| name | build |
| description | End-to-end feature workflow (requirements -> planning -> implementation -> release) in a single command. |
| allowed-tools | Bash(echo *), Bash(rp1 *) |
| metadata | {"category":"development","is_workflow":true,"workflow":{"run_policy":"resumable","identity_args":["FEATURE_ID"]},"version":"3.0.0","tags":["core","feature","orchestration"],"created":"2025-12-30T00:00:00.000Z","updated":"2026-05-05T00:00:00.000Z","author":"cloud-on-prem/rp1","arguments":[{"name":"FEATURE_ID","type":"string","required":true,"description":"Feature identifier (kebab-case)"},{"name":"REQUIREMENTS","type":"string","required":false,"default":"","description":"Raw requirements text","variadic":true},{"name":"PHASE_PLAN_PATH","type":"string","required":false,"default":"","description":"Optional phase-plan artifact path for child-feature traceability"},{"name":"PHASE_ID","type":"string","required":false,"default":"","description":"Optional parent phase identifier for child-feature traceability"},{"name":"AFK","type":"boolean","required":false,"default":false,"description":"Non-interactive mode","aliases":["afk","no prompts","unattended"]},{"name":"GIT_COMMIT","type":"boolean","required":false,"default":false,"description":"Commit changes after build"},{"name":"GIT_PUSH","type":"boolean","required":false,"default":false,"description":"Push branch to remote","implies":["GIT_COMMIT"]},{"name":"GIT_PR","type":"boolean","required":false,"default":false,"description":"Create PR after build","implies":["GIT_PUSH","GIT_COMMIT"]}],"sub_agents":["rp1-dev:feature-requirement-gatherer","rp1-dev:feature-architect","rp1-dev:hypothesis-tester","rp1-dev:feature-tasker","rp1-dev:task-builder","rp1-dev:task-reviewer","rp1-dev:code-checker","rp1-dev:feature-verifier","rp1-dev:comment-cleaner","rp1-dev:build-verify-aggregator","rp1-dev:feature-archiver"]} |
Build Command
YOU ARE A PURE ORCHESTRATOR. Spawn agents for all work. NEVER write/edit/read files yourself. NEVER implement code, requirements, designs, or tests. Use exact agent references per step. If agent fails, retry it — never do its work.
§CTX
Use the pre-resolved projectRoot, kbRoot, workRoot, and codeRoot values from the generated Workflow Bootstrap section. Do not hardcode .rp1/work/ or .rp1/context/ paths.
Feature dir: {workRoot}/features/{FEATURE_ID}/
FEATURE_ID slug: FEATURE_ID names the feature directory and is the resume key, so it MUST be a short kebab-case slug. When you run the §0 bootstrap, make the first --args token a clean slug:
- If the user already gave a kebab-case slug, pass it unchanged.
- Otherwise derive a short slug (3-6 words, lowercase, hyphens only — no slashes, spaces, or file extensions) from the request and pass the rest after it:
--args "<slug> <remaining request>". The leading token resolves to FEATURE_ID; the remainder resolves to REQUIREMENTS.
- Never pass a file path, URL, or full sentence as the slug. If the request points at a doc (e.g. a research file), derive the slug from its subject, not its path.
The bootstrap sanitizes FEATURE_ID to a safe slug as a fallback, but derive a meaningful one yourself — the fallback slug of raw prose is ugly. Your /build {FEATURE_ID} resume instruction surfaces the chosen slug for the user.
References
| File | Purpose | When to Load |
|---|
references/build-redirected.md | Oversized-scope redirect handling when feature-architect returns needs_phase_planning | When feature-architect returns status = "needs_phase_planning" |
references/parallel-builders.md | Worktree lifecycle protocol for parallel-wave concurrent builders | When parallel-wave mode preconditions are met during section 4.3 |
§0-FIRST-ACTION
After the generated Workflow Bootstrap section resolves RUN_ID, RUN_RESUMED, and the canonical directories, the first prompt-authored action MUST be:
rp1 agent-tools workflow-state \
--run-id {RUN_ID} \
--workflow build \
--feature {FEATURE_ID} \
--parent-phases requirements,planning,implementation,release
Do NOT read files, load KB, analyze requirements, or spawn agents before this completes.
Parse the JSON ToolResult.
- If the tool fails or returns malformed output: emit
requirements failed, report the tool error, STOP.
- If
data.summary.contract_gaps is non-empty: choose the first gap by phase order, emit waiting_for_user and status_change waiting on that gap phase, report missing registered outputs, STOP. Do not inspect feature files or infer success from filenames.
- Set
WORKFLOW_STATE = data.
- Set
START_PHASE = data.summary.next_phase.
- If any
WORKFLOW_STATE.phases[] entry has status = "waiting" and there are no contract gaps for that phase, set WAITING_PHASE to the earliest waiting parent phase and set START_PHASE = WAITING_PHASE.phase.
- When resuming a
WAITING_PHASE, return to that phase's recorded checkpoint/decision handler. Do not rerun that phase's producer agents unless the resumed decision is Revise, Add Task, Repair, or another explicit update path.
- If
START_PHASE is null: output an already-complete summary from registered workflow state and STOP.
