Zhachory1 Plan to Launch
Ensure a thought-out plan exists before implementation. Move from product intent to technical design, council stress-test, executable plan, ship, code review, final council, and human launch approval.
Core Principles
Intent before code. PRD and DD must be good enough before implementation starts.
Cheap failures first. Catch wrong assumptions in docs and council, not after code exists.
Ship follows accepted plan. Implementation should satisfy planned tasks and flag plan/design mismatches instead of silently inventing fixes.
Material expansion renews approval. Discovery does not amend accepted scope. Pause implementation and obtain human approval for recorded material scope delta before continuing.
Review checks fidelity. Code-reviewer judges correctness and whether code still matches PRD/DD intent.
Loops are bounded and delta-only. Re-review only raised issues and changed material.
Inputs
- product or user goal.
- target repo and branch.
- existing docs, tickets, issues, prior decisions, code context.
- success metrics or constraints.
- run id and budget caps.
Tool Awareness
Use grounding-brief with long-term memory MCP, code RAG, git, tickets, PRs, and agent-fleet journals. Use agent-fleet /council for required minimal/full councils after AGENT_FLEET_HOME preflight. Use agent-fleet /ship or local ship skill for implementation when plan is accepted. Use run-telemetry at each gate.
Council Policy
Use agent-fleet council levels: none, single-lens, minimal-council, or full-council. Low-risk, reversible work may record council_level: none or use single-lens; docs and final code councils are required only when risk, ambiguity, conflicting review, unresolved blocker, design mismatch, security/privacy/reliability concern, or leadership-visible tradeoff appears. Required minimal-council/full-council must block if agent-fleet preflight fails.
Workflow Modes
lite: accepted small change; produce scope lock, focused plan, patch/review handoff. Budget: focused refs, one artifact,council_levelmaxsingle-lens, loop cap 1, focused validation.default: normal feature/change; PRD/DD may be lightweight, task plan required, council only if risk triggers. Budget: selected docs, up to two artifacts,council_levelmaxminimal-council, loop cap 2.deep: high-risk architecture, launch, security/reliability, cross-team, or leadership-visible work. Budget: full docs, required gates, council allowed with reason.- Escalate: ambiguity, design mismatch, unresolved blocker, risky rollout, or conflicting review.
- Load: start with this skill and user artifact; load doc templates, agent-fleet council protocol, handoff, or telemetry references only when that gate will run.
Execution
Run only the steps required by selected mode. lite may use scope lock instead of formal PRD/DD; default may use lightweight PRD/DD; deep uses full docs and gates.
Ground context. Invoke grounding-brief over user goal, memory, code RAG, existing docs, related PRs, issues, and prior runs. Do not propose infrastructure until its capability reuse inventory is complete.
Lock scope. Before implementation, record a scope fingerprint: intended files, effort, systems, trust/deployment/rollback boundaries, and issue-wide implementation PR count.
Author PRD. Use structured-doc-authoring to define what, why, users, scope, non-goals, success criteria, and launch readiness. Use success-criteria-metrics to lock the primary outcome and guardrails.
Author DD. Use structured-doc-authoring to translate PRD into architecture, boundaries, data/control flow, tradeoffs, tests, observability, rollout, risks, and alternatives.
Council on docs. Select council level from risk: none for low-risk clear docs, single-lens for one domain concern, and agent-fleet minimal-council/full-council only for multi-domain or high-risk design decisions. Required council verdict must pass or pass-with-nits before planning only when council level is minimal/full.
Scope expansion gate. Pause implementation and invoke human-approval-gate when review or discovery adds infrastructure or a trust boundary; doubles expected files, effort, or systems touched; raises issue-wide implementation PR count above two; turns configuration/integration into platform capability; or changes deployment/rollback semantics. Package original scope fingerprint, discovery, smallest options, added cost/risk, and recommendation. Emit gate/escalation telemetry with reuse candidate outcomes/count, expansion reason, and renewed approval state. approve resumes only when tied to fingerprint; reject ends paused run; revise returns to design and remains paused; extend budget approves recorded cap only. No response remains pending, and later delta repeats gate.
Write plan. Use task-decomposition-planning to create ordered implementation tasks with dependencies, acceptance criteria, validation, rollout, and stop conditions.
Ship tasks. Use agent-fleet ship for each accepted task or slice, with at most two issue-wide implementation PRs unless human approval extends cap. If implementation exposes design error, stop and return to DD instead of patching around it.
Run code-reviewer. Review for correctness, safety, maintainability, security, performance, tests, observability, and fidelity to PRD/DD. Re-review only deltas after fixes.
Council on review and code. Use agent-fleet /council only when review findings conflict, blockers remain unresolved, implementation reveals design mismatch, or risk level justifies a final merge/no-merge council against original intent.
Launch PR. Use handoff-packaging to create PR package with linked docs, rationale, tests, council verdicts, and known limitations. Use human-approval-gate for final signoff.
Loop Controls
Doc loop. PRD/DD and agent-fleet council gate iterate only when council level is minimal/full; otherwise record skipped or single-lens reason and proceed when doc acceptance criteria are met.
Implementation loop. ship and code-reviewer iterate until review has no blockers, cap hit, or plan mismatch discovered.
Outer design kickback. If final council or review evidence finds design intent wrong, return to DD and preserve implementation learnings as evidence. A material delta pauses the run until human approval for its scope fingerprint; telemetry is evidence, not gate authority.
Delta-only rule. Re-council and re-review focus on required changes, not settled material.
State carry-forward. Maintain decisions, rejected alternatives, review issues, and plan mismatches in the run directory.
Output Contract
# Dev Workflow Run
## Artifacts
- PRD:
- DD:
- council docs level/verdict:
- implementation plan:
- shipped tasks:
- code review:
- final council level/verdict:
- PR package:
## Gate State
- docs: pass | pass-with-nits | block
- implementation: clean | needs-fix | design-kickback | paused
- scope fingerprint:
- scope renewal: not-required | pending | approved | rejected | revision-requested
- paused action / resume condition:
- delivery PR budget: count, approved cap, deferred work
- final: launch | revise | escalate-human
## Evidence
- tests:
- validation:
- linked PR/docs:
- telemetry path:
Success Criteria
- PRD/DD or scope lock exists before implementation, according to selected mode.
- council docs gate either records skipped/one-lens reason or passes when minimal/full council is triggered.
- task plan or lite task list has acceptance criteria and dependencies.
- material scope expansion emits a human approval gate and pauses implementation.
- delivery uses no more than two issue-wide implementation PRs without explicit approval.
- telemetry records reuse candidates, scope expansion, and renewed approval state.
- shipped code maps to planned tasks.
- code-reviewer blockers resolved or escalated.
- final council confirms merge/no-merge call only when escalation matrix triggers it; otherwise skipped reason is recorded.
- launch package and human approval ask exist.
run-telemetryevents emitted for all major gates.
Common Failure Modes
Skipping docs for speed. Fix by writing lightweight PRD/DD, not by jumping to code.
Council churn. Fix by agent-fleet caps, delta-only review, and human escalation.
Implementation invents new design. Fix by DD kickback.
Review drifts from intent. Fix by linking every blocker to PRD/DD, code evidence, or safety risk.
Self-Improvement
Record which gates found real defects, which council blockers were false positives, where plan missed implementation reality, and where human reviewers edited PR package. Feed these into backprop for workflow optimization.