Source file: docs/LIFECYCLE_SKILL_ROUTING.md (maintainer entry). Regenerate this HTML page after changing the source file.
Lifecycle And Skill Routing
DHF_PUBLIC_STATUS_V1 · 2026-08-03: This router documents the current stage-based source architecture. Source availability does not prove runtime activation or publication. See the canonical status in English or 中文. The proposed light/standard/governed simplification profiles are not active.
Purpose
This document explains which workflows MyCodexEnv covers, how the Harness Runtime maps work to lifecycle stages, and which Codex/gstack skills should be used at each stage.
MyCodexEnv has three layers:
- Runtime layer:
AGENTS.md, hooks, tool policy, evidence schema, verification scripts, and checkpoint helpers. - Router layer:
delivery-harness-framework, which reads durable state, classifies the lifecycle stage, and selects the next workflow. - Specialist layer: repo-specific lifecycle adapters, gstack skills, local planning/testing/review skills, and deterministic helper scripts.
The generic lifecycle skill should stay generic. Project-only paths, fixtures, checks, smoke matrices, and business safety boundaries belong in repo-specific adapter skills.
DHF Source-Stage Contract Mirror
Canonical profile and Result Invariant semantics live in
codex/skills/delivery-harness-framework/SKILL.md; routing and rollout-switch
behavior live in codex/hooks/dhf_preprompt.py.
- Governance profiles are
light,standard, andgoverned. - Generic injection requires explicit generic activation; an ordinary non-project prompt stays continue-only. Opt-out is evaluated before every route. The dispatcher recognizes the ShipQ project boundary and directly lazy-delegates; generic DHF must not recover, read shared state, or pre-classify before the ShipQ adapter owns the task.
- Every completion preserves exactly
result,scope_and_constraints,verification_receipt, andremaining_risk_or_next_action; profile-specific ceremony is conditional rather than repeated by default. - The source simplified route uses
DHF_PREPROMPT_SIMPLIFIED_PROFILES=1; values0,false,off, andlegacyroll back. Runtime evidence must report eithersource_stage_unsyncedorruntime_promotedand reject drift.
Visual Guides
docs/dhf-architecture-status-en.html: canonical English source/runtime/publication status.docs/dhf-architecture-status-cn.html: canonical Chinese source/runtime/publication status.docs/index.html: Chinese Delivery Harness Framework public landing page for GitHub Pages.docs/index-en.html: English Delivery Harness Framework public landing page for English readers.docs/dhf-workflow-skills-cn.html: 中文 DHF workflow skills matrix,逐步列出每个阶段使用的 skill、职责说明和 GitHub 源码链接。docs/dhf-workflow-skills-en.html: English DHF workflow skills matrix with stage-by-stage skill explanations and GitHub source links.docs/delivery-harness-beginner-guide-cn.html: 中文 beginner guide,用一句话定义、三层架构、五步流程、例子和术语表解释 Delivery Harness Framework。docs/project-lifecycle-harness-flow-cn.html: 中文纵向主流程图,帮助用户从入口、状态恢复、阶段分类一路看到验证、发布和交接。docs/project-lifecycle-harness-flow-skills.html: 中文 skill/helper 路由速查页,帮助用户理解每个生命周期阶段应使用哪个 skill,以及 helper 的执行职责。docs/project-lifecycle-harness-flow-skills-zh-status-style.html: 新版中文 Delivery Harness Framework skill/helper 路由图,使用和 LinkedIn 素材一致的状态风格。docs/project-lifecycle-harness-flow-skills-en-status-style.html: English status-style Delivery Harness Framework flow map for public sharing.
Related Documentation
README.md: top-level quick start and Harness Runtime overview.docs/repo-index.md: low-token repo navigation and runtime surface index.docs/index.html: Chinese public Delivery Harness Framework docs entry.docs/index-en.html: English public Delivery Harness Framework docs entry.docs/dhf-workflow-skills-cn.html: Chinese workflow skill inventory with GitHub links.docs/dhf-workflow-skills-en.html: English workflow skill inventory with GitHub links.docs/delivery-harness-beginner-guide-cn.html: beginner-oriented Chinese explanation.docs/HARNESS_RUNTIME.md: lifecycle, evidence, checkpoint, permission, and subagent contracts.docs/AGENT_HARNESS_STATUS.md: Agent Harness workflow/infra status map.docs/CODEX_ENV_REPRODUCTION.md: Codex + Claude environment reproduction guide.docs/project-lifecycle-harness-flow-cn.html: Chinese vertical lifecycle flow.docs/project-lifecycle-harness-flow-skills.html: Chinese lifecycle skill/helper routing visual guide.docs/project-lifecycle-harness-flow-skills-zh-status-style.html: current styled Chinese lifecycle skill/helper routing visual guide.