- Initialize
PLANNING_UPDATE_CONTEXT = "", TASK_REGENERATION_REASON = "", and ARCHIVE_RETRY_PATH = "" unless restored from a resumed checkpoint event.
- Phase order:
requirements -> planning -> implementation -> release.
STATE-MACHINE
stateDiagram-v2
[*] --> requirements
requirements --> planning : requirements_accepted
requirements --> requirements : requirements_revised
requirements --> [*] : stopped
planning --> implementation : plan_accepted
planning --> planning : plan_revised
planning --> [*] : oversized_or_stopped
implementation --> implementation : add_task_or_repair
implementation --> release : readiness_ready
implementation --> [*] : unrecoverable_failure
release --> implementation : add_task
release --> release : archive_chosen
release --> [*] : release_complete
§PARENT-EMIT-DISCIPLINE
Only parent steps are requirements, planning, implementation, and release.
Parent status events are limited to:
running: broad phase starts or resumeswaiting: user decision or contract gapcompleted: broad phase is accepted/readyfailed: no planned recovery remains
Subagents emit namespaced detail steps (task-builder:building, task-reviewer:reviewing, etc.). Retryable subagent failures keep the parent phase running or waiting; they MUST NOT emit parent failed.
Before executing each non-skipped phase, emit running for that phase. First emit (entering START_PHASE) includes --name to label the run:
rp1 agent-tools emit \
--workflow build \
--type status_change \
--run-id {RUN_ID} \
--step {STATE} \
--name "Feature: {FEATURE_ID}" \
--data '{"status": "running", "feature": "{FEATURE_ID}"}'
Subsequent routine transitions use the same command without --name (set-once semantics; the DB keeps the first value). Replace {STATE} and the data JSON per this table:
| Trigger | Step | Data |
|---|
| Phase entry (subsequent) | {phase} | {"status": "running", "feature": "{FEATURE_ID}"} |
| Phase accepted / AFK continues | {phase} | {"status": "completed", "feature": "{FEATURE_ID}"} |
| User stops at checkpoint | {phase} | {"status": "waiting", "feature": "{FEATURE_ID}"} |
| Release: no archive | release | {"status": "completed", "feature": "{FEATURE_ID}", "archive_status": "declined"} |
| Release: archived | release | {"status": "completed", "feature": "{FEATURE_ID}", "archive_status": "completed", "archive_path": "..."} |
| Unrecoverable failure | {failing step} | {"status": "failed", "feature": "{FEATURE_ID}"} |
Emit blocks with additional contextual data (reason, task_unit, readiness_status, etc.) remain verbatim at their point of use below. waiting_for_user, artifact_registered, and end-run emits are always shown inline.
RUN_ID comes from the generated Workflow Bootstrap section. Do NOT override it.
Producer agents register their own artifacts. Do not scan feature directories to register markdown files.
§PROGRESS
| Phase | Owns | Agent(s) |
|---|
| requirements | Requirements artifact + scope redirect handoff | feature-requirement-gatherer |
| planning | Design, hypotheses, task generation | feature-architect, hypothesis-tester (opt), feature-tasker |
| implementation | Task execution, reviews, checks, feature verification, comment cleanup, readiness aggregation | build-task-plan tool, task-builder, task-reviewer, code-checker, feature-verifier, comment-cleaner, build-verify-aggregator |
| release | Manual checklist, archive choice, final run closure | feature-archiver |
Symbols: [ ]=PENDING [~]=RUNNING [x]=COMPLETED [-]=SKIPPED [!]=FAILED
Requirements/planning fail fast on unrecoverable contract failures. Implementation retries recoverable builder/reviewer failures once before waiting (interactive) or failing per AFK policy. NEVER delete artifacts.
AFK mode: skip prompts, auto-select defaults, retry once on failure, auto-archive.
§CHECKPOINT-OPTIONS
AskUserQuestion shows at most 4 options. At every checkpoint:
- Present that checkpoint's canonical options verbatim — never rename, merge, or drop one to make room for another choice.
Review feedback from Arcade and Stop must always be offered.
- If an agent's result adds a decision the user must make (e.g. a scope choice the gatherer surfaced), or the canonical list exceeds four, do not fold it into the menu. Split into sequential
AskUserQuestion calls: present the canonical proceed/stop menu first, then ask the surfaced or overflow choice after the user opts to proceed.
§PHASE-1: Requirements
Skip if: START_PHASE is after requirements.
Resume checkpoint: If WAITING_PHASE.phase == "requirements", jump directly to this phase's Checkpoint options below. Do not dispatch feature-requirement-gatherer unless the resumed decision is Revise.
Spawn agent — do NOT gather requirements yourself:
{% dispatch_agent "rp1-dev:feature-requirement-gatherer" %}
FEATURE_ID={FEATURE_ID}, REQUIREMENTS={REQUIREMENTS}, AFK_MODE={AFK}, PHASE_PLAN_PATH={PHASE_PLAN_PATH}, PHASE_ID={PHASE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
If PHASE_PLAN_PATH and PHASE_ID were passed explicitly, forward them unchanged.
