@harness-engineering/cli 2.8.0 → 3.0.0

package/dist/agents/skills/claude-code/initialize-harness-project/SKILL.md

@@ -29,6 +29,8 @@ Detect plugin-only state by checking whether `harness` resolves on PATH (`comman
 
 ## Process
 
+**Prompt the human in plain text** — every framework confirmation, migration check, and telemetry-identity question in this skill is plain text only. Do not elevate to `AskUserQuestion`: the framework list (~10 options) exceeds its 4-option cap and natural headers like "Confirm framework" exceed its 12-char cap, rendering the call as ERR.
+
 ### Phase 1: ASSESS — Determine Current State
 
 1. **Check for existing harness configuration.** Look for `.harness/` directory, `AGENTS.md`, `harness.config.json`, and any skill definitions. Their presence determines whether this is a new project or a migration.
@@ -96,7 +98,7 @@ Detect plugin-only state by checking whether `harness` resolves on PATH (`comman
 5. **Configure i18n (all levels).** Ask: "Will this project support multiple languages?" Based on the answer:
    - **Yes:** Invoke `harness-i18n-workflow` configure phase to set up i18n config in `harness.config.json` (source locale, target locales, framework, strictness). Then invoke `harness-i18n-workflow` scaffold phase to create translation file structure and extraction config. Set `i18n.enabled: true`.
    - **No:** Set `i18n.enabled: false` in `harness.config.json`. The `harness-i18n-process` skill will still fire gentle prompts for unconfigured projects when features touch user-facing strings.
-   - **Not sure yet:** Skip i18n configuration entirely. Do not set `i18n.enabled`. The project can enable i18n later by running `harness-i18n-workflow` directly.
+   - **Not sure:** Skip i18n configuration entirely. Do not set `i18n.enabled`. The project can enable i18n later by running `harness-i18n-workflow` directly.
 
 5b. **Configure design system (non-test-suite projects).** Ask: "Will this project have a UI requiring a design system?" Mirror the i18n step's three-way response shape. Use `emit_interaction`:
 
@@ -136,7 +138,7 @@ Detect plugin-only state by checking whether `harness` resolves on PATH (`comman
     Based on the answer:
     - **Yes:** Ask a follow-up: "Which platforms? `web`, `mobile`, or both?" Write `design.enabled: true` and `design.platforms: [...]` (a non-empty array of `web` and/or `mobile`) to `harness.config.json`. Inform the user: "Design tokens will be generated when you start your first design-touching feature — `harness-design-system` fires automatically via `on_new_feature`."
     - **No:** Write `design.enabled: false` to `harness.config.json`. Do not write `design.platforms`. The `on_new_feature` trigger respects this flag and will not fire `harness-design-system`.
-    - **Not sure yet:** Do not write `design.enabled` or `design.platforms`. The project can enable design later by running `harness-design-system` directly; `on_new_feature` will prompt gently when a feature touches user-facing UI.
+    - **Not sure:** Do not write `design.enabled` or `design.platforms`. The project can enable design later by running `harness-design-system` directly; `on_new_feature` will prompt gently when a feature touches user-facing UI.
 
     **Skip this step entirely if Phase 1 step 5 classified the project as a test suite.** Test-suite projects will be dispatched at step 6 below to `initialize-test-suite-project` and have no UI to govern.
 
@@ -183,7 +185,7 @@ Detect plugin-only state by checking whether `harness` resolves on PATH (`comman
     - **Absent (most common on init).** Present the prompt above. Apply the answer:
       - **Yes:** delegate to `harness-strategy` (which routes via its own Phase 0 to the first-run interview). When `harness-strategy` completes, continue with step 6.
       - **No:** write `init.strategy.declined: true` to `.harness/state.json` (merge into existing JSON; do not clobber). This is the explicit decline location; re-running init detects the flag and skips the prompt.
-      - **Not sure yet:** no state write. `/harness:strategy` remains available standalone, and a future re-run of init will re-offer.
+      - **Not sure:** no state write. `/harness:strategy` remains available standalone, and a future re-run of init will re-offer.
 
     - **Present and valid.** Skip the prompt silently. Surface a one-line note: `STRATEGY.md detected — downstream skills will pick it up as grounding`. No `init.strategy.declined` write.
 
@@ -303,7 +305,7 @@ This phase closes the parity gap that the marketplace plugin install does not co
 
    Based on the answer:
    - **Yes:** Invoke `harness-roadmap` (skill) or run `/harness:roadmap --create` to create `docs/roadmap.md`. Verify the file exists. The `manage_roadmap` MCP tool is for managing entries in an existing roadmap, not for creating one.
-     - **If `design.enabled === true` in `harness.config.json`** (set by Phase 3 step 5b), call `manage_roadmap` with `action: add`, `feature: "Set up design system"`, `status: "planned"`, `milestone: "Current Work"`, `summary: "Run harness-design-system to define palette, typography, and generate W3C DTCG tokens. Deferred from project init — fires on first design-touching feature via on_new_feature."`. Skip silently if `manage_roadmap show` reports a duplicate `(feature, milestone)` pair. This closes the loop between deferred design intent and visible planning work.
+     - **If `design.enabled === true` in `harness.config.json`** (set by Phase 3 step 5b), call `manage_roadmap` with `action: add`, `feature: "Set up design system"`, `status: "planned"`, `milestone: "Intake"`, `summary: "Run harness-design-system to define palette, typography, and generate W3C DTCG tokens. Deferred from project init — fires on first design-touching feature via on_new_feature."`. Skip silently if `manage_roadmap show` reports a duplicate `(feature, milestone)` pair. This closes the loop between deferred design intent and visible planning work.
    - **No:** Skip silently. The user can still run `/harness:roadmap --create` later — that informational fallback remains valid.
 
 2. **Commit the initialization.** All generated, configured, and instrumentation files in a single commit. Include `harness.config.json`, `AGENTS.md`, `.harness/arch/baselines.json`, `.harness/telemetry.json` (if created), updates to project `.mcp.json` (if Tier-0 integrations were wired), and any roadmap files.
@@ -323,11 +325,11 @@ This phase closes the parity gap that the marketplace plugin install does not co
 - **`harness integrations list` / `harness integrations add <name>`** — Phase 5 step 6. Lists Tier-0 (zero-config) and Tier-1 (API-key) MCP integrations and adds them to project `.mcp.json`.
 - **`harness-i18n-workflow configure` + `harness-i18n-workflow scaffold`** — Invoked during Phase 3 if the project will support multiple languages. Sets up i18n configuration and translation file structure.
 - **`harness-design-system` (deferred via `on_new_feature`)** — Phase 3 step 5b records `design.enabled` + `design.platforms` in `harness.config.json` but does NOT run the full design-system skill. Token generation defers to the first design-touching feature, where `harness-design-system` fires via `on_new_feature` and reads `design.enabled` to decide whether to proceed.
-- **`harness-strategy`** — Phase 3 step 5c delegates to this skill on "Yes". The skill conducts a first-run interview (pushback rules with a 2-round cap per section) and writes a valid `STRATEGY.md` at repo root via `writeStrategyDoc`. On "No" init writes `init.strategy.declined: true` to `.harness/state.json`. On "Not sure yet" no state is written; the user can run `/harness:strategy` standalone. When `STRATEGY.md` already exists and is valid the prompt is skipped; when present-but-invalid the user gets the three-path repair offer.
+- **`harness-strategy`** — Phase 3 step 5c delegates to this skill on "Yes". The skill conducts a first-run interview (pushback rules with a 2-round cap per section) and writes a valid `STRATEGY.md` at repo root via `writeStrategyDoc`. On "No" init writes `init.strategy.declined: true` to `.harness/state.json`. On "not sure" no state is written; the user can run `/harness:strategy` standalone. When `STRATEGY.md` already exists and is valid the prompt is skipped; when present-but-invalid the user gets the three-path repair offer.
 - **`validateStrategy`** — `@harness-engineering/core` helper used by Phase 3 step 5c to detect present-but-invalid `STRATEGY.md`. Invoked through a Node one-liner so init does not require importing core directly.
 - **`initialize-test-suite-project`** — Sub-skill. Invoked during Phase 3 step 6 when Phase 1 step 5 classified the project as a test suite. Owns archetype selection, shared-library vs in-repo decision, layer variants, tag taxonomy, reporter stack, custom report, and "prove the guards fire" verification.
 - **`harness-roadmap` skill** — Phase 6 step 1 invokes this skill (or `/harness:roadmap --create`) when the user opts in to creating `docs/roadmap.md`. The `manage_roadmap` MCP tool does not create roadmaps; it manages entries in an existing one.
-- **`manage_roadmap` MCP tool** — Phase 6 step 1, when `design.enabled === true`, calls `manage_roadmap` with `action: add` to insert a `planned` "Set up design system" item under milestone `Current Work` with a summary describing the deferred work.
+- **`manage_roadmap` MCP tool** — Phase 6 step 1, when `design.enabled === true`, calls `manage_roadmap` with `action: add` to insert a `planned` "Set up design system" item under milestone `Intake` with a summary describing the deferred work.
 
 ## Success Criteria
 
@@ -340,29 +342,29 @@ This phase closes the parity gap that the marketplace plugin install does not co
 - The adoption level matches what was agreed upon with the human
 - All generated files are committed in a single atomic commit
 - i18n configuration is set if the human chose to enable it during init
-- For non-test-suite projects, the design-system question was asked and `harness.config.json` reflects the answer: `design.enabled: true` (with `design.platforms` populated) for yes, `design.enabled: false` for no, or absent for not-sure.
-- The strategy question (Phase 3 step 5c) was asked unless `STRATEGY.md` already existed and was valid (in which case the prompt was skipped silently with a one-line detection note). The answer was recorded according to the documented semantics: Yes → `STRATEGY.md` exists and `harness validate` passes against `StrategyDocSchema`; No → `.harness/state.json` contains `init.strategy.declined: true`; Not sure yet → no `STRATEGY.md` and no `init.strategy.declined` flag.
+- For non-test-suite projects, the design-system question was asked and `harness.config.json` reflects the answer: `design.enabled: true` (with `design.platforms` populated) for yes, `design.enabled: false` for no, or absent for not sure.
+- The strategy question (Phase 3 step 5c) was asked unless `STRATEGY.md` already existed and was valid (in which case the prompt was skipped silently with a one-line detection note). The answer was recorded according to the documented semantics: Yes → `STRATEGY.md` exists and `harness validate` passes against `StrategyDocSchema`; No → `.harness/state.json` contains `init.strategy.declined: true`; Not sure → no `STRATEGY.md` and no `init.strategy.declined` flag.
 - **Phase 5 (INSTRUMENT) outputs:** `.harness/graph/` is populated; `.harness/arch/baselines.json` exists; `harness check-perf` ran without errors (intermediate and above); `.harness/telemetry.json` exists if the human opted into identity tagging; legacy layout warnings were surfaced via `harness migrate --dry-run` and either resolved or explicitly deferred; Tier-0 MCP integrations (context7, sequential-thinking, playwright) are present in the project's `.mcp.json` (or the human declined and that decision is recorded).
 - The roadmap question was asked. If the user answered yes, `docs/roadmap.md` exists and was created via `harness-roadmap` (or the documented `/harness:roadmap --create` fallback).
-- When `design.enabled === true` AND the user answered yes to the roadmap question, `docs/roadmap.md` contains a `planned` entry titled "Set up design system" under milestone `Current Work` with a summary describing the deferred work. The entry is absent in all other answer combinations.
+- When `design.enabled === true` AND the user answered yes to the roadmap question, `docs/roadmap.md` contains a `planned` entry titled "Set up design system" under milestone `Intake` with a summary describing the deferred work. The entry is absent in all other answer combinations.
 - For test suites: `initialize-test-suite-project` ran to completion and its Success Criteria are also met
 - For plugin-only invocations: every CLI invocation in this skill ran successfully via `npx @harness-engineering/cli <cmd>` without requiring a global install. If any invocation failed because the npx download failed, the human was prompted to retry or `npm install -g @harness-engineering/cli` instead.
 
 ## Rationalizations to Reject
 
-| Rationalization                                                              | Why It Is Wrong                                                                                                                                                                                                                                                                                                                                                       |
-| ---------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| "The generated AGENTS.md template looks fine -- no need to customize it"     | Phase 3 says do not blindly accept generated content. Without project-specific descriptions, agents receive generic instructions.                                                                                                                                                                                                                                     |
-| "We should start at the advanced level since we want full coverage"          | The skill recommends basic for new projects. Each level builds on the previous. Jumping to advanced creates misconfigured rules.                                                                                                                                                                                                                                      |
-| "I will skip the i18n question to keep setup fast"                           | Phase 3 requires asking about i18n and recording the decision. Skipping creates ambiguity about whether the omission was intentional.                                                                                                                                                                                                                                 |
-| "I will skip the design-system question to keep setup fast"                  | Phase 3 step 5b requires asking about design and recording the answer in `design.enabled`. Skipping creates ambiguity about whether the omission was intentional and bypasses the linkage between init and the deferred `harness-design-system` invocation on `on_new_feature`.                                                                                       |
-| "I will skip the strategy question — it's just paperwork"                    | Phase 3 step 5c is the only point in the workflow where init asks the user to capture the strategic anchor. Skipping bypasses the grounding signal that brainstorming, ideate, and roadmap-pilot read; downstream skill output degrades silently. Even a "No" or "Not sure yet" answer is better than no answer because it locks in the decision (or absence of one). |
-| "STRATEGY.