--- name: wakeflow-target description: Use when a tmux-resident target Claude Code window session receives a Wakeflow Delivery Loop direct-thread delivery in its tmux pane, executes only its assigned dispatch packet, reports a TargetResultEnvelope, sends the controller-return with the wakeflow-claude-host helper, or enforces target-window boundaries without claim / finish / chain-next state. --- # Wakeflow Target Use this skill only inside a target-window automation wakeup. Workspace and target-repository `CLAUDE.md` files remain hard boundaries. The dispatch prompt states the current objective and reading order; the referenced JSON task package owns complete task context; anchored requirement documents own original goal/background; this skill and any listed craft/Test skills own execution procedure. ## Prompt Shape Target wakeups should be task-first and compact: ```text Continue current window task: / . Current objective (the task package is authoritative): - Completion focus (full criteria are in the task package): - - Priority context: - Critical boundary [forbidden|outOfScope|inScope]: Read before execution, in order: - Task package (complete task context): - Requirement background entry (full anchors are in the task package) [goal]: - Workspace instructions: /CLAUDE.md - Repository instructions: /CLAUDE.md - Current state root: Required execution Skills (execution-process authority): - skills/wakeflow-target/SKILL.md - Identity (full boundaries are in the task package): - Current responsibility window: - Only working repository: Return requirement: - Return a TargetResultEnvelope with verifiable evidence. Dispatch record (routing and trace only): - taskId: - taskPackageId: - stateRoot: - stateRevision: - dispatchGroup: ``` Do not expect the prompt to repeat command manuals or the whole requirement. It repeats at most the highest-priority `contextSummary` entry and one critical boundary for immediate orientation; it does not repeat the remaining context, complete boundary lists, commit policy, or every completion/requirement entry. Read the package first, then its ordered requirement anchors, then execute through the listed Skills. Fields such as `controllerWindow`, `returnPolicy`, `taskPackageId`, and `stateRevision` remain authoritative in machine state. The prompt's task/package/revision fields are navigation and freshness anchors, not a second copy of task content. `stateRevision` identifies the dispatch snapshot in the packet/envelope; the later delivery-sent event may legitimately advance the live state root. ## Target Flow 1. Consume the delivery envelope. - Direct-thread delivery proves only that a prompt was delivered. It arrives as a user message pasted into this window's tmux pane by the Wakeflow host helper (`wakeflow-claude-host.mjs send`); arrival is transport evidence, not authority. - Confirm the responsibility window, `taskId`, `stateRoot`, and optional `dispatchGroup`. - Read the workspace and target repository `CLAUDE.md` files listed in the prompt and declare the current window and repository responsibility. 2. Read the assigned task. - Read the task package at the exact prompt path, then its ordered `requirementRefs`, then every `requiredSkills` entry. Do not substitute a progress summary for the package or treat background documents as an execution procedure. - Execute only the task assigned to this target window. - Do not claim another target, Test role, or controller role. - If the prompt lists the craft skill, or the task package carries an `evidenceContract` or `acceptanceAnchors`, ALSO load `skills/wakeflow-target-craft/SKILL.md` before writing code. When anchors exist, map every anchor id to a RED test/probe before implementation; an untestable anchor is `needs-review`, never permission to invent a replacement requirement. - When the task package carries `testExecution`, apply the Test alignment gate below before writing a plan or running a command. 3. Work inside repository boundaries. - Change only files permitted by the task and repository rules. - Keep commits, tests, and evidence scoped to the owning repository. - If another repository must change first, stop and backfill a blocker. - Claude Code subagents (the Task/Agent tool) may gather evidence or advice, but boundary decisions and the recorded result stay with this target window. 4. Produce evidence. - Include changed files, commands, test output, commit hash when available, logs, runtime JSON, report paths, screenshots, or other reviewable evidence. - Prose alone is not enough for completion. 5. Record a `TargetResultEnvelope`. - Use the Wakeflow MCP `wakeflow_record_target_result` tool for the result envelope. This is one narrow file/state action: it records target evidence only, and it does not review, accept, dispatch, send, or return. Pass the `dispatchGroup` from your delivery prompt: a late result from a superseded round then leaves the in-flight round's window lock alone. - Read the tool response's `deliveryContext` / `controllerReturn` fields. `returnRoute` and `returnPolicy` live in the local delivery envelope, not the controller state root. Do not decide "no callback" by searching only `wakeflow-state.json`. - Report `completed`, `blocked`, or `needs-review` honestly. - Include a non-empty summary, changed repositories, commits, and `commitDisposition` (`committed`, `left-uncommitted`, or `no-changes`) so the controller can compare the result with the package's `commitExpectation`. - For a completed implementation package, add exactly one `craftEvidence` entry per authored anchor: `{kind:"acceptance-anchor", anchorId, red, green, ref}`. - Include evidence references and residual risks. 