If phase-plan handoff tokens remain embedded inside REQUIREMENTS using the legacy PHASE_PLAN_PATH=... PHASE_ID=... form, leave them untouched so feature-requirement-gatherer can normalize them before writing requirements.md.
Validate: parse JSON; accept status: "success" with artifact path ending in features/{FEATURE_ID}/requirements.md, or text Requirements completed: with matching path. status: "error" = intentional failure (abort, no retry). Mentions of commits, code edits, tests, or implementation = contract failure: retry once with scope reminder. Second failure = abort.
Checkpoint (skip if AFK):
Emit waiting_for_user on requirements with prompt "Continue, Revise, Review feedback from Arcade, or Stop?" and context "Requirements gathering complete".
{% ask_user "Continue, Revise, Review feedback from Arcade, or Stop?", options: "Continue", "Revise", "Review feedback from Arcade", "Stop" %}
If feature-requirement-gatherer surfaced a decision needing your input (e.g. a scope choice), keep these four options verbatim; on Continue, ask the surfaced decision as a separate AskUserQuestion before entering planning (§CHECKPOINT-OPTIONS). Never displace Review feedback from Arcade or Stop.
On Revise: get feedback, append to REQUIREMENTS, re-invoke §PHASE-1.
On Review feedback from Arcade: load arcade-collab skill, process all feedback for RUN_ID, then return to this checkpoint with original options.
On Stop: emit requirements waiting per §PARENT-EMIT-DISCIPLINE table, output summary, exit with /build {FEATURE_ID} resume instruction.
On Continue, or immediately when AFK skips the checkpoint, emit requirements completed per §PARENT-EMIT-DISCIPLINE table before entering planning.
§PHASE-2: Planning
Skip if: START_PHASE is after planning.
Resume checkpoint: If WAITING_PHASE.phase == "planning", inspect the latest parent waiting/status event from WORKFLOW_STATE.recent_events:
reason = "rejected_hypotheses" -> jump directly to §2.2 Hypothesis Gate.- otherwise jump directly to the §2.3 planning Checkpoint.
(An oversized-scope redirect terminates the run as cancelled, so a redirected build never resumes here — re-invoking /build starts a fresh run. See references/build-redirected.md.)
Do not dispatch feature-architect, hypothesis-tester, or fresh feature-tasker on a waiting resume unless the resumed decision is Revise.
Spawn agent — do NOT design yourself:
{% dispatch_agent "rp1-dev:feature-architect" %}
FEATURE_ID={FEATURE_ID}, AFK_MODE={AFK}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, UPDATE_MODE={design.md exists}, UPDATE_CONTEXT={PLANNING_UPDATE_CONTEXT}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
Parse the response as JSON.
- Accept
status = "success" to continue with design follow-on work.
- Accept
status = "needs_phase_planning" as an oversized-scope redirect. Read references/build-redirected.md and follow the redirect handling procedure. Do NOT run hypothesis-tester, feature-tasker, or enter implementation.
- Treat
status = "error" or malformed output as a planning failure. Abort the build instead of guessing.
After a success response, check whether {workRoot}/features/{FEATURE_ID}/hypotheses.md exists on disk. If it exists:
{% dispatch_agent "rp1-dev:hypothesis-tester" %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
If the file does not exist, skip hypothesis validation regardless of flagged_hypotheses or artifacts.hypotheses in the response.
§2.2 Hypothesis Gate
If hypothesis-tester ran, inspect its response before task generation:
- If it reports an error, malformed rejection JSON, or cannot determine rejected hypotheses safely: abort on
planning.
- If it reports completion/no pending hypotheses and no rejected block: continue.
- If it includes JSON with
type = "rejected_hypotheses" and non-empty hypotheses[]: do NOT run feature-tasker yet.
- Treat a rejected hypothesis as high impact when
impact = "HIGH", risk = "HIGH", or the field is missing/unknown.
If rejected hypotheses exist and AFK=true:
- If any rejected hypothesis is high impact, emit
planning failed with reason = "rejected_high_impact_hypothesis" and STOP.
- Otherwise continue with risk, but include rejected IDs in the final planning summary. Do not silently hide the risk.
If rejected hypotheses exist and AFK=false, run the interactive rejection gate:
Emit waiting_for_user on planning with prompt "Rejected planning hypotheses found. Revise plan, Continue with risk, or Stop?" and context about paused task generation. Then emit planning waiting with reason: "rejected_hypotheses".
{% ask_user "Rejected planning hypotheses found. Revise plan, Continue with risk, or Stop?", options: "Revise plan", "Continue with risk", "Stop" %}
- Revise plan: collect feedback, set
TASK_REGENERATION_REASON = "Rejected hypotheses: {ids}; revision requested: {summary}", set PLANNING_UPDATE_CONTEXT = TASK_REGENERATION_REASON, emit planning running with task_regeneration_reason and update_mode: true, then re-invoke §PHASE-2 before any task generation.
- Continue with risk: proceed to the single normal
feature-tasker dispatch below and preserve the rejected IDs in the final planning summary.
- Stop: output the rejected hypothesis IDs and
/build {FEATURE_ID} resume instruction, leave planning waiting, and STOP.
