@agentskit/harness 0.4.0 → 0.5.0

package/docs/PRD-0.4.0.md

@@ -0,0 +1,639 @@
+# Harness 0.4.0 PRD and architecture roadmap
+
+Status: proposal for implementation and release
+
+This document consolidates the ideas extracted from `agentskit-devflow`, the
+AgentsKit Playbook, and the current `@agentskit/harness` implementation. It is
+the working PRD for the Harness as a modular SDLC control engine and loop
+engineer.
+
+## Product objective
+
+The Harness must make agent-assisted delivery predictable, auditable,
+configurable, and measurable without becoming a monolithic replacement for
+agents, orchestrators, trackers, or developer tools.
+
+1. High-quality structure: explicit contracts, predictable organization,
+   directional dependencies, deterministic tests, and current documentation.
+2. Modular composition: capabilities are replaceable through profiles and
+   adapters without editing the kernel.
+3. Controlled SDLC: state machines, workflows, bounded parallelism, caching,
+   agent memory, gates, recovery, and auditability.
+4. Loop engineering: every run measures outcome, cost, speed, failures, and
+   evidence so the next iteration can improve the process.
+
+## Kernel mission and boundary
+
+The Harness is the **SDLC control plane**. It coordinates, validates, measures,
+and enforces contracts. Concrete capabilities are supplied by adapters.
+
+```text
+Harness kernel
+  ├─ controls: phases, state machines, gates, policies, budgets, evidence
+  ├─ optimizes: context, memory, cache, parallelism, and recovery
+  ├─ observes: tokens, cost, duration, CPU, memory, precision, regressions
+  └─ connects: coding agents, Orca, Doc Bridge, code review, trackers, runtime
+```
+
+The kernel must not contain a coding agent, model, UI, mandatory Docker
+sandbox, or concrete GitHub/Linear client. It provides protocols so those
+components can be swapped without rewriting workflows.
+
+### Kernel responsibilities
+
+- start from the issue contract and run the discovery/`grill-me` cycle;
+- surface ambiguities, options, and recommendations before development;
+- stop only at material human decisions defined by the contract;
+- validate documentation, code, context, DoD, and acceptance criteria before
+  opening a PR;
+- orchestrate agents and sub-agents with concurrency and budget limits;
+- persist decisions, artifacts, hashes, events, and evidence;
+- prevent incomplete PRs, changed approved content, and invalid transitions;
+- resume idempotently after failure, timeout, or interruption;
+- measure quality, cost, speed, memory, cache, parallelism, and machine load.
+
+### Out of the kernel
+
+- deciding product ambiguity or business rules autonomously;
+- replacing the coding agent, model, or review tool;
+- maintaining a shadow copy of issues or PRs;
+- requiring Docker when a process runtime is sufficient;
+- treating missing telemetry as zero;
+- closing an issue or claiming production validation without evidence.
+
+## Plugin and adapter contract
+
+Every plugin declares capabilities, inputs, outputs, effects, lifecycle,
+cancellation, errors, and telemetry. The kernel provides run identity, bound
+context, policy, budget, cancellation, and event emission.
+
+| Adapter | Provides | Must return |
+|---|---|---|
+| Coding agent | Analysis, edits, and tests | Structured result, diff, usage, failures |
+| Orca | Dispatch, leases, worktrees | Lease, state, events, recovery data |
+| Doc Bridge | Documentation and knowledge | Sources, hashes, relevance, context used |
+| Code review | Adversarial review | Verdict, findings, severity, confidence, evidence |
+| Memory | Reusable facts and decisions | Hits, origin, validity, context cost |
+| Cache | Reusable results/context | Hit/miss, key, validity, savings estimate |
+| Runtime | Process or Docker execution | Command, limits, attestation, output, resources |
+| GitHub/Linear | PRs, issues, transitions | Remote state, SHA, idempotency key, confirmation |
+
+Adapters may not bypass the state machine, policy gate, budget, or evidence
+binding. Missing telemetry is `unknown`, never an implicit success.
