Delivery Harness Framework / Current Skill Routing Map
Current Skill Routing Map for DHF.
MyCodexEnv uses delivery-harness-framework as the Delivery Harness Framework (DHF) entry point. DHF restores runtime facts, probes the Codex environment, identifies the lifecycle stage, and routes the work to a repo-specific harness, a gstack specialist skill, local development and verification skills, or deterministic runtime helpers. The latest routing gates also cover domain docs / ADRs, vertical slices, feedback loops, throwaway prototypes, and durable agent briefs.
8standard lifecycle stages
9+helper / verification / brief entry points
3generic, repo-specific, gstack roles
4completion evidence fields
Current Flow
Vertical Mermaid flowchart: restore runtime facts, select the right skill/helper, then checkpoint
ShipAI.ca
flowchart TD Start(["User task enters
DHF"]) AGENTS["Read AGENTS.md
README and docs/repo-index.md"] Recover["Run harness_recover.py --json
restore phase, dirty state, next safe task"] Env["Run harness_env_probe.py --json
observe hooks, policy, schema"] NeedsReq{"Are goal, scope,
and acceptance clear?"} Req["Requirements lock
planner / req-to-dev
CONTEXT / ADR"] ReqVal["harness_requirements.py validate PATH"] Classify{"Classify lifecycle stage
and risk boundary"} Repo{"Repo-specific harness available?"} RepoSkill["Hand off to project adapter
business paths, commands, smoke matrix"] Product["Product / scope judgment
gstack-plan-ceo-review / gstack-office-hours"] Plan["Engineering / design planning
vertical slices / AFK / HITL"] Proto["Prototype learning / deep modules
throwaway prototype / deep module"] Dev["Implementation
tdd-guide / atdd-guide / scoped workers"] Team{"Parallel agents?"} TeamVal["harness_agent_team.py validate PLAN.json
optional durable brief"] Val["Verification and evidence
feedback loop / verification-loop / harness_report.py"] QA["QA / security / review
gstack-qa / gstack-cso / gstack-review / code-reviewer"] Ship["Ship / deploy
gstack-ship / gstack-land-and-deploy / gstack-canary"] Docs["Documentation release
gstack-document-release / doc-updater"] Handoff["Handoff
harness_checkpoint.py append"] Done(["DHF output
skill routing, verification evidence, next safe task"]) Start --> AGENTS --> Recover --> Env --> NeedsReq NeedsReq -- "No" --> Req --> ReqVal --> Classify NeedsReq -- "Yes" --> Classify Classify --> Repo Repo -- "Yes" --> RepoSkill --> Dev Repo -- "No, product-oriented" --> Product --> Plan Repo -- "No, engineering/design-oriented" --> Plan --> Proto Repo -- "No, directly implementable" --> Dev Proto --> Team Dev --> Team Team -- "Yes" --> TeamVal --> Val Team -- "No" --> Val Classify -- "Verification" --> Val Classify -- "QA/security/review" --> QA --> Val Classify -- "Ship/deploy" --> Ship --> Val Val --> Docs --> Handoff --> Done classDef runtime fill:#e9f7f5,stroke:#0f9b8e,color:#17241f; classDef decision fill:#f0eaff,stroke:#8a4fff,color:#17241f; classDef build fill:#fbf0dc,stroke:#b26a00,color:#17241f; classDef verify fill:#eaf6f0,stroke:#12805c,color:#17241f; classDef ship fill:#faeaf1,stroke:#d13574,color:#17241f; class AGENTS,Recover,Env,ReqVal,TeamVal,Handoff runtime; class NeedsReq,Classify,Team decision; class Req,Plan,Proto,Dev,RepoSkill,Product build; class Val,QA,Docs verify; class Ship ship;
Skill Routing Map
Helpers provide objective runtime facts; skills provide cognitive workflow routing
| Flow Step | Primary Skill | Helper Command | When to Route Here |
|---|---|---|---|
| Start / recover | delivery-harness-framework | scripts/harness_recover.py --jsonscripts/harness_env_probe.py --json |
Any complex task, resumed session, dirty worktree, cross-stage handoff, or ambiguous request. |
| Requirements | plannerreq-to-devtask-flow-orchestrator | docs/templates/harness-requirements.mdscripts/harness_requirements.py validate PATHCONTEXT.md / CONTEXT-MAP.md / ADR |
The goal, scope, audience, constraints, domain vocabulary, success criteria, or acceptance tests are not locked. |
| Product / business judgment | gstack-plan-ceo-reviewgstack-office-hours | scripts/harness_requirements.py validate PATH when a requirement artifact existsgstack planning skills run gbrain preflight when already configured |
User value, scope, demo boundary, priority, or business tradeoffs need to be clarified; gbrain does not replace repo state. |
| Engineering / design planning | gstack-plan-eng-reviewgstack-plan-design-reviewgstack-plan-tuneplanner | scripts/harness_agent_team.py validate PLAN.json when parallel work is proposeddocs/templates/harness-agent-brief.md when a worker needs a durable contract |
Architecture, API, data flow, migration, visual structure, vertical slices, AFK/HITL, or multi-agent ownership must be decided; repeated question preferences route to plan-tune. |
| Development | tdd-guideatdd-guidescoped worker agents | scripts/harness_agent_team.py validate PLAN.jsongit status --short --branch |
Acceptance criteria are clear and lifecycle policy allows scoped repo writes; worker briefs can validate current/desired behavior, key interfaces, and acceptance criteria. |
| Verification / evidence | verification-loopharness_report.py | python3 test_runner.pygit diff --checkscripts/verify_codex_env.sh ... |
Before declaring completion, after code/config/docs changes, or when a debug task needs a runnable feedback loop first. |
| QA / review / security | gstack-qagstack-qa-onlygstack-reviewcode-reviewergstack-csosecurity-reviewer | git diff, scripts/harness_env_probe.py --json for runtime-sensitive work |
A diff exists, behavior is user-visible, a safety/privacy boundary is involved, or the work is close to handoff or release. |
| Ship / deploy / canary | gstack-shipgstack-land-and-deploygstack-canary | scripts/verify_codex_env.sh ...scripts/harness_checkpoint.py append |
The user explicitly asks for commit, PR, merge, release, deploy, or production validation. |
| Docs / handoff | gstack-document-releasedoc-updaterdelivery-harness-framework | scripts/harness_checkpoint.py appendscripts/harness_report.py --json |
Docs need to reflect delivered behavior, or the session must persist next-safe-task state. |
Routing Rules
1. Runtime facts outrank memory
Run recovery and env probe first. Chat history and subconscious output are hints, not the only source of truth.
2. Requirements gate before planning
If scope, acceptance criteria, or domain vocabulary are unstable, create or validate the requirements artifact and read
CONTEXT.md / ADRs.
3. No parallel writes without a validator
Any worker team must pass
harness_agent_team.py validate to prove write sets do not overlap and do not claim docs/harness-state.md; that global checkpoint is appended by the integrator after verification, while single-line main-agent tasks are outside this worker rule.
4. Completion claims require fresh evidence
Before declaring completion, record
command, exit_code, key_output, and timestamp; before debugging fixes, prove the feedback loop exists.
5. Handoff is a runtime event
At the end of meaningful work slices, use
harness_checkpoint.py append to record command, exit code, key output, timestamp, and next safe task.
6. Repos cannot force global sandbox truth
Env probe may report observable sandbox and approval settings; it must not pretend to know hidden Desktop thread state.
Handoff Assets
Primary English Docs
Chinese Docs
Published by ShipAI.ca as a public DHF reference. The watermark identifies the publisher of this public reference asset.