§2.3 Task Generation
Normal fresh path invariant: dispatch feature-tasker exactly once, after feature-architect succeeds and after the hypothesis gate is either skipped, clear, or explicitly continued with risk.
{% dispatch_agent "rp1-dev:feature-tasker" %}
FEATURE_ID={FEATURE_ID}, WORK_ROOT={workRoot}, UPDATE_MODE=false, UPDATE_CONTEXT={TASK_REGENERATION_REASON}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
Validate: parse JSON; accept status: "success" with feature_id, task_plan_path, and artifacts[] for both tasks.md and tasks.json (storageRoot: "work_dir"). status: "error" = abort planning; do NOT enter implementation or release. Malformed/missing artifacts = failure. Do not continue without confirmed results.
Checkpoint (skip if AFK):
Emit waiting_for_user on planning with prompt "Continue, Revise, Review feedback from Arcade, or Stop?" and context "Design and task generation complete".
{% ask_user "Continue, Revise, Review feedback from Arcade, or Stop?", options: "Continue", "Revise", "Review feedback from Arcade", "Stop" %}
On Revise: get feedback. If the feedback changes scope, requirements, assumptions, or design, set TASK_REGENERATION_REASON to one sentence before regeneration, emit planning running with task_regeneration_reason and update_mode: true, set PLANNING_UPDATE_CONTEXT = TASK_REGENERATION_REASON, re-invoke §PHASE-2, and dispatch feature-tasker with UPDATE_MODE=true and UPDATE_CONTEXT={TASK_REGENERATION_REASON}. Do not regenerate tasks before the reason is recorded.
On Review feedback from Arcade: load arcade-collab skill, process all feedback for RUN_ID, then return to this checkpoint with original options.
On Stop: emit planning waiting per §PARENT-EMIT-DISCIPLINE table, output summary (requirements complete, planning waiting), exit with /build {FEATURE_ID}.
On Continue, or immediately when AFK skips the checkpoint, emit planning completed per §PARENT-EMIT-DISCIPLINE table before entering implementation.
§PHASE-3: Implementation
Skip if: START_PHASE is after implementation.
Resume checkpoint: If WAITING_PHASE.phase == "implementation", inspect the latest parent waiting/status event from WORKFLOW_STATE.recent_events:
reason = "review_retry_exhausted" -> resume the repair/skip/stop decision before any new builder dispatch.reason = "readiness_add_task" or reason = "release_add_task" -> run §4.1 against the updated tasks.json, then continue implementation from remaining task units.reason = "missing_readiness_contract" -> jump to §4.6 verification/readiness.- otherwise jump directly to the Implementation checkpoint if a registered
features/{FEATURE_ID}/build-readiness.md artifact exists.
Do not dispatch task-builder, validators, or comment-cleaner before the matching resumed decision path is selected.
You MUST spawn task-builder — do NOT write code yourself.
§4.1 Plan Task Units
Use the schema-backed task plan sidecar. Do not parse tasks.md for machine planning.
rp1 agent-tools build-task-plan \
--tasks-path "{workRoot}/features/{FEATURE_ID}/tasks.json" \
--max-simple-batch 3 \
--complex-isolated true
Parse the JSON ToolResult.
- Extract
task_units, implementation_tasks, documentation_tasks, and warnings.
- Set
TASK_PLAN = data.
- For each
task_unit, derive TASK_UNIT_IDS = task_unit.task_ids.join(","). This is the only source for builder/reviewer TASK_IDS.
- Preserve
warnings as task_plan_warnings for readiness/release notes.
- If the tool fails or returns malformed output, emit
implementation waiting with the tool error and STOP. Do not infer task state from markdown.
- If
task_units is empty, skip task-builder/task-reviewer and continue to documentation follow-ups, cleanup manifest handling, and verification.
§4.2 Cleanup Manifest Baseline
Before the first task-builder unit, snapshot the build-start repository state:
rp1 agent-tools change-manifest snapshot \
--code-root "{codeRoot}" \
--out "{workRoot}/features/{FEATURE_ID}/change-manifest-baseline.json"
Parse the ToolResult envelope. On failure: continue the build but record cleanup_manifest_result as skipped (skipReason: "baseline_snapshot_failed", files: 0, ownedLineCount: 0). Do not dispatch comment-cleaner later unless change-manifest generate explicitly returns status: "created" with non-empty ownership.
§4.3 Builder-Reviewer Loop
Process TASK_PLAN.task_units with dependency-aware pipelining. Never derive task IDs from tasks.md; use TASK_UNIT_IDS from the current task_unit. Do not edit tasks.json from the orchestrator; the reviewer owns task-plan persistence.
Ready-Set Derivation
Initialize before the first dispatch:
COMPLETED_UNIT_TASK_IDS = {} (set of task IDs from units whose reviewer returned SUCCESS).UNIT_LOOKUP = map from each task ID to its parent task_unit.unit_id, built once from all task_units.
A unit is ready when every task ID in its depends_on is in COMPLETED_UNIT_TASK_IDS. Units with empty depends_on are ready immediately.