+
+## Reference workflow
+
+```mermaid
+flowchart TD
+  Issue[Linear issue + GitHub code] --> Preflight[Contract and grill-me]
+  Preflight -->|material ambiguity| Human[One human decision packet]
+  Preflight -->|defined| Context[Doc Bridge + memory + cache]
+  Context --> Plan[Structured plan]
+  Plan --> Review[Adversarial code review adapter]
+  Review --> Implement[Coding agent adapter]
+  Implement --> Verify[Tests, lint, eval, runtime]
+  Verify --> Audit[DoD, evidence, quality matrix]
+  Audit -->|failure| Repair[Bounded repair and resume]
+  Repair --> Verify
+  Audit -->|pass| PR[Deterministic PR]
+  PR --> QA[GitHub/Linear: QA]
+  QA --> Production[Real production validation]
+  Production --> Learn[Metrics, memory, profile learning]
+  Learn --> Preflight
+```
+
+`yolo` changes operational pause thresholds only. It never removes product
+decisions, security gates, evidence requirements, or mandatory HITL points.
+
+## Lifecycle and failure semantics
+
+The kernel owns the execution state; GitHub and Linear show projections of that
+state. A feature is not complete when a PR is opened.
+
+```text
+CLARIFYING → PLANNED → IMPLEMENTING → VERIFYING
+     ↑             │          │            │
+     └─ decision ──┘          │            ├─ failed → IMPLEMENTING
+                              │            └─ passed → READY_FOR_QA
+                              └─ failure ───────────────┘
+
+READY_FOR_QA → PRODUCTION_VALIDATING → COMPLETE
+      │                 │
+      └─ QA failure ────┴→ VERIFYING (new source revision invalidates evidence)
+```
+
+`AWAITING_HUMAN_APPROVAL`, `AWAITING_AUTHORIZATION`, and `BLOCKED` are explicit
+states that can interrupt any gated transition. Every transition records the
+actor, reason, source revision, contract hash, evidence IDs, and run ID.
+
+The tracker projection must update Linear to `QA` only after feature validation
+passes. Production validation is a separate gate. A failed QA or production
+check returns the run to verification and invalidates downstream evidence.
+
+## Traceability contract
+
+Every requirement and decision receives a stable ID. The release report must
+map each ID through the complete chain:
+
+```text
+requirement → issue → phase → artifact → check/eval → evidence → gate → release
+```
+
+An item without a current evidence reference is `unverified`. An item changed
+after evidence was produced is `stale`. Aggregate dashboards may summarize this
+chain, but cannot replace criterion-level evidence.
+
+## Current Harness state
+
+| Capability | Evidence | Assessment |
+|---|---|---|
+| Workflow DAG and concurrency | `src/workflow.ts` | Strong base; needs configurable SDLC phase profiles |
+| State machine and verification contract | `src/state-machine.ts`, `src/verification.ts`, `src/types.ts` | Covers transitions and stale evidence |
+| Event log, locks, recovery | `src/events.ts`, `src/agent.ts`, `src/resilience.ts` | Base for diary, replay, and resume |
+| Policy and delivery gates | `src/policy.ts`, `src/delivery.ts` | Base for pre-PR gates and HITL |
+| Process/Docker runtime | `src/runtime.ts` | Docker optional; process mode supported |
+| Doc Bridge, Orca, tracking adapters | `src/adapters/` | Correct direction; keep providers outside the kernel |
+| Metrics, eval, memory, cache | `src/metrics.ts`, `src/eval.ts`, `src/memory.ts`, `src/cache.ts` | Initial instrumentation; bind it to phases |
+| Profiles and context | `src/profiles.ts`, `src/context.ts` | Base for modes and composition |
+
+The current registry in `src/plugins.ts` is the starting point for capability
+composition. The current adapters are projections/planners rather than a full
+integration suite; H-045 and H-047A must prove real adapter behavior wherever a
+component is changed.
+
+Runtime guarantees are mode-specific. Process mode provides shell-free
+execution, timeouts, output limits, and configured environment handling; it does
+not provide filesystem or network isolation. Docker mode adds the configured
+sandbox controls and attestation. An adapter must declare its assurance level
+instead of implying that every runtime has the same isolation.
