agentleFS
PhlegonlabsPhlegonlabs/product-delivery-harness/skills/delivery-harness/SKILL.md

delivery-harness

Skill1 starsChanged 18 days agoMITView on GitHub

Use only when the user explicitly requests delivery-harness or a requested skill workflow requires it. Route engineering work through planning, authorization, execution, verification and integration. Use direct delivery for one writer or a Direct Fan-Out with one verification sequence; use PLAN-v6/RUN-v11 for durable coordination, isolated integration or a bounded correction graph. Use agent-configured executors through observed tools or explicitly bound bridges.

The file

Show its rules and text

We split this skill into the rules it gives an agent

16 rules. About 0 are common, 13 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈13

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run only on the user's explicit request, continuation, or a required dependency of that requested skill workflow.

    line 8
  2. Very specific:

    Before Product-package execution, apply references/pre-delivery-self-review.md#harness-entry.

    line 18
  3. Distinctive:

    Use PLAN/RUN only when coordination requires it.

    line 24
  4. Distinctive:

    Before any frontend dispatch, apply Frontend Visual Inputs to the actual author task.

    line 107
  5. Distinctive:

    Create an authorized local commit when requested.

    line 111
  6. Very specific:

    Use this routing only when the size gate classifies work as large.

    line 120
  7. Distinctive:

    Before authoring PLAN, read Managed Route, Reference Routing, and Plan Large Work.

    line 122
  8. Distinctive:

    Before readiness and dispatch, read Default Runtime And Wave Policy, Default Mission Topology, and [Pass Plan

    line 123
  9. Distinctive:

    Before managed runtime use, read Adapter Routing.

    line 124
  10. Distinctive:

    Before execution, read Execute And Integrate.

    line 125
  11. Distinctive:

    Before verification, read Verify Local-First.

    line 126
  12. Distinctive:

    Before closeout, read Complete.

    line 127
  13. Distinctive:

    Follow the current host's effective instruction precedence.

    line 135
  14. Distinctive:

    Keep automatic context discovery enabled.

    line 136
  15. Very specific:

    Read references/ui-implementation-contract.md before UI implementation or review.

    line 140
  16. Distinctive:

    Run System Review And Route.

    line 153

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Delivery Harness

Run only on the user's explicit request, continuation, or a required dependency of that requested skill workflow. A matching task, installation, discussion, source edit or existing project artifact does not activate it. Required checks and dependencies stay within the requested outcome; a new top-level stage needs an explicit request.

Commands: references/installed-commands.md; stage cards: references/stages/stage-cards.md; launch brief: assets/templates/HARNESS_GOAL.template.md.

At every invocation, run the entry checklist. For new delivery work, direct or managed, acceptance follows references/delivery-acceptance-contract.md. English PRD/architecture remain implementation authority, not Chinese review copies.

Design And Maintenance Routing

Before Product-package execution, apply references/pre-delivery-self-review.md#harness-entry; missing or stale evidence blocks execution, not inspection or planning. Classify UI work under Product Contracts.

Optional context: references/reference-selection.md (references/option-library/).

Purpose

Use PLAN/RUN only when coordination requires it. Preserve upstream ownership:

  • product-definition-builder owns approved PRD.md, architecture.md, and stack-decisions.md, including complete frontend/backend architecture and technology decisions.
  • ui-design-builder owns UI Design Intake, PRD-led direction studies, HiFi checks, Visual Approval and the HiFi target. design-system-compiler owns a required design-system pair.
  • Implement current approved product/stack and applicable UI sources. Every must-have UX-* trace still needs objective evidence. Recommended and Provisional stack rows are proposals. Do not invent product, copy, stack or design decisions.
  • The security lens in references/security-review.md owns read-only review of the fixed integrated SHA; it neither remediates nor probes live targets.

Project Size Gate

Classify work as small when one writer can own one bounded outcome in one checkout and verify it with one coherent local sequence; that writer may be the required delegated specialist. small also covers a few independent outcomes in different files (Direct Fan-Out). A high file count, language count, test duration, or reasoning difficulty does not make work large by itself.

Classify work as large only when at least one condition is true:

  • a mission needs another mission's integrated output;
  • a shared contract must be frozen, reviewed and integrated before dependents start;
  • execution needs durable cross-session or cross-host handoff;
  • a migration, destructive operation, or correction loop needs a durable bounded graph; or
  • the user explicitly requests managed Harness orchestration.

Planning preference may choose the stricter route, but cannot override a large classification or any safety boundary.

System Review And Route (parent-only, read-only)
  small -> direct inspect -> implement -> local verify -> review -> authorized Git actions
  large -> planner -> PLAN v6 + RUN v11 -> readiness -> managed execution

Small work creates no PLAN/RUN files. If scope grows, preserve evidence and plan the remainder. The Sequential Parent Route runs one mission at a time (execution-state-model.md). Preserve required independent reviews.