Before every dispatch cycle, recalculate READY_UNITS from unbuilt units, sorted by lowest unit_id. Do not require one builder to finish before checking for a ready wave. If no unit is ready and uncompleted units remain, emit implementation waiting with reason: "dependency_deadlock" and STOP.
If READY_UNITS has 2+ entries and all Parallel-Wave Mode preconditions pass, use Parallel-Wave Mode immediately. Otherwise pick the first ready unit by lowest unit_id and use the serial-pipelined dispatch below.
Pipelined Dispatch
For each selected ready unit k, dispatch the builder:
{% dispatch_agent "rp1-dev:task-builder" %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, TASK_IDS={TASK_UNIT_IDS}, GIT_COMMIT={GIT_COMMIT}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
After builder(k) completes, determine pipelining eligibility:
- Compute the next ready unit k+1 (lowest
unit_id among remaining ready units, excluding k).
- Unit k+1 is pipeline-eligible when it exists AND none of unit k's
task_ids appear in unit k+1's depends_on AND the two units' task file lists share no source file path (a failed reviewer(k) triggers a retry rebuild of k's files on the shared tree, so any overlap with k+1's in-flight edits is unsafe regardless of GIT_COMMIT).
When k+1 is pipeline-eligible: dispatch reviewer(k) AND builder(k+1) as parallel agents in a single message:
{% dispatch_agent "rp1-dev:task-reviewer", background %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, TASK_IDS={TASK_UNIT_IDS_K}, GIT_COMMIT={GIT_COMMIT}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
{% dispatch_agent "rp1-dev:task-builder", background %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, TASK_IDS={TASK_UNIT_IDS_K1}, GIT_COMMIT={GIT_COMMIT}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
Wait for both to complete before any new dispatch. Process reviewer(k) result first.
When k+1 depends on k or no k+1 is ready: dispatch reviewer(k) alone and wait:
{% dispatch_agent "rp1-dev:task-reviewer" %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, TASK_IDS={TASK_UNIT_IDS}, GIT_COMMIT={GIT_COMMIT}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
Reviewer Success
Reviewer contract: SUCCESS + task_plan_updated = true completes the unit. Add k's task_ids to COMPLETED_UNIT_TASK_IDS, recalculate the ready set, and continue to the next unprocessed unit.
Reviewer Failure -- Sequential Path
When reviewer(k) returns FAILURE or malformed and no builder(k+1) is in flight, enter the retry path immediately.
attempt = 1, max = 2. On FAILURE with attempt < max: build PREVIOUS_FEEDBACK JSON from issues/summary, re-spawn task-builder with feedback (if GIT_COMMIT=true, pass REWRITE_COMMITS=true for atomic commit rewrite):
{% dispatch_agent "rp1-dev:task-builder" %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, TASK_IDS={TASK_UNIT_IDS}, GIT_COMMIT={GIT_COMMIT}, REWRITE_COMMITS={GIT_COMMIT}, PREVIOUS_FEEDBACK={PREVIOUS_FEEDBACK}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
Re-run task-reviewer for the same unit. On second failure, escalate per §4.3.7.
Reviewer Failure -- Pipelined Path
When reviewer(k) returns FAILURE or malformed while builder(k+1) is in flight:
- Wait for builder(k+1) to complete. Do not dispatch anything new.
- Mark dependencies stale: any unit whose
depends_on includes task IDs from unit k is not-ready until k passes review. The already-built k+1's review is deferred until k is resolved.
- Run the retry path for unit k using the same attempt/max logic as the sequential path above.
- If k passes retry: add k's task IDs to
COMPLETED_UNIT_TASK_IDS, recalculate the ready set, then dispatch reviewer for the already-built k+1 unit.
- If k's retry is exhausted: escalate per §4.3.7. The already-built k+1 unit is abandoned along with k.
Exhausted-Retry Escalation {#s4-3-7}
Escalate without marking parent implementation failed while recovery remains.
- Interactive: emit
waiting_for_user on implementation with prompt "Task review failed after retry. Repair, Skip task, or Stop?" and context about the failing task unit. Then emit implementation waiting with task_unit: "{TASK_UNIT_IDS}" and reason: "review_retry_exhausted". STOP with /build {FEATURE_ID} resume instructions.
- AFK: if an explicit skip policy exists, record the skipped
TASK_UNIT_IDS as release follow-ups; otherwise emit parent implementation failed only because no skip/repair path remains.
Pipelining Rules
- Never two builders concurrently (serial mode). At most one task-builder may be in flight at any time in the default serial-pipelined path. The only concurrency is one reviewer alongside one builder on different units.
- Never reviewer and builder on the same unit. A unit's reviewer dispatches only after that unit's builder completes.
- Dependency gate. A unit whose
depends_on contains task IDs from unit k must not begin building until k's reviewer returns SUCCESS.
- Failure isolation. When a pipelined reviewer(k) fails, wait for the in-flight builder(k+1) to finish, then resolve k's retry before any new dispatch.
- Max two builders (parallel-wave mode only). Parallel-wave mode dispatches at most two builders concurrently -- one on the primary
codeRoot, one on a worktree codeRoot. Never three or more.
- Parallel mode requires GIT_COMMIT=true and clean tree. Without atomic commits the integration rebase has nothing to replay. A dirty working tree prevents worktree creation.
