Execute planned or planless work

Execute the selected work through its existing plan. An explicit current instruction may instead authorize one planless quick slice: an understood canonical feature story, or a contextual instruction with no story or plan. The coordinator owns delivery; implementation agents return uncommitted changes.

Establish execution context

Identify the execution source before changing project state:

• Planned: require an executable plan. For a feature story, read its seed section. For a bounded correction, require its complete correction input — its story and plan, or a plan-homed correction's plan — under planning scope and lifecycle.
• Quick: require an explicit current instruction that authorizes planless execution. The source is either a canonical feature story with understood goal, scope, key examples, and no blocking decision, plus the instruction to skip slice planning; or a contextual instruction that supplies the objective, known expectations, remaining uncertainty, and authority. A seed alone or apparent story size cannot select this path. Corrections require plans. In Story Branch or Trunk Mode, first admit an accepted independent mission; otherwise create no story, plan, or queue entry.

Name missing authority or source context and stop before backlog changes, delegation, or implementation. Successful quick execution keeps its scope in its story, when admitted, and its decisions, progress, and proof in the conversation; create no plan, completion note, or substitute record. A contextual instruction may leave expectations unresolved. Carry its goal and uncertainty, resolve what a behavior change requires, and allow an evidenced no-change conclusion through the explained-empty-change path.

Use planning level for source ownership, proof ownership when mapping or accepting proof, and active-plan refinement for plan updates. Judge proof during execution; retain completed plans, source history, and proof decisions for retrospective and story wrap-up. Quick work retains its source, conversation, and execution identity. Use slice decomposition for Behavior/Structure types and slice sizing for budgets and learning escalation.

At startup, obtain enough authoritative context to select and delegate the first slice. Reuse instructions while their sources and applicability assumptions remain unchanged. Resolve project context at the first boundary that needs it:

• execution-source kind, slice target, hard limit, and exceptions; replanning permission; integration checkout, when one exists, and branch for a claim, using the project's configured integration branch or main when none is supplied; execution mode and location: default Story Branch Mode; --trunk or a clear equivalent selects Trunk Mode; explicit caller selection uses the current branch. Resolve contradictions before changing state. Mode never creates execution authority; for planned work, plan path and status vocabulary;
• backlog path and selected entry for work selected from Backlog list;
• selective formatter and commit hook contract before taking or admitting work and its claim commit;
• navigation, focused tests, runtime wrapper, and workflow precedence for the selected slice;
• authorized push destination before delivery; for Story Branch or Trunk Mode, also trunk publication's Preconditions before selecting the owned workspace and taking or admitting work, and before publishing a claim, validated increment, or owned repair — those preconditions resolve publication inputs and defer shared-checkout preservation to maintain the default checkout;
• generation triggers and commands when affected; and
• refactor context before refactor delegation.

Missing context stops its affected boundary, including first-slice delegation when needed there. Other invoking tools, including GSD, retain this slice delivery contract; a phase or task cannot replace a slice.

Before first implementation delegation, read delegation and the common reassessment/human-judgment rules, replanning permission, plus currently triggered sections of execution decisions. Before accepting a return, read proof acceptance; before delivery, read delivery. Before arming observation, read CI observation and only the current host's notification adapter. Arm from the execution checkout against the authorized target branch; do not wait for CI. On Codex, arm the yielded stream at execution start. Before creating the execution workspace, read execution location. Before a claim, validated increment, or owned repair publication, read trunk publication. Use targeted retrieval and disposable research for omitted/truncated passages or bounded investigations; another step alone needs no reload.

Take or admit work

After resolving execution source and authority, inspect the backlog before plan-status changes, observer startup, delegation, or implementation. Resolve the selective formatter and claim commit hook contract. An absent or understood check-only hook permits the transition; an unknown, mutating, failing, or disputed hook stops with the queue unchanged. Resolve publication preconditions, including the authorized remote/trunk, before a Story Branch or Trunk Mode startup that publishes a claim. Existing current-branch and host-owned checkout restrictions still apply.

When your instruction carries an established start, follow established start instead of the start command below.

For authorized queued Story Branch or Trunk Mode work, invoke the installed scripts/execution-start.mjs start once with the originating integration checkout as --integration when one exists. Without one, an existing owned workspace supplies repository access; to create a new workspace, supply an owned worktree of the repository or its common Git directory as --repository, which is only read from, never refreshed. Also supply the owned workspace path and branch, selected identity, stable execution publisher ID, mode (trunk or story-branch), actual remote and trunk branch, and the established --push-authorized --workspace-authorized flags. Supply your own --host (claude, codex, or cursor) and --model; omit either you cannot state rather than guess. Supply --plan as a path relative to the backlog directory when explicitly selected; the command also resolves the canonical published plan. The command fetches trunk, checks the published selected source and preparation, selects the workspace or reuses it under refresh eligibility, names you as an agent, commits an isolated Take that publishes your agent profile, makes that agent the author of your workspace commits (workspaceAuthorship: "not-configured" means only the Take commit names the agent), confirms publication on remote trunk, publishes a Story Branch Mode execution branch at that Take so the branch the profile names exists on the remote, and reports local refresh separately.