System Review And Route

Within an explicitly requested delivery-harness workflow, complete this checkpoint before loading any task-specific skill. Read the request, repository instructions, Git state, requested scope, and relevant frozen product/design inputs. Do not create or edit PLAN.md, RUN.md, tasks.md, branches, worktrees, evidence, or implementation files during this checkpoint.

Record:

Project size: small | large
Delivery profile: standard | mvp | program
Intent: plan-only | plan-then-stop | plan-then-execute | execute-ready-plan
Route: direct | plan-backed graph
Host adapter: none | general (observed host identity or generic)
RUN landing: local_only
Post-archive publication: none | exact candidate branch
Upstream inputs: present | missing | needs owner decision
Gitignore impact: none | update | needs owner decision
Planned review points: <n> (<reasons>)

For UI work, inspect only docs/design/, a user-named design folder, and obvious in-scope images from the normal scan, record candidate paths and hashes, then apply references/design-input-updates.md; only PRD-approved consequences and any required compiled design-system pair influence implementation.

Classify each task's Gitignore impact with references/gitignore-contract.md; it applies to both direct and managed routes.

When an existing RUN is running, perform the Resume Reconciliation Gate in references/execution-state-model.md before selecting work. Start with python "<delivery-harness-skill-root>/scripts/inspect_harness_run.py" --repo-root <target-root> --session-id <id>; its Next block names the state and one recommended command, and a running RUN with no lock needs one owner decision before selection. Inspect host process/session evidence separately. Canonical state, live process state, Git heads, and dirty worktrees are separate evidence; never assume worker_running proves a live worker. For a package pinned to an earlier Harness version, also read Legacy Pins.

After capability detection, apply references/runtime-upgrades.md. Only an old compatible runtime's active wave may finish. An incompatible or restarted runtime waits for a fresh probe, then re-orchestrates every remaining task onto the new runtime through new attempts using that RUN's exact pinned bundle. Provider changes require explicit replanning; never hot-upgrade a worker or silently mutate installed runtime software.

MVP Week

The owner-selected mvp delivery profile targets one week from approved PRD to a verified release candidate for a small project or an MVP. It runs one combined review plus one re-check in delivery, folds Final UI Acceptance into the combined review, isolates every post-re-check defect, runs the broad suite once, and keeps at most four owner gates across the flow. Speed changes nothing else: the authorization ledger, the security lens on executable diffs, data safety, and exact-SHA binding still apply.

Non-Negotiable Boundaries

These rules apply to both routes:

  • Selecting this skill grants no mutation permission. Bind each state-changing action to the user's exact instruction and target.
  • Preserve all 12 managed action keys in the Authorization Action Ledger.
  • Execution intent covers only applicable local setup, branch, commit, and integration actions; it does not authorize push. Harness 0.38+ RUNs stay local_only with push false; after close, archive, publication and promotion follow stage Card 8.
  • Freeze the remote main base in PLAN branch_policy; retain historical RUN pins. Never edit/commit on a protected branch or recreate one as a run branch. Follow references/branch-promotion-contract.md#delivery-kind-and-base and references/branch-promotion-contract.md#protected-branches; never force-push. Ask before creating that branch unless a signed launch brief names it; never add a fixed prefix.
  • Archival, worktree removal, and branch deletion are separate actions, never implied by completion.
  • The parent owns routing, authorization, PLAN/RUN, dispatch, leases, integration, and lifecycle actions. Workers and reviewers receive only their assigned actions. Pinned managed protocols retain parent-owned PLAN/RUN, integration and lifecycle transitions.
  • A managed runtime review launches only from a persisted reserve-review-dispatch receipt for the selector's current directive. A raw runtime spawn is unplanned work; do not accept its result or reconstruct a receipt afterward.
  • Non-mission nodes reserve with reserve-node-attempt, execute outside the RUN lock, then record evidence with record-node-result; interrupted attempts become blocked.
  • A review PASS binds one exact SHA. Any repair invalidates it for the paths the repair touches; on the direct route, scope-disjoint bookkeeping commits carry it forward, and a scoped repair PASS re-binds the repaired paths (references/verification-gates.md#review-topology).
  • Preserve failed, interrupted, cancelled, dirty, and partial worktree evidence. Never reset or remove it automatically.

Direct Route

For initial product implementation, read Initial Product Foundations and order prerequisites before dependent features; the Surface Catalog lists each surface class's required early missions. Implement local hooks first; record fixed-build external setup as separately authorized action rows from references/external-setup.md in docs/DEPLOYMENT.md; deployment operations stay separate.

For unproven cross-layer dependencies, apply Early Journey And Engineering Evidence within existing tasks and checks.