- Overlapping file lists never concurrent. If two units' task file lists share any source file path, they MUST NOT be in flight together — this applies to pipelined reviewer(k) ∥ builder(k+1) dispatch as well as parallel-wave builders. Fall back to strictly serial dispatch for those units.
- Serial fallback. When in doubt about preconditions, file overlap, or worktree health, fall back to the serial pipelined path. Parallel-wave is an optimization, not a requirement.
- Emit namespacing unchanged. Both primary and secondary builders emit with the same
task-builder: step prefix. The --unit parameter distinguishes their events.
- Default no-commit builds are serial. When
GIT_COMMIT=false, parallel groups in tasks.md describe dependency windows, but builder execution stays serial-pipelined because there is no atomic commit to replay from a worktree.
Parallel-Wave Mode
When ALL of the following preconditions are met, dispatch two builders concurrently instead of one:
GIT_COMMIT is true.- The working tree at
codeRoot is clean (no unstaged or uncommitted changes).
- The ready set contains 2+ units with no mutual dependency (neither unit's
depends_on includes the other's task IDs).
- The two candidate units' task file lists have no overlapping source file paths.
Trigger: At the start of each dispatch cycle, before selecting a serial unit, compute the current ready wave. If at least two ready units are mutually independent and file-disjoint, dispatch the first unit as primary builder(k) on codeRoot and the second unit as secondary builder(k+1) on a worktree codeRoot in a single message.
If fewer than two ready units qualify, use the standard serial-pipelined dispatch.
Dispatch:
{% dispatch_agent "rp1-dev:task-builder", background %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, TASK_IDS={TASK_UNIT_IDS_K}, GIT_COMMIT={GIT_COMMIT}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
{% dispatch_agent "rp1-dev:task-builder", background %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={worktreePath}, TASK_IDS={TASK_UNIT_IDS_K1}, GIT_COMMIT={GIT_COMMIT}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
Wait for both builders to complete. Dispatch reviewer(k) on the primary codeRoot before integrating the secondary worktree. If reviewer(k) fails or is malformed, abandon the secondary worktree, retry or escalate k per the Reviewer Failure sections, and rebuild k+1 later from the serial path after k passes. If reviewer(k) succeeds, add k's task IDs to COMPLETED_UNIT_TASK_IDS, then integrate the secondary worktree per references/parallel-builders.md and dispatch reviewer(k+1) on the primary codeRoot.
Integration order: After both builders finish and reviewer(k) succeeds, integrate the worktree branch (k+1) onto the primary branch per the Integration section in references/parallel-builders.md. If integration fails, apply the Conflict Fallback procedure (discard worktree, rebuild k+1 serially). Then dispatch reviewer(k+1) using the standard serial or pipelined path.
See references/parallel-builders.md for the full worktree lifecycle protocol: creation, CODE_ROOT routing, integration, conflict fallback, cleanup, and failure handling.
§4.4 Post-Build
Documentation tasks from TASK_PLAN.documentation_tasks:
- Complete only through a declared supported workflow.
- Current build has no declared docs implementation workflow; create
documentation_followups from each docs task with id, title, target, acceptance_refs, dependencies, and blocks_release = false.
- Carry
documentation_followups into readiness/release manual_items.
- Do not spawn undeclared documentation agents.
§4.5 Cleanup Manifest Generation
After builders, reviewers, and documentation follow-up collection finish, generate the durable cleanup handoff before verification:
rp1 agent-tools change-manifest generate \
--code-root "{codeRoot}" \
--out "{workRoot}/features/{FEATURE_ID}/change-manifest-001.json" \
--status-out "{workRoot}/features/{FEATURE_ID}/change-manifest-status.json" \
--source build \
--baseline "{workRoot}/features/{FEATURE_ID}/change-manifest-baseline.json"
Parse the ToolResult envelope into cleanup_manifest_result. If data.status == "created" with files > 0 and ownedLineCount > 0, comment-cleaner may participate using data.manifestPath. If data.status == "skipped", preserve statusPath/skipReason for the aggregator. On failure: set skipped with skipReason: "change_manifest_generate_failed", files: 0, ownedLineCount: 0.
§4.6 Verification And Readiness
Step 1 — Resolve comment-cleaner participation
Evaluate the cleanup manifest result BEFORE dispatching any verification agent. Comment-cleaner participates only when ALL of: cleanup_manifest_result.data.status == "created", cleanup_manifest_result.data.files > 0, cleanup_manifest_result.data.ownedLineCount > 0, and cleanup_manifest_result.data.manifestPath is present. Do not dispatch comment-cleaner with branch, unstaged, commit-range, base-branch, mode, or commit parameters; the generated manifest is the only safe cleanup boundary.
If comment-cleaner will NOT participate, set the comment_cleaner phase result now (before dispatch): status: "WARN", empty blocking_issues, one warning noting "Automatic comment cleanup skipped because no non-empty generated manifest was available" with evidence from cleanup_manifest_result.data.statusPath, one artifact entry for the status path with storageRoot: "absolute", one evidence entry with status: "not_applicable" and summary from skipReason, files_checked: 0, manifest_path: null, and manifest_status_path/skip_reason from the cleanup manifest result.