+
+Cache measurements must distinguish result cache, context cache, and provider
+prefix cache. A hit is not a token saving unless provider usage or a measured
+equivalent is available.
+
+## Agents Playbook practices to adopt
+
+The Playbook is modular; adopt a practice because it prevents a real failure,
+not to create completeness theater. Reference:
+<https://playbook.agentskit.io/llms.txt>.
+
+### Day-zero kernel invariants
+
+- typed boundaries for profiles, plugins, events, artifacts, and adapters;
+- named exports and stable public entry points;
+- stable typed error hierarchy and machine-readable error codes;
+- ADR before architecture changes and RFC before breaking public contracts;
+- verify-first checks for issue, branch, SHA, context, and remote state;
+- honest confidence: `automated-verified`, `claimed`, `not-verified`,
+  `known-not-done`;
+- fail-loud defaults for mandatory dependencies;
+- structured PR intent (`adds`, `changes`, `removes`, `tests`, `docs`);
+- fast quality gates before push and in CI;
+- deny-by-default egress and dependency hygiene.
+
+### Agent behavior
+
+- bootstrap and routing documents that tell an agent what to read and where to
+  edit;
+- versioned prompt and tool registries with hashes and eval results;
+- explicit context selection, ordering, compaction, retention, and cost;
+- grounding, constrained generation, and abstention when evidence is weak;
+- scoped sub-agent contracts; child results never approve the parent task;
+- three-tier eval: deterministic, LLM-as-judge, and production monitoring.
+
+## Event Bridge
+
+Event Bridge is useful, but an external broker is not a day-zero dependency.
+
+Define the contract now:
+
+```text
+eventId, eventType, schemaVersion, occurredAt,
+runId, issueRef, sourceRevision, correlationId,
+payload, idempotencyKey, provenance
+```
+
+Rules: at-least-once delivery, idempotent consumers, compatible schema
+evolution, replay, inspection, dead-letter handling, and backpressure that
+cannot stall the main workflow.
+
+Implementation order:
+
+1. Use the existing append-only event log as the source of truth and define an
+   `EventSink`/`EventBridge` interface.
+2. Add a queue, pub/sub, or stream adapter only when real independent
+   consumers require it.
+
+Do not select Kafka, NATS, Redis, or another broker before measuring volume,
+durability, and latency needs.
+
+## MCP
+
+MCP is an integration boundary, not the internal orchestration layer. The
+Harness must remain fully usable without MCP.
+
+Initial read-only surface:
+
+- discover profiles and capabilities;
+- inspect runs and their lifecycle state;
+- read artifacts, evidence, decisions, and blockers;
+- inspect gate status and quality metrics;
+- request bounded Doc Bridge or memory context.
+
+Start, pause, resume, and cancel are mutating operations even when they do not
+touch an external service. They require a separate gated MCP surface with
+authorization, idempotency, and audit evidence. Arbitrary shell execution, PR
+publication, remote issue transitions, worktree cleanup, secrets, and
+out-of-scope context are never exposed by the default profile.
+
+Implement the capability manifest first, then a local/stdio read-only MCP
+adapter. Add mutating operations only after authorization, idempotency, and
+audit evidence are proven.
+
+## Consolidated execution plan
+
+### Phase 0 — Baseline and architectural contract
+
+- inventory modules, imports, and public exports;
+- classify modules as kernel, execution, context, delivery, profile, or adapter;
+- record forbidden dependencies and justified exceptions;
+- capture typecheck, test, build, pack, and metric baselines;
+- create the versioned eval manifest, golden corpus, rubrics, and no-Harness
+  baseline before feature implementation;
+- record provider/model, prompt/tool versions, repetitions, seed/temperature
+  where applicable, and thresholds for every eval;
+- write the kernel/adapters ADR and plugin contract.
+
+Output: module map, dependency rules, extension contract, and reproducible eval
+baseline.