Use its compact one-line result directly; do not filter or fetch it again. Values you supplied stay in your execution context and are not echoed. An accepted result (ok: true, status published or resumed) carries publishedSha, startingRevision, candidateSha, created, agent when the claim names one, plan or remote when resolved rather than supplied, and the default checkout's maintenance (result, plus reason when not refreshed). Retain them; the first increment's managed delivery uses publishedSha as its previously published base. existing (work already Taken under your claim) writes nothing and returns that claim's publishedSha. No maintenance result or earlierMaintenance issue undoes accepted publication. A refusal or unconfirmed result (ok: false, non-zero exit) stops before implementation; report and act on its status, error, and any recovery or provenance, handling developer-identity-refused as under agent commits. Inspect current Git state only when a reported reason needs it; never repeat a mutating command to obtain diagnostics. If publication is interrupted, invoke the same installed command with the retained workspace, branch, publisher ID, identity, --starting-revision and --candidate-sha from the last result (or its recovery) or confirmed pre-push candidate. Use the latest candidate SHA after a replay. A resumed result confirms current ownership through remote ancestry, even when trunk has advanced; it may finish eligible local refresh without another Take or push. A rival or ambiguous provenance stops implementation. Preserve the stopped candidate and exact recovery fields on an uncertain result.

Every accepted start, new or resumed, then requires this project's checkout-bound setup and applicable command under execution location before implementation. Setup failure preserves the accepted claim and workspace. The separate product-backlog Take tool keeps its local domain purpose; a local Taken entry alone does not satisfy this startup boundary. Queued current-branch work keeps its existing local Take contract and gains no publication authority. Leave taken work through pauses, failures, completion, and retrospective; wrap-up removes it.

Admit accepted work that no backlog list holds

When the current instruction accepts a mission that no backlog list holds, follow admit accepted work, which admits it with this start command and --admit before its substantive work. Explicitly selected one-shot work starts with --one-shot instead.

Choose the execution location

Follow execution location for mode, workspace creation, project-command readiness, reuse of host-established preparation for the selected checkout, retained identity, execution resume, push destination, and checkout-bound runtime. That reference applies the shared checkout ownership lifecycle for selection, local checkout role, and target selection.

Continue or recover at an execution boundary

Planned work uses the existing plan and conversation under execution and resume state. Quick work uses its source and conversation, without a recovery artifact. Retain the resolved replanning permission with that context. During uninterrupted work, reuse decisions, accepted proof, and delivery progress while their sources, assumptions, and covered boundaries hold. After confirmed push, obtain only newly relevant next-slice detail; a slice transition alone needs no full recovery read.

After interruption or a changed observation from Git, agents, or observer coverage, verify execution identity first and reconcile only affected worktree/index, branch/commits, implementation/refactor return, and exact observer identity. Preserve unrelated or ambiguously owned work. Reuse proof only while promise, boundary, implementation, setup, and observations match.

Resume at the first delivery obligation not established by evidence. An incomplete or oversized return still needs oversized-slice handling before proof acceptance. Otherwise implementation returns still need proof acceptance/refactoring; completed refactors need remaining delivery; uncommitted plan edits need staging/commit. Classify an interrupted claim, increment, or repair with interrupted publication before any further commit or push. Plan status or a compact report proves none of those later boundaries. When pushed commit, retained delivery result, and required registration agree, select the next unfinished slice in plan order. Missing/contradictory execution identity requires the recovery decision above.

Execute the next slice

1. In the execution checkout, use the plan's current statuses, decisions, learnings, proof, and selected existing-solution finding/evidence when present. Read retained state on initial entry/recovery; reuse valid readings during continuation. For quick work, reread the source and conversation's scope, decisions, progress, proof, and remaining uncertainty; it is the only slice, with no separate state artifact. Confirm current execution authority. Recover an existing CI observer before considering a new one.
2. Select the next unfinished planned slice in plan order, or the one quick slice. Apply execution decisions; for behavior/state removal or disablement, also run the destructive later-outcome check.
3. When planned refinement is needed and learning escalation permits, invoke slice-plan refinement in place, then restart at step 1, unless replanning is disabled; then apply oversized-slice decisions and stop without retry. If quick work no longer fits one coherent slice, apply oversized-slice decisions. A no-replan return stops without planning or retry. When replanning is allowed, use ordinary slice planning for remaining work, and restart as planned execution. Before delegating a change that invalidates a required pre-change observation, apply proof ownership: reuse an adequate baseline with known matching revision/environment/selection conditions, or obtain it first. Missing/failed prerequisites stop only dependent work. Apply on entry and resume without a new startup audit or repeated recovery read. Otherwise delegate under delegation.
4. On return, recheck execution decisions; handle incomplete/oversized work there before delivery, including a no-replan overrun. Otherwise accept proof and confirm uncommitted work or an explained empty change.
5. Run delivery end to end. That delivery publishes through increment and repair publication. After successful delivery, restart for the next planned slice, one at a time: each finishes its delivery before the next starts; a quick slice has no successor.

Finish or stop

Follow finish or stop for the execution-complete record, the completion operation, reporting and its markers, automatic retrospective invocation, and incomplete-work reporting.