Step 2 — Parallel dispatch
CRITICAL: Emit ALL dispatch blocks below back-to-back in ONE message with NO text between them. Always include code-checker and feature-verifier. Include comment-cleaner only when Step 1 resolved it as participating; when not, omit its block entirely. No prose, conditionals, or explanatory text may appear between dispatch blocks.
{% dispatch_agent "rp1-dev:code-checker", background %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}
{% enddispatch_agent %}
{% dispatch_agent "rp1-dev:feature-verifier", background %}
FEATURE_ID={FEATURE_ID}, KB_ROOT={kbRoot}, WORK_ROOT={workRoot}, CODE_ROOT={codeRoot}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
{% dispatch_agent "rp1-dev:comment-cleaner", background %}
CHANGE_MANIFEST={cleanup_manifest_result.data.manifestPath}, CODE_ROOT={codeRoot}
{% enddispatch_agent %}
Step 3 — Wait for all before aggregation
Wait for ALL dispatched verification agents to complete before proceeding. Do not begin aggregation until every agent result is available. Collect each agent's response into its corresponding slot below.
Build PHASE_RESULTS_JSON with keys: code_checker (validation envelope or legacy), feature_verifier (validation envelope or legacy), comment_cleaner (validation envelope or synthetic warning), and implementation_context containing task_plan_warnings and documentation_followups arrays. These MUST be the preserved arrays from §4.1 and §4.4 -- never hardcode to [] unless actually empty.
{% dispatch_agent "rp1-dev:build-verify-aggregator" %}
PHASE_RESULTS={PHASE_RESULTS_JSON}, FEATURE_ID={FEATURE_ID}, WORK_ROOT={workRoot}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
Extract readiness_status, release_behavior, ready_for_release, blocking_issues, warnings, and manual_items. Preserve compatibility fields overall_status and ready_for_merge when present.
Readiness release behavior:
- PASS/proceed: release may start.
- WARN/proceed_with_notes: release may start with warnings/manual notes visible.
- FAIL/return_to_implementation: keep parent
implementation running for planned repair or waiting for a user decision.
- WAITING/wait_for_human: emit parent
implementation waiting for required manual evidence.
If readiness has blocking failures or missing required components, keep parent implementation running for planned repair or waiting for a user decision. Emit parent implementation failed only when no repair/decision path remains.
If readiness is FAIL or WAITING in interactive mode, present the readiness evidence before stopping: emit waiting_for_user on implementation with prompt "Readiness needs work. Repair, Add Task, Review feedback from Arcade, or Stop?" and readiness context (status, blocker/warning/manual-item counts, readiness artifact path). Then emit implementation waiting with readiness_status. STOP with /build {FEATURE_ID} resume instructions.
If AFK and readiness is FAIL or WAITING, emit implementation failed unless an explicit repair/skip policy is already available.
When readiness is PASS or WARN and can proceed to release, present the human gate before release.
Implementation checkpoint (after readiness; skip if AFK):
Emit waiting_for_user on implementation with prompt "Release, Add Task, Review feedback from Arcade, or Stop?" and readiness context.
{% ask_user "Release, Add Task, Review feedback from Arcade, or Stop?", options: "Release", "Add Task", "Review feedback from Arcade", "Stop" %}
On Release: continue.
On Add Task: collect ADDED_TASK_REQUEST, dispatch feature-tasker with UPDATE_MODE=true and UPDATE_CONTEXT={"source":"implementation_checkpoint","request":"{ADDED_TASK_REQUEST}"}, validate the same success contract as §2.3, then emit implementation waiting with reason: "readiness_add_task" and added_task_request. STOP with /build {FEATURE_ID} resume instructions. On resume, build-task-plan must consume the updated tasks.json.
{% dispatch_agent "rp1-dev:feature-tasker" %}
FEATURE_ID={FEATURE_ID}, WORK_ROOT={workRoot}, UPDATE_MODE=true, UPDATE_CONTEXT={"source":"implementation_checkpoint","request":"{ADDED_TASK_REQUEST}"}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
On Review feedback from Arcade: load arcade-collab skill, process all feedback for RUN_ID, then return to this checkpoint with original options.
On Stop: emit implementation waiting with reason: "stopped_at_implementation_checkpoint" and STOP with /build {FEATURE_ID} resume instructions.
After the user chooses Release, or AFK skips this checkpoint, emit implementation completed per §PARENT-EMIT-DISCIPLINE table.
Git Operations (conditional)
If GIT_COMMIT: stage+commit. If GIT_PUSH: push. If GIT_PR: create PR.
§PHASE-4: Release
Skip if: START_PHASE is after release.
Resume checkpoint: If WAITING_PHASE.phase == "release", inspect the latest parent waiting/status event from WORKFLOW_STATE.recent_events:
reason = "add_task_requested" -> jump to §PHASE-3 Implementation and consume the updated tasks.json.reason = "archive_incomplete" -> set ARCHIVE_RETRY_PATH from the prior archiver result's exact archive_path, then jump directly to Archive retry.- otherwise jump directly to the Release gate.