+
+### Phase 1 — Kernel/adapters organization
+
+Target capability-first layout:
+
+```text
+src/
+  kernel/       # workflow, machine, policy, events, evidence, hashes
+  execution/    # runtime, recovery, coordination, budgets, metrics
+  context/      # context contracts, memory, cache
+  delivery/     # gates and deterministic composition
+  adapters/     # Orca, Doc Bridge, GitHub, Linear, review, providers
+  profiles/     # declarative capability composition and modes
+```
+
+Move mechanically by capability, preserve `src/index.ts` re-exports, run
+typecheck/tests after each group, and do not create folders for one-file layers
+without a real boundary.
+
+### Phase 2 — Deterministic SDLC engine
+
+- define profiles with phases, inputs, outputs, dependencies, gates, retries,
+  and budgets;
+- run preflight/`grill-me` before mutation;
+- aggregate ambiguities into one human decision packet;
+- continue automatically when the contract is complete;
+- support `safe`, `yolo`, and `dry-run` through one engine;
+- reject cycles and unbounded retries.
+
+### Phase 3 — Artifacts, decision log, and recovery
+
+- implement a versioned `ArtifactEnvelope` in JSON plus a readable form;
+- bind artifacts to run, issue, SHA, contract, configuration, and context;
+- record plan, findings, decisions, repairs, and blockers;
+- resume from the event log without repeating completed effects;
+- invalidate stale artifacts and evidence.
+
+### Phase 4 — Operational adapters
+
+- coding agent: structured output, diff, usage, and failures;
+- Orca: leases, deterministic worktrees, issue locks, remote SHA confirmation,
+  and safe cleanup;
+- Doc Bridge: source, hash, relevance, and context cost;
+- memory/cache: hit/miss, validity, scope, and measured savings;
+- process/Docker runtime selected by profile;
+- GitHub/Linear effects idempotent and evidence-bound.
+
+### Phase 5 — Delivery gates
+
+- review and audit before opening a PR;
+- derive PR intent from structured fields;
+- bind approved PR content to a hash;
+- publish only after checks, DoD, and audit pass;
+- transition Linear to QA after feature validation;
+- validate production before closing the delivery;
+- confirm remote branch/SHA before worktree cleanup.
+
+### Phase 6 — Efficiency and observability
+
+- measure input/output/cache tokens, cost, and duration per phase;
+- measure memory/context retrieval, cache hit/miss, and relevance;
+- measure parallelism, contention, CPU, RAM, and saturation;
+- classify failures and enforce watchdog budgets;
+- fix provider/model for comparison experiments;
+- never convert missing telemetry to zero.
+
+### Phase 7 — Loop engineer
+
+- maintain real issues and a baseline without the Harness;
+- run evals by version, profile, and fixed provider/model;
+- score each quality dimension from 0 to 100;
+- produce periodic results and blocker packages;
+- promote approved learnings into profile or adapter changes;
+- repeat: run → measure → review → adjust → validate.
+
+## Release scope for 0.4.0
+
+### In scope
+
+- kernel/adapters separation with compatible public exports;
+- versioned capability, event, and error contracts;
+- configurable phase profiles and bounded retries;
+- versioned artifacts, decision log, and idempotent resume;
+- structured preflight/`grill-me` and adversarial verdicts;
+- deterministic PR intent and immutable content binding;
+- per-phase token, cost, cache, memory, duration, parallelism, CPU, RAM, and
+  failure metrics;
+- contracts for coding agents, Orca, Doc Bridge, code review, memory, cache,
+  runtime, GitHub, and Linear;
+- `safe`, `yolo`, and `dry-run` profiles;
+- quality gates, eval battery, documentation, and extension examples;
+- real integration evals for every touched AgentsKit component, with pinned
+  versions and release evidence;
+- Event Bridge contract and capability manifest for future MCP integration.
+
+### Out of scope
+
+- external event broker;
+- mutating or remote MCP server;
+- dynamic plugin marketplace/loader;
+- custom dashboard or UI;
+- mandatory concrete provider/tracker clients;
+- unmeasured optimizations or breaking migration of all consumers.
+
+## Deliverables and issue breakdown
+
+Each issue is a vertical unit with a contract, evidence, and rollback plan.