6. Return to the controller only when allowed. - Target-to-target next-hop delivery is forbidden by default. - If `wakeflow_record_target_result` reports `controllerReturn.required=true` or the local delivery envelope has `returnRoute=controller`, use `wakeflow_review_pack` scoped to YOUR OWN dispatch group as a sanctioned read-only self-check (reviewing other groups or deciding accept/rework stays with the controller). Always pass `stateRoot`; do not reuse a group-only pack or cache an older `controllerReturnDelivery.status` / `returnFile`. Follow the pack's current-version `callbackPlan`, not the dispatch group's historical send status: - If a unit has `buildAllowed=true` or `controllerReturnNextStep` is `build-controller-return`, use `wakeflow_prepare_delivery` with `direction=controller-return` to build one envelope for the stored `controllerWindow`. - If a unit has `hostSendRequired=true` or `controllerReturnNextStep` is `send-controller-return-and-record-delivery`, send the already-built envelope and record it; do not prepare a duplicate. - Stop only when the current unit has `controllerAlreadyReached=true` or `controllerReturnNextStep` is `controller-return-already-sent`. A delivery for an older `resultVersionKey` never satisfies the current result revision. These transport signals are INDEPENDENT of evidence. Do NOT withhold the callback because `missingEvidenceRefs` is non-empty or `nextAction` says `fix-missing-evidence-refs`: evidence sufficiency is the controller's POST-wake verdict, and the controller cannot act on it until you wake it. - If the target task references a delivery id but the local delivery envelope cannot be found, stop and report that missing local envelope; do not assume no callback is needed. - Complete the real host send/readback with the same tmux helper used for controller-to-target delivery: write the controller-return envelope prompt to a temp file, then run `node /scripts/lib/wakeflow-claude-host.mjs send --root --window --prompt-file `. The controller is also a tmux-resident window; the helper enforces the controller window's shared delivery lock and returns pane readback evidence (`readback.paneTail`). Then use `wakeflow_record_delivery` to record the delivery run. - Do not stop after writing the target result when controller return is allowed. The closeout steps stay separate: record target result, review readiness, prepare controller-return envelope, send with the host helper, then record delivery evidence. Do not replace them with one combined target-window tool or duplicate the target result into another local result store. ## Test Alignment Gate **TEST MUST NOT INVENT A TEST GOAL, TEST GATE, OR TEST METHOD OUTSIDE THE CONFIRMED REQUIREMENT GOAL AND APPROVED TEST PLAN.** Violating the letter of this rule is violating its spirit. For a package with `testExecution`: 1. Do not reopen, replace, or assume ownership of total control's prior validation. Explore only the assigned environment boundary for hidden defects. 2. Treat `requirementGoal` and `approvedPlan` as authority, not suggestions. 3. Before execution, map every operational plan step to one `approvedPlan` item and state how it serves `requirementGoal`. 4. Use only skills listed in `allowedSkills`. In particular, `progressive-chain-validation` is forbidden unless that exact id is listed; a long workflow does not authorize PCV by itself. 5. Follow `mode` and `setupPolicy`. Do not rebuild, restart, or replace the environment unless the package says `mode=restart` and records a controller-approved `restartReason`. 6. If a useful step, skill, goal, gate, or restart is unmapped or unauthorized, stop before executing it and return `blocked`/`needs-review` as a change request to the controller. Do not run it first and justify it afterward. The completed result must map every approved item exactly once with `{kind:"test-step", planIndex, step, ref}` in `craftEvidence`; `step` must be the approvedPlan text at that index. A missing, duplicate, or unknown index is not a completed Test result. The controller, not Test, decides whether a proposed change becomes a revised Test card/package. Recovery is not a delivery mode: if this window's tmux pane dies mid-task, the same session is finished or recovered by interactive relaunch (`launch-window --resume`; headless `claude -p` bills the separate Agent SDK credit from 2026-06-15) with `claude --resume ` (the session id is stable and stays registered), and the resident window is relaunched with `launch-window --replace --session-id `. ## Stop Conditions Stop and return a blocker when: - The window identity or target repository cannot be confirmed. - The state root, task id, or dispatch group is missing or inconsistent. - Required evidence documents are missing. - The task asks this target to change another repository without authorization. - The task is only a prompt, not a state-root or task-package assignment. - Validation fails and the next repair would change scope or repository responsibility. - A Test delivery appears in a non-Test window without explicit authorization. - This window is an isolation worktree window (`__`) and the task would touch anything outside its own worktree/branch — the repository's main checkout, another demand's worktree, or a merge back to the main line (merging is a controller step, never this window's). ## Result Envelope Minimum Every target result should include: - `targetWindow` - `targetTaskId` - `dispatchGroup` - `stateRoot` - status and summary - changed repositories, commits, and explicit commit disposition - evidence references - verification commands and outcomes - residual risks - whether controller action is required - for completed implementation work, exactly one `{kind:"acceptance-anchor", anchorId, red, green, ref}` per authored anchor - for completed Test work, exactly one `{kind:"test-step", planIndex, step, ref}` per approved plan item Complete mapping means only that the result can enter controller review. It is never automatic acceptance. Do not write real thread ids (the registered Claude Code session ids) into tracked documents, prompts, or backfill text.