  1. Inspect the bounded component and relevant instructions. Before any frontend dispatch, apply Frontend Visual Inputs to the actual author task.
  2. For a product-definition-builder package, verify its Product Definition Approval and Stack Decision Checkpoint plus the applicable self-review and UI handoff under references/pre-delivery-self-review.md; then implement with one writer or a Direct Fan-Out, under references/delegation-contract.md. Agent policy designates the UI writer; preserve its complete design inputs, required skills and independent acceptance. When feature files exist, render the acceptance register (scripts/render_delivery_acceptance.py) and record its hash before implementing.
  3. Probe the toolchain, browser and reviewer route once (a failed row is an environment block), then run focused checks selected from changed files, security tests included.
  4. Review the complete diff and run git diff --check.
  5. Create an authorized local commit when requested.
  6. Dispatch one independent read-only reviewer on that SHA carrying every lens the diff needs: code; security, with references/security-review.md loaded; and, when UI changed, visual with a real browser and one Final UI Acceptance. It returns one consolidated finding set of blocking findings and non-blocking notes. The security lens applies to every executable diff (references/verification-gates.md#review-topology); a docs-only diff records security: not_applicable (<reason>) with the changed path list, and an executable diff without the security lens is UNVALIDATED, not PASS.
  7. Make one consolidated repair, then present the focused verifier PASS at the new head before one targeted re-check of the prior blocking findings and the repair range (references/gen-verify-correct.md). After the re-check, every remaining defect enters the Isolated Repair Protocol.
  8. Perform only remaining authorized Git actions; for promotion, read references/branch-promotion-contract.md#candidate-gate and references/branch-promotion-contract.md#promote-to-main.

A text-only markup/style review cannot claim a visual PASS or close a required visual gate. Obey references/ui-implementation-contract.md, including rule-8 UI-impact classification and same-change doc updates.

Managed Stage Routing

Use this routing only when the size gate classifies work as large. For direct work, do not read the managed stage file. Start each stage from its card in references/stages/stage-cards.md; read the sections below when the card names them or a command fails.

  • Before authoring PLAN, read Managed Route, Reference Routing, and Plan Large Work.
  • Before readiness and dispatch, read Default Runtime And Wave Policy, Default Mission Topology, and Pass Plan Readiness.
  • Before managed runtime use, read Adapter Routing.
  • Before execution, read Execute And Integrate.
  • Before verification, read Verify Local-First.
  • Before closeout, read Complete.

Repository Context Contract

Discover the effective instruction chain from repository root to the selected checkout.

  • Existing AGENTS.md/AGENTS.override.md/CLAUDE.md are user-owned; never overwrite, normalize, silently copy.
  • Entry runs scripts/configure_project_context.py --root <target-root> --check --merge-agents; without --check, only an authorized bootstrap creates missing files from PROJECT_AGENTS.template.md and PROJECT_CLAUDE.template.md; generated files differ. Existing AGENTS stays proposal-only until a reviewed plan binds original/template hashes, selected headings and acknowledged conflicts. Same-heading differences are unresolved semantic divergences. Resolve new Skill Bindings from observed skills with owner confirmation.
  • Follow the current host's effective instruction precedence; never guess it from a provider name or inject another host's instructions.
  • Keep automatic context discovery enabled. A nested instruction may narrow a mission, never widen write scope or authorization.

UI Implementation Contract

Read references/ui-implementation-contract.md before UI implementation or review.

  • design-system-compiler alone owns compilation when the Design System Need Gate is required.
  • UI implementation runs the bound frontend author under this Harness's conformance contract. Do not claim that skill defines conformance mode or reopen Style Integration.
  • Initial/enhancement system-conformance uses current approved PRD, UI design, HiFi and pair; their responsive sets must agree. Target-conformance uses the approved target when the gate is not_required. Routine maintenance follows references/ui-implementation-contract.md; a missing required input is a design-input delta.
  • A page-faithful target binds implementation only after the user explicitly requests faithful conformance.
  • Final UI Acceptance runs the Final Visual Parity Loop and the page-quality pass once, on the exact head, in the one review (references/verification-gates.md). Impeccable is not the default reviewer.
  • Bundled defaults exist only for bundled skills. A project's owner-confirmed Skill Bindings table may bind installed external visual-direction, frontend-authoring, or UI-quality tools after their full trees and side effects are checked; an unresolved or incompatible slot blocks its dependent node.

Workflow

1. Intake And Route

Run System Review And Route; the Resume Reconciliation Gate above covers a running RUN.

If the user pauses or cancels a managed run, apply the durable control transition subcommand of scripts/harness_transition.py — pause, resume, or cancel, each requiring --source. A conversational stop is not scheduler state. Preserve active and dirty worktrees, then reconcile interrupted mission workers with reconcile-interrupted and stopped review workers together with reconcile-interrupted-reviews before any resume.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.