+
+### H-040 — Baseline and boundary map
+
+- [ ] inventory modules, imports, and public exports;
+- [ ] classify every module and record dependency exceptions;
+- [ ] record reproducible baselines;
+- [ ] create the versioned eval manifest and no-Harness baseline;
+- [ ] define rubrics, thresholds, repetition counts, and comparison policy;
+- [ ] create the kernel/adapters ADR.
+
+DoD: reviewed map, reproducible test/metric/eval baseline, and no new runtime
+behavior.
+
+### H-041 — Capability, event, and error contracts
+
+- [ ] version plugin/capability interfaces;
+- [ ] standardize lifecycle, cancellation, timeout, effects, and telemetry;
+- [ ] standardize event envelope and idempotency key;
+- [ ] standardize stable error codes;
+- [ ] validate schemas at boundaries;
+- [ ] publish a static capability manifest.
+
+DoD: contracts compile and pass round-trip/compatibility tests.
+
+### H-042 — Physical organization and API compatibility
+
+- [ ] move modules only across real capability boundaries;
+- [ ] preserve `src/index.ts` re-exports;
+- [ ] remove kernel imports of adapters;
+- [ ] add directional dependency tests;
+- [ ] update organization docs and ADRs.
+
+DoD: typecheck, consumer tests, and package checks pass without public API
+breakage.
+
+### H-043 — Phase profile and deterministic executor
+
+- [ ] define phase schema with inputs, outputs, dependencies, gates, retries,
+      and budgets;
+- [ ] compose over existing workflow, state machine, and policy primitives;
+- [ ] run preflight before mutation;
+- [ ] aggregate human ambiguity decisions;
+- [ ] implement safe, yolo, and dry-run profiles;
+- [ ] reject cycles and infinite retries.
+
+DoD: the same contract yields the same route; pass/block/escalate/retry,
+cancel, and resume are covered.
+
+### H-044 — Artifacts, decision log, and resume
+
+- [ ] implement versioned `ArtifactEnvelope`;
+- [ ] bind artifacts to run, issue, SHA, contract, config, and context;
+- [ ] record plans, findings, decisions, repairs, and blockers;
+- [ ] resume idempotently from the event log;
+- [ ] invalidate stale evidence;
+- [ ] provide human-readable and JSON inspection.
+
+DoD: interruptions at every phase are resumable or honestly blocked.
+
+### H-045 — Execution and context adapters
+
+- [ ] adapt coding agents with structured results, diffs, usage, and failures;
+- [ ] complete Orca lease/lock/worktree and remote SHA confirmation;
+- [ ] enrich Doc Bridge with relevance, source, hash, and cost;
+- [ ] measure memory and cache scope, validity, hits, and savings;
+- [ ] retain process/Docker selection by profile;
+- [ ] keep GitHub/Linear effects idempotent.
+
+DoD: every adapter has a fake/dry-run path and cannot bypass policy or evidence.
+
+### H-046 — Adversarial review and delivery gates
+
+- [ ] add configurable parallel review lenses;
+- [ ] require evidence/reproduction for findings;
+- [ ] treat missing reviewers as unverified/block;
+- [ ] generate PR intent from structured fields;
+- [ ] bind PR content to G2/G3 hashes;
+- [ ] publish only after checks, DoD, and audit pass;
+- [ ] transition Linear to QA after feature validation.
+
+DoD: incomplete, modified, or unaudited PRs cannot be published.
+
+### H-047 — Telemetry, eval, and quality matrix
+
+- [ ] record per-phase and per-adapter metrics;
+- [ ] measure tokens, cost, duration, CPU, RAM, memory, cache, and parallelism;
+- [ ] classify failures and enforce watchdog budgets;
+- [ ] run the eval battery on every relevant change;
+- [ ] produce 0–100 scores and baseline deltas;
+- [ ] keep missing telemetry explicitly unknown.
+
+DoD: reproducible report covers quality, cost, speed, precision, and resources.
+
+### H-047A — Eval battery and touched-component impact
+
+Normal tests validate implementation. Evals validate behavior and quality;
+both are mandatory.
