AI harness pipeline — chain contract, progressive disclosure, DoD gates

frameworkactive owner: philip · updated 2026-11-14

relates: build-process · agent-skill-packaging · dual-deployment-serverless-container

AI harness pipeline — chain contract, progressive disclosure, DoD gates

The delivery pipeline is itself a product: two pipeline skills (feature-analysis, development) ride a chain contract — every SKILL.md declares what it consumes, what it produces, its gate, and the next step. Context is progressively disclosed per step; quantitative DoD measures live in one place; human interaction is bunched and asynchronous.

Problem it solves

• Multi-step AI delivery drifts: context bloat between steps, unmeasured quality, drip-fed questions that babysit the user. The harness makes the pipeline inspectable (each step's contract is data) and the human burden minimal by design.

Shape

• Pipeline skills: pipeline/skills/feature-analysis (collect → refine → prioritise → allocate → issue file in issues/) and pipeline/skills/development (design → build → store → verify → deploy → secure). One directory per step, one SKILL.md playbook each.
• Chain contract: every step declares consumes / produces / gate / next — the hand-off is explicit, machine-checkable, and resumable.
• Progressive disclosure: machine/harness.dsl is always-on (the harness's own KB); each step loads only its SKILL.md + declared pre-read: files. Same T0→T3 discipline as the knowledge base (kb:kb-retrieval-token-budget).
• Definition of Done (issues/definition-of-done.md): the flow of skill steps + quantitative outcome measures; each skill lists the measures it owns — quality is measured per step, not vibes at the end.
• Tone/goals (TONE.md): minimise bot-sitting (user answers, never watches) and end-user fatigue (few interactions, business language only, no jargon); questions batch across sub-steps and issues into one async Q/A set (fillable PDF or Q&A session) that feeds back so flow resumes (issues/async-questioning.md).
• Hook gates (hooks/modules/): dod_gate.py and dsl_drift.py run as PreToolUse hooks with per-agent adapters (hooks/adapters/claude.settings.json, devin.hooks.v1.json) — the DoD and DSL drift checks fire automatically, not by memory.
• Issue files (issues/): async-questioning, definition-of-done, linear-seam, looper — the pipeline's durable state lives in files, not chat.

Rules

• Every pipeline step declares consumes/produces/gate/next — no implicit hand-offs.
• Questions to the human are bunched and async; the AI does everything it can between touchpoints.
• User-facing language is business language — if a question can't be asked in business terms, the AI isn't ready to ask it.
• DoD measures are quantitative and owned per step; the loop (looper.md) repeats until measures pass.
• Gates run as hooks with per-agent adapters — enforcement is mechanical, not remembered.
• Pipeline state lives in issue files — resumable across sessions and agents.

Anti-patterns

• Steps that produce undocumented outputs (next step guesses).
• Drip-fed questions — one question per turn babysits the user.
• Jargon in user-facing questions.
• Quality checked only at the end (no per-step measures).
• Gates that depend on the agent remembering to run them (hooks exist for this).
• State kept in chat instead of issue files (lost on session end).

Sources

• factory/ai-harness/pipeline/README.md, pipeline/skills/{feature-analysis,development}
• factory/ai-harness/issues/{definition-of-done,async-questioning,linear-seam,looper}.md
• factory/ai-harness/TONE.md, machine/harness.dsl
• factory/ai-harness/hooks/modules/{dod_gate.py,dsl_drift.py}, hooks/adapters/
• knowledge-base/factory/build-process.md