Do not emit release completed on a waiting resume until the resumed release decision succeeds.
Release MUST start only after readiness aggregation has completed.
Before emitting release running:
- Set
READINESS_CONTRACT from the build-verify-aggregator JSON returned in §4.6 when this invocation ran implementation.
- If resuming directly at
release, set READINESS_CONTRACT from WORKFLOW_STATE.artifacts only when a registered artifact exists with path = "features/{FEATURE_ID}/build-readiness.md" and storageRoot = "work_dir".
- If no readiness contract or registered readiness artifact exists, emit
implementation waiting with reason = "missing_readiness_contract" and STOP. Do not emit release running.
- If
READINESS_CONTRACT.readiness_status is FAIL or WAITING, or ready_for_release is false, return to §4.6 readiness handling. Do not start release.
Missing readiness: emit implementation waiting with reason: "missing_readiness_contract".
Emit release running per §PARENT-EMIT-DISCIPLINE table before presenting release options.
Output: Feature ID, phase status table, registered artifacts, readiness artifact, readiness status, blockers, warnings, and manual items. Show manual checklist status before archive options: satisfied, not applicable, or still visible as release notes. Do not claim manual items are complete unless the readiness contract says so.
Release gate (skip if AFK; AFK defaults to archive):
Emit waiting_for_user on release with prompt "Release, Add task, Review feedback from Arcade, or Stop?" and readiness context.
The canonical release options are five (Archive, Complete without archive, Add task, Review feedback from Arcade, Stop) — over the 4-option cap — so present them as two steps per §CHECKPOINT-OPTIONS.
{% ask_user "Release, Add task, Review feedback from Arcade, or Stop?", options: "Release", "Add task", "Review feedback from Arcade", "Stop" %}
On Release: ask the archive sub-decision as a separate question:
{% ask_user "Archive the feature now, or complete without archiving?", options: "Archive", "Complete without archive" %}
On Archive: proceed to the Archive step below.
On Complete without archive: emit release completed per §PARENT-EMIT-DISCIPLINE table with archive_status: "declined" and STOP. Do not run feature-archiver.
On Add task: collect ADDED_TASK_REQUEST, dispatch feature-tasker with UPDATE_MODE=true and UPDATE_CONTEXT={"source":"release_gate","request":"{ADDED_TASK_REQUEST}"}, validate the same success contract as §2.3. Emit release waiting with archive_status: "deferred", reason: "add_task_requested", and added_task_request. Emit implementation waiting with reason: "release_add_task" and added_task_request. STOP. Parent release MUST NOT complete until release is re-entered after implementation and readiness re-aggregation.
{% dispatch_agent "rp1-dev:feature-tasker" %}
FEATURE_ID={FEATURE_ID}, WORK_ROOT={workRoot}, UPDATE_MODE=true, UPDATE_CONTEXT={"source":"release_gate","request":"{ADDED_TASK_REQUEST}"}, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
On Review feedback from Arcade: load arcade-collab skill, process all feedback for RUN_ID, then return to this checkpoint with original options.
On Stop: emit release waiting with archive_status: "deferred" and STOP with /build {FEATURE_ID} resume instructions.
Archive
{% dispatch_agent "rp1-dev:feature-archiver" %}
MODE=archive, FEATURE_ID={FEATURE_ID}, ARCHIVE_PATH={ARCHIVE_RETRY_PATH}, WORK_ROOT={workRoot}, SKIP_DOC_CHECK=false, WORKFLOW=build, RUN_ID={RUN_ID}
{% enddispatch_agent %}
Parse the feature-archiver response. Accept success from JSON with status: "success", mode: "archive", archive_status: "completed", archive_path, and an artifacts[] entry beginning with archives/features/ using storageRoot: "work_dir". Require registration_status: "registered" since WORKFLOW/RUN_ID were passed.
On failure (needs_confirmation, malformed, missing archive result/registration): do NOT emit release completed. If registration_retry_required: true, set ARCHIVE_RETRY_PATH = response.archive_path. Interactive: emit release waiting with reason: "archive_incomplete" and archive_path, STOP. AFK: emit release failed only when no recovery remains.
Archive-incomplete: emit release waiting with archive_status: "incomplete", reason: "archive_incomplete", and archive_path: "{ARCHIVE_RETRY_PATH}".
After feature-archiver succeeds and registers the actual archived output, emit release completed per §PARENT-EMIT-DISCIPLINE table with archive_status: "completed" and archive_path.
§TERMINAL-STATES
Every exit path MUST emit a terminal status. No run may remain in "running" after the skill finishes.
| Exit Path | Status | Step |
|---|
| Archive completes successfully | completed | release |
| User selects "Stop" at any checkpoint | waiting | current step |
| Oversized scope redirected to /phase-plan | cancelled (end-run) | planning |
| User selects "Complete without archive" at release gate | completed | release |
| Unrecoverable agent failure | failed | failing parent phase |
| AFK mode abort | failed | failing parent phase |
On any unrecoverable failure, emit per §PARENT-EMIT-DISCIPLINE table with status: "failed" on the failing parent phase.
§ANTI-LOOP
Single-pass execution. Parse → detect → run steps → STOP.