+
+The eval manifest is created in H-040 and extended by every issue. A feature
+cannot enter implementation without a manifest entry describing its expected
+behavior, affected components, grader, threshold, and evidence output.
+
+Required layers:
+
+1. contract eval: schemas, exports, errors, events, idempotency, compatibility;
+2. deterministic behavior: state, workflow, gates, retries, cancel, resume,
+   and dry-run;
+3. integration eval for every touched real adapter;
+4. quality eval for completeness, precision, grounding, evidence, escalation,
+   and PR quality;
+5. regression/golden eval against previous version and no-Harness baseline;
+6. resource eval for tokens, cache, memory, duration, parallelism, CPU, RAM,
+   and cost.
+
+| Component | Required eval focus |
+|---|---|
+| Core/state machine | Contracts, determinism, replay, policy, compatibility |
+| Workflow | Legal/illegal transitions, cycles, retry, cancel, fan-out/fan-in |
+| Memory | Scope, relevance, contamination, TTL, redaction, context budget |
+| Cache | Keys, validity, invalidation, isolation, measured savings |
+| Doc Bridge | Recall/precision, source/hash, stale rejection, cost |
+| Agent/model adapter | Output schema, tool calls, usage, timeout, abort, failures |
+| Orca/worktree | Claims, locks, deterministic branch, conflicts, resume, cleanup |
+| Runtime | Limits, process/Docker parity, egress, attestation, cancellation |
+| Code review | Independent lenses, refutation, reproduction, missing reviewer |
+| GitHub/Linear | Idempotency, states, SHA, PR intent, transitions |
+| Eval/metrics | Calibration, reproducibility, false positives/negatives |
+
+Minimum thresholds:
+
+- critical contracts, security, policy, provenance, and idempotency: 100% pass;
+- required deterministic corpus: 100% pass;
+- subjective quality: manifest threshold, initially 80/100;
+- no dimension regresses more than 5 points without a recorded decision;
+- `unknown`, `unverified`, and `stale` never count as approval;
+- a failure in a touched component blocks dependent work;
+- prompt, model, memory, cache, or adapter changes require comparative eval.
+
+### H-047B — AgentsKit ecosystem compatibility eval
+
+This issue closes the gap between adapter interfaces and real ecosystem
+behavior. It is required whenever a change touches an AgentsKit package,
+repository, prompt, or provider integration.
+
+- [ ] pin the exact versions/revisions of every touched AgentsKit component;
+- [ ] run the component's own tests and documented evals;
+- [ ] run the Harness integration eval using the real adapter boundary;
+- [ ] compare behavior with the previous version and no-Harness baseline;
+- [ ] record compatibility, migration, and rollback evidence;
+- [ ] publish or update the affected package only after its eval gate passes;
+- [ ] record the published artifact/version in the Harness release manifest.
+
+The initial component matrix includes core, memory, adapters, runtime,
+Doc Bridge, code review, and eval. Untouched components receive a compatibility
+smoke test; touched components require the full relevant battery. A passing
+Harness test with a failing upstream component eval is a release blocker.
+
+### H-048 — Documentation and adoption
+
+- [ ] update README with kernel/adapters architecture;
+- [ ] document minimum profile, minimum adapter, and dry-run;
+- [ ] add agent, Doc Bridge, code-review, and tracking examples;
+- [ ] document extension and troubleshooting;
+- [ ] update ADRs, changelog, migration notes, and capability manifest.
+
+DoD: a consumer can install, run a fake profile, and write an adapter without
+reading internals.
+
+### H-049 — Hardening and release candidate
+
+- [ ] run unit, contract, integration, CLI, and consumer-tarball tests;
+- [ ] run `ak-verify` against the current contract;
+- [ ] validate build, pack, exports, docs, and generated output;
+- [ ] run the complete eval battery and fixed-provider pilot benchmark;
+- [ ] run every required AgentsKit component eval for the actual touched
+      versions/revisions;
+- [ ] verify that CI release gates consume the same eval manifest and evidence
+      bundle as the release candidate;
+- [ ] perform adversarial review of the full diff;
+- [ ] close or classify every blocker.
+
+DoD: no P0 criterion is pending and the release candidate is reproducible from
+a clean revision.
+
+## 0.4.0 quality gates
+
+| Dimension | Release gate |
+|---|---|
+| Contracts/provenance | 100% external interfaces schema-validated and hash-bound |
+| Determinism | 100% reference-profile cases reproducible |
+| Eval battery | 100% required layers and touched-component evals executed |
+| Security | Zero policy, egress, or secret violations |
+| Evidence | 100% pilot criteria have current evidence |
+| Delivery | Zero PRs without intent, DoD, and audit |
+| Recovery | Every interruption point resumes or blocks honestly |
+| Efficiency | Baseline recorded; no >10% regression without justification |
+| Technical quality | Typecheck, tests, build, pack, and consumer checks green |
+| Documentation | README, organization, ADRs, changelog, and examples current |
+| Observability | Missing token/cost/resource data remains explicitly unknown |
+
+An aggregate score cannot compensate for a failed P0 gate. The release cannot
+ship with any required criterion `blocked`, `unverified`, or `stale`.
+
+### Scoring and reproducibility
+
+Each dimension is a versioned set of weighted criteria. For criterion `i`,
+`statusScore(i)` is `1` for pass, `0.5` for explicitly partial, and `0` for
+fail, unknown, unverified, or stale. `not-applicable` is removed from the
+denominator only with a recorded reason.
+
+```text
+dimensionScore = round(100 × Σ(weight[i] × statusScore(i)) / Σ(weight[i]))
+overallScore   = round(Σ(dimensionWeight[d] × dimensionScore[d]) /
+                      Σ(dimensionWeight[d]))
+delta          = candidateScore − baselineScore
+```
+
+The manifest fixes the criteria, weights, corpus version, provider/model,
+prompt/tool hashes, profile, configuration hash, repetition count, and
+seed/temperature where supported. Deterministic checks run once; nondeterministic
+provider or judge checks use the manifest repetition count and report
+min/median/max. A critical criterion is a veto: it must be `1`, regardless of
+the aggregate score. A score without these inputs is not comparable.
+
+## Dependency order
+
+```text
+H-040 → H-041 → H-042 → H-043 → H-044
+                              ↘ H-045 → H-046 → H-047 → H-047A → H-047B → H-048 → H-049
+```
+
+H-045 may start after H-041 when it uses stable contracts. H-046 requires
+H-043 and H-044. H-047/H-047A start early for baseline collection but close
+only after every touched adapter and AgentsKit component is evaluated. H-047B
+may run in parallel for independent components, but must close before H-048 and
+H-049. H-049 is hardening and release, not a place for new features.
+
+## 0.4.0 publication checklist
+
+1. Merge H-040 through H-049 into `main` in dependency order.
+2. Update `package.json` to `0.4.0` and add the changelog entry.
+3. Run `pnpm typecheck`, `pnpm test`, `pnpm build`, and `pnpm pack`.
+4. Run the real CLI flow, `ak-verify`, the complete eval battery, and pilot
+   benchmarks.
+5. Confirm revision, contract, configuration, and run IDs in the final report.
+6. Merge to `main`; Trusted Publishing releases only when the package version
+   changes and never uses `NPM_TOKEN`.
+7. Verify the registry tarball and `@agentskit/harness@0.4.0`.
+8. Run a smoke test as a consumer of the published package.
+9. Record publication evidence and start the next baseline.
+
+Any failure leaves the release `BLOCKED`; there is no manual bypass.
+
+## Open decisions
+
+- final profile format: JSON, YAML, or validated TypeScript;
+- artifact persistence: local filesystem only or storage adapter;
+- mandatory phases for the pilot profile;
+- quality-score weights and HITL thresholds;
+- retention and redaction policy for memory/context logs;
+- first real-issue corpus and no-Harness baseline.
+
+These decisions require a human choice when they change product behavior, risk,
+cost, or policy. Agents should provide options, evidence, and a recommendation.