill /implementing-staged-plans

Follow the invocation with a target program or manifest, the action you want, the desired approval mode when creating new state, and any explicit stop.

Core terms

Program

A repository-backed record that ties an authoritative source to requirements, planned increments, approvals, state, workspace, and evidence.

Increment

One bounded, reviewable unit of the program. An increment has its own plan, authority, implementation, review, verification, and acceptance state.

Exact-file plan

The current implementation plan for one increment, including the files and behavior expected to change. It is bound to the current program, workspace, and repository observation rather than acting as a timeless checklist.

Execution baseline

The immutable pre-product-change snapshot of the exact plan, repository observation, path dispositions, and preserved user work. New-program status cannot become authorized until this baseline and its plan-bound action authorization are durable.

Workspace binding

The approved writable repository path, branch, base, current head, and recorded pre-existing work. It prevents work prepared for one tree from being applied to another without revalidation.

Action authorization

A current grant for one named action and scope, such as modifying the workspace, running local verification, or creating a local commit. It is separate from the approval mode.

Evidence

The concrete result that supports a claim: for example, a command result, accepted review artifact, state digest, or repository observation. Evidence is limited to what it actually exercised.

Review packet

The bound collection of review findings, dispositions, verification results, and the exact diff proposed for acceptance.

Handoff

A navigation record for continuing later. It points to the controlling state and evidence but grants no authority on its own.

Reconciliation

The final accounting across every requirement, amendment, increment, finding, deferral, and program-level verification result.

Closure packet

The bound summary presented for explicit program-closure approval after reconciliation succeeds.

Lifecycle at a glance

The implemented order is: Create a New Program, Activate a Generated Program, Before Production Modification, Prepare Review and Diff Disposition, Dispose the Current Diff, Continue an Accepted Program when requested, Authorize a Successor Increment when one is allocated, Resolve a Blocked Program when typed recovery is required, and Close a Final Program when no successor remains. Creation and first-increment behavior remain owned by Plan A. Plan B adds exact accept-continue, the distinct accepted-state-continuation route, status-current successor grants through current_increment_authority_binding, and prompt-bound blocked-recovery. Final programs reuse the unchanged Plan A closure transaction and derive paths from implementation-closure-storage/v1.

Every typed transaction writes controlling status last and adopts only byte-identical prefixes. A divergent prefix stops for recovery without cleanup.

The skill always revalidates current repository facts and controlling records before relying on an earlier state.

Approval modes

Approval modes define routine interruption policy. New-model typed dispositions in version 0.1.2 always offer accept-stop and conditionally offer exact accept-continue for one satisfied successor. Modes do not grant action authority or automatic successor rollover. Legacy approval:full and approval:full-diff modes retain their automatic acceptance behavior.

Mode Scope Routine plan pause Diff acceptance Continuation and mandatory stop
approval:standard One increment Yes User Stops for plan approval, user-owned material decisions, contradictions, hard stops, and diff acceptance
approval:pre-approve One increment No User Stops for user-owned decisions, program amendments, contradictions, hard stops, and diff acceptance
approval:full-increment One increment No User Runs through verification, then stops for diff acceptance unless a hard stop occurs
approval:full-diff Legacy current increment No Legacy policy permits automatic acceptance only after verification and a valid bound packet Cannot begin another increment
approval:full Legacy current increment No Legacy policy permits automatic acceptance only after verification and a valid bound packet Cannot begin another increment automatically; discovery stops at legacy-rollover-upgrade-required

New-program proposal construction, bootstrap, and launch reject approval:full-diff and approval:full before every write. Those modes are dual-read compatibility for already persisted legacy programs only.

When a direct creation request omits a mode, proposal construction defaults to approval:full-increment before producing any bytes. The selected mode is then persisted explicitly. Persisted state with a missing or unknown mode is invalid and is not silently defaulted. An explicit user-requested gate always remains in force, even if the selected mode would normally omit that pause.

Three different kinds of permission

Approval mode sets the routine interruption policy. It answers questions such as whether to pause for the exact-file plan and who accepts the reviewed diff.

Implementation action authorization permits one current program action and scope. Examples include writing a specific program artifact, modifying the bound workspace, running named local verification, or creating a local commit.

Later or external-action authorization covers consequential actions after implementation, such as a pull request, merge, publication, release, deployment, migration, destructive operation, permission change, or provider mutation. The program must be in a suitable state, and the grant must name that exact action.

None of these permissions substitutes for the others. In particular, accepting a diff does not authorize a commit, and closing a program does not authorize a release.

Common hard stops

The workflow stops instead of guessing when it finds:

  • more than one possible program or workspace;
  • a source, plan, approval, status, brief, handoff, or packet digest mismatch;
  • a branch, base, head, path, or pre-existing-work observation that has drifted;
  • an active or conflicted Git operation;
  • a requested transition that is not legal from the current state;
  • missing, expired, revoked, rejected, ambiguous, or mismatched authority;
  • a user-owned decision or program amendment required by the selected mode; or
  • a requested Git, publication, provider, or external action outside the grant.

Typed continuation and blocked recovery have their own exact prompts and retry/recovery routes. legacy-rollover-upgrade-required still quarantines the legacy rollover writer, and blocked-transaction-required still rejects generic direct blocked edges. program-revision-workflow-required and unsupported-program-mutation continue to stop unsupported revision, supersession, cancellation, or other program mutation. Accepted legacy programs and historical terminal records remain readable, but those read paths do not reactivate a mutation sink.

The result should name the failed invariant and return the smallest action that can resolve it. The workflow does not manufacture replacement state to continue.

What can be checked mechanically

The bundled scripts can validate schemas, exact bindings, digests, declared state transitions, file constraints, deterministic packet structure, and specific local command evidence supplied to them.

They do not prove that a reviewer was genuinely independent, a human approval was well informed, a design is semantically correct, a live service behaves as expected, or an external action occurred. Those claims still need evidence from the relevant human or runtime process.

For copyable task prompts, continue with Common workflows. For a blocked run, use Troubleshooting.