Covered Workflows
| Workflow | What It Covers | Primary Entry |
|---|---|---|
| Environment reproduction | Clone, bootstrap, sync Codex/Claude homes, install pinned prerequisites, verify local setup. | README.md, bootstrap.sh, scripts/verify_codex_env.sh |
| Global rules and repo routing | Keep cross-repo Codex rules in one source and route repo-local constraints through AGENTS.md. |
codex/AGENTS.md, root AGENTS.md, scripts/manage_agents.py |
| Skill synchronization | Keep codex/skills/* as source of truth and copy managed skills into ~/.codex/skills/*. |
scripts/sync_codex_home.sh |
| Harness lifecycle routing | Recover state, classify phase, choose generic/repo-specific/gstack workflow, define gates. | delivery-harness-framework |
| Requirements and planning | Capture success criteria, scope, constraints, domain docs, ADR conflicts, vertical slices, risks, and validation gates before implementation. | planner, req-to-dev, task-flow-orchestrator, scripts/harness_requirements.py |
| Backlog-ready issue/spec authoring | Turn vague implementation intent into a bounded issue or executable spec before coding starts. | vendored gstack spec, delivery-harness-framework |
| Development execution | Make scoped repo changes, use tests first when behavior changes, preserve unrelated user work. | tdd-guide, atdd-guide, repo-local workflow |
| Validation and evidence | Run fresh tests/checks and record check, how it exited, what it showed, when it ran. |
verification-loop, scripts/harness_evidence.py, scripts/harness_report.py |
| Review and QA | Inspect diffs, browser behavior, security/privacy boundaries, and user-visible regressions. | gstack-review, code-reviewer, gstack-qa, security-reviewer |
| Committee review loop | Run an explicit expert-committee scoring loop with a separate revision worker until a target rating is met. | committee-review-loop |
| Ship and deployment | Commit/PR/release/deploy/canary only when explicitly requested and verified. | gstack-ship, gstack-land-and-deploy, gstack-canary |
| Handoff and recovery | Append state, preserve next safe task, recover without chat history, and optionally save/restore user-facing gstack context. | scripts/harness_checkpoint.py, scripts/harness_recover.py, vendored gstack context-save / context-restore |
| Documentation release | Keep README/docs aligned with shipped behavior and skill/runtime changes; generate missing docs, polished artifacts, or source-backed diagrams when requested. | gstack-document-release, vendored gstack document-generate, vendored gstack make-pdf, vendored gstack diagram, doc-updater |
Lifecycle Stage Map
| Stage | Signals | Default Permission | Skill Or Helper | Purpose |
|---|---|---|---|---|
research |
Unknown repo, stale handoff, unclear source ownership, missing context. | Read-only. | delivery-harness-framework, scripts/harness_recover.py, scripts/harness_env_probe.py |
Read durable sources and recover the next safe task before acting. |
requirements |
Goal, audience, acceptance criteria, scope, constraints, or domain language are unclear. | Read-only. | planner, req-to-dev, scripts/harness_requirements.py |
Turn ambiguity into success criteria, read CONTEXT.md / CONTEXT-MAP.md and relevant ADRs when present, then validate requirements artifacts before treating them as source of truth. |
planning |
The user wants a backlog-ready GitHub issue, ticket, or executable spec before implementation begins. | Read-only by default. | vendored gstack spec, delivery-harness-framework |
Preserve runtime state, lane boundaries, and validation gates while turning vague intent into a bounded spec or issue. |
planning |
Architecture, API shape, migration, runtime, data flow, cross-module decisions, or multi-step work breakdown. | Read-only by default. | gstack-plan-eng-review, planner, task-flow-orchestrator |
Produce an implementation plan, vertical slice breakdown, risk list, and validation gate before edits. |
development |
Acceptance criteria are clear and touched files/modules are bounded. | Scoped writes. | tdd-guide, atdd-guide, repo workflow, scoped worker agents |
Implement the change with tests appropriate to the risk. |
validation |
Work is implemented or docs/config changed and needs fresh evidence. | Tests/checks by default. | verification-loop, scripts/harness_report.py, scripts/verify_codex_env.sh |
Prove the result with fresh check output and required evidence fields. |
review |
Diff exists, work is near handoff/PR/landing, or user asks for review. | Read-only by default. | gstack-review, code-reviewer, review-swarm, security-reviewer |
Find bugs, regressions, security risks, and missing tests before shipping. |
review |
User explicitly asks for a committee, expert panel, subagent reviewer/worker split, rating loop, or improvement until a score such as 9.5/10. |
Committee read-only; revision worker scoped writes. | committee-review-loop, scripts/harness_agent_team.py |
Keep ordinary review routes separate while iterating committee findings into bounded, verified revisions. |
ship |
User requests commit, push, PR, merge, release, deploy, production verification, or version-slot queue visibility. | Requested release actions only. | gstack-ship, gstack-land-and-deploy, gstack-canary, vendored gstack landing-report |
Complete release steps, verify deployment, and preserve rollback or version-queue context. |
handoff |
Session is ending, task crosses phases, work spans sessions, next safe task must survive, or the user asks to save/restore working context. | Docs/state only unless a gstack context skill is explicitly invoked. | scripts/harness_checkpoint.py, vendored gstack context-save / context-restore, gstack-retro, gstack-learn |
Append harness checkpoint state, capture blockers, preserve resumable user context, and leave a recoverable next step. |
Specialist Skill Map
| Skill | Use For | Notes |
|---|---|---|
delivery-harness-framework |
Generic startup/resume routing, durable state reads, phase classification, helper selection, evidence expectations. | Use first for complex or resumed work; delegate after phase selection. |
| repo-specific lifecycle harnesses | Project paths, local checks, business fixtures, deployment topology, smoke matrices. | These adapters take over after the generic router identifies a repo-specific boundary. |
gstack-plan-ceo-review |
Product framing, user value, scope, demo boundaries, strategic tradeoffs. | Use when product judgment is the work. |
vendored gstack spec |
Backlog-ready GitHub issues, tickets, and executable specs with explicit acceptance criteria before implementation begins. | Use when the work needs a durable spec artifact rather than immediate coding. |
gstack-office-hours |
Founder-style pressure testing, market/user/business clarity. | Useful before committing to a product direction. In gstack 1.52.x, it can use configured gbrain context during planning preflight. |
gstack-plan-eng-review |
Architecture, data model, API contract, migration, performance, test strategy, and distribution/publish readiness for new artifacts. | Use before implementation on complex engineering plans; the current upstream workflow is explicitly search-aware, completeness-biased, and brain-aware when gbrain is configured. |
gstack-plan-design-review |
UX structure, visual direction, responsive behavior, design acceptance, and mockup-first plan refinement. | Use before UI implementation or major design changes. When a design binary/mockup path exists, prefer visual review over prose-only critique; when gbrain is configured, let the planning skill consume it as specialist context. |
gstack-qa |
Browser QA, screenshots, console/network checks, responsiveness, accessibility smoke, then iterative bug fixing with regression evidence. | Use for user-facing web behavior when fixes are allowed. |
gstack-qa-only |
Report-only QA when fixes should not be applied automatically. | Keeps QA read-only. |
gstack-ios-qa |
Real-device iOS QA on a live app via the embedded StateServer and agent loop. | Use for "test my iPhone app", device-only bug hunts, or live iOS smoke runs. |
gstack-ios-design-review |
Real-device iOS visual audit against Apple HIG and app design intent. | Use for shipped iOS UI review, not pre-implementation design planning. |
gstack-ios-fix |
Closed-loop iOS bug fixing after a real-device repro exists. | Use when an iOS defect should be fixed, redeployed, and reverified autonomously. |
vendored gstack ios-sync / ios-clean |
Maintain or remove the gstack iOS debug bridge and debug-only instrumentation. | Use only for explicit iOS bridge regeneration or release-cleanup requests. |
gstack-cso |
Infrastructure-first security and privacy review. | Use for auth, tokens, secrets, PII, or public/private boundaries. |
security-reviewer |
Code-level security review for validation, auth, injection, secret leakage, error disclosure. | Good for narrower security audits. |
gstack-review |
Pre-landing diff review, specialist checks, and fix-first handling of mechanical findings. | Use near PR, handoff, or ship; current upstream review is explicitly geared toward auto-fixing low-risk findings before asking. |
code-reviewer |
Local code review of staged or unstaged changes. | Use for focused findings with file/line references. |
review-swarm |
Parallel read-only review across behavioral/security/performance/test risks. | Use when the user asks for multi-agent review. |
committee-review-loop |
Explicit expert-committee/subagent loops with scoring, revision briefs, and a target rating such as 9.5/10. |
Do not use for one-off design/code/QA review; preserve read-only committee roles and bounded revision-worker write sets. |
gstack-ship |
Opinionated release workflow before PR/landing, including coverage audit, plan-completion audit, pre-landing review, PR title/version hygiene, and documentation sync. | Only use when user asks to ship. |
gstack-land-and-deploy |
Merge and deploy workflow with post-deploy checks. | Requires explicit release/deploy intent. |
gstack-canary |
Post-deploy canary monitoring. | Use after deploy when live behavior must be watched. |
gstack-document-release |
Post-ship documentation update across README/docs/CHANGELOG/TODOS/VERSION, with coverage-map and documentation-debt checks. | Use when docs must reflect shipped behavior. |
vendored gstack document-generate |
Missing tutorials, how-to docs, reference pages, or explanatory docs for a feature/module/project. | Use when documentation must be created from scratch rather than just updated after release. |
vendored gstack make-pdf |
Convert Markdown into a polished PDF artifact. | Use for publication-quality exports, not routine repo docs. |
vendored gstack diagram |
Turn an English description or mermaid source into a diagram triplet (.mmd, editable .excalidraw, rendered SVG/PNG). |
Use for offline source-backed diagrams that may later be embedded into docs or exported through make-pdf. |
vendored gstack landing-report |
Read-only version queue and PR landing visibility before ship. | Use when asking what version slot or landing order comes next. |
vendored gstack context-save / context-restore |
User-facing save/resume of working context across sessions or Conductor workspaces. | Use alongside, not instead of, scripts/harness_checkpoint.py; the harness checkpoint remains the repo-local append-only state source. |
vendored gstack scrape / skillify |
Read-only browser data extraction and codifying repeated scrape flows into browser-skills. | Use for extraction workflows; use gstack-qa for behavior testing and bug fixing. |
vendored gstack setup-gbrain / sync-gbrain |
Configure or refresh optional gbrain-backed repo memory/search. | Use only when the user explicitly wants gbrain setup, refresh, or repo re-indexing. |
gstack-plan-tune |
Tune repeated planning questions, developer profile preferences, and question-memory behavior. | Use only when the user asks to tune how gstack asks questions or to inspect/update developer profile signals; this is operator preference state, not repo lifecycle state. |
doc-updater |
Targeted docs/codemap/README updates for feature, API, or structure changes. | Use for repo-local documentation maintenance. |
verification-loop |
Build/static/test/coverage/security verification before completion. | Use when a full validation loop is required. |
tdd-guide |
Test-first implementation for behavior changes and bug fixes. | Use before writing production code. |
atdd-guide |
Acceptance-test-first workflow for business-facing criteria. | Use when external behavior and acceptance scenarios matter. |
planner |
Requirement analysis, architecture sorting, task decomposition, risk detection. | Use for complex features and refactors. |
task-flow-orchestrator |
Ordered execution across Karpathy, Planner, TDD, and verification workflows. | Use when the task needs workflow sequencing. |
skill-evaluator |
Skill existence, routing, eval matrix, baseline comparison, off-target load diagnosis. | Use before expanding or revising critical skills, especially when upstream workflow text changed and routing precision may drift. |
visual-explainer |
Self-contained HTML diagrams, flowcharts, architecture views, tables, and visual explanations. | Use for visual documentation and system explanations. |
Skillset Gap Routes
| Pattern | Lifecycle Slot | Harness Rule |
|---|---|---|
| Domain docs and ADRs | requirements, planning |
Read CONTEXT.md, CONTEXT-MAP.md, and relevant ADR files when present; use their domain vocabulary and flag conflicts instead of inventing terms. |
| Vertical slices | planning, development |
Break multi-step work into independently verifiable vertical slice units; mark each unit AFK when a worker can proceed from durable context or HITL when human judgment is required. |
| Feedback loop first | debug/investigation |
Establish a runnable feedback loop before hypotheses or fixes: failing test, CLI fixture, curl script, browser check, trace replay, throwaway harness, fuzz loop, or differential run. |
| Prototype learning | planning |
Use a throwaway prototype only to answer one named logic, state, interface, or UI question; delete it or capture the durable decision before handoff. |
| Durable agent brief | planning, development, handoff |
Use docs/templates/harness-agent-brief.md for a durable agent brief when worker contracts need current behavior, desired behavior, key interfaces, acceptance criteria, and out of scope. |
| Expert committee rating loop | review, development |
Route only explicit committee/subagent rating loops to committee-review-loop; capture expert domains, target rating, revision worker scope, verification gate, and stopping condition. |
| Deep module review | planning, review |
Prefer a deep module shape with small public interfaces and meaningful behavior behind them; use the interface as the test surface and preserve locality/leverage. |
Runtime Helper Map
| Helper | Lifecycle Slot | Purpose |
|---|---|---|
scripts/harness_recover.py |
research, handoff |
Read repo index, harness state, git status/log, and local evidence summary to recover current phase and next safe task. |
scripts/harness_env_probe.py |
research, validation |
Report observable Codex runtime config, hooks, tool policy, evidence schema, sandbox fields, and approval fields. |
scripts/harness_requirements.py |
requirements |
Validate requirements artifacts before planning or development treats them as source of truth. |
scripts/harness_agent_team.py |
planning, development |
Validate agent role, scope, write set, verification check, protected harness-state boundaries, and optional durable worker brief before parallel worker dispatch. |
scripts/harness_evidence.py |
validation |
Validate and append structured local evidence events. |
scripts/harness_report.py |
validation, review, handoff |
Summarize local JSONL evidence by phase, event type, time window, or JSON output. |
scripts/harness_checkpoint.py |
handoff, phase transitions |
Append checkpoint entries to docs/harness-state.md with changed surfaces, verification, blockers, and next safe task. |
scripts/verify_codex_env.sh |
validation |
Verify Codex/Claude runtime sync, managed skills, hooks, runtime policies, and supported Codex CLI version. |
main repository regression suite |
validation |
Main repo regression suite for sync, hooks, runtime helpers, skill boundaries, eval fixtures, and docs/config behavior. |
Routing Rules
- Start with
delivery-harness-frameworkwhen work is complex, resumed, cross-session, security-sensitive, release-facing, or ambiguous. - Prefer repo-specific lifecycle harnesses when a known project adapter owns the business domain.
- Prefer gstack skills when the task is product review, engineering review, backlog/spec authoring, design review, browser QA, real-device iOS QA, security review, ship, deploy, canary, release documentation, retro, or learning capture.
- Prefer the richer upstream gstack posture when available:
gstack-office-hoursfor problem discovery, mockup-firstgstack-plan-design-reviewfor UI planning, fix-firstgstack-reviewfor near-landing diffs, coverage-awaregstack-shipfor release readiness, and debt-awaregstack-document-releasefor post-ship docs. With gstack1.52.x, planning skills may use configured gbrain digests as specialist context, but DHF still owns repo state, lane, checkpoint, and validation boundaries; usegstack-plan-tuneonly for explicit question/developer-profile tuning. Also prefergstack-ios-qa/gstack-ios-design-review/gstack-ios-fixfor live-device iOS workflows, vendoredios-sync/ios-cleanonly for explicit iOS debug-bridge maintenance, vendoredcontext-save/context-restorefor user-facing session continuity,landing-reportfor version queue visibility,document-generate/make-pdf/diagramfor generated docs and artifacts, andsetup-gbrain/sync-gbrainonly for explicit persistent memory setup. - Prefer deterministic helper scripts when the runtime already provides one; do not reimplement their parsing or validation manually.
- Use
verification-loopor repo-native checks before any completion claim. - Keep
docs/harness-state.mdas an integrator-owned checkpoint surface for delegated worker plans; workers should report or write slice-local handoff artifacts instead. Single-line main-agent tasks may still append verified checkpoints directly. - Use
harness_checkpoint.py appendbefore ending long-running work, after a meaningful validated slice, or before risky remote/release actions.
Evidence Standard
Any claim that work is complete, fixed, or passing must include fresh evidence:
check
how it exited
what it showed
when it ran
If a check is blocked by credentials, network, remote approval, sandbox, or an external dependency, report the exact blocker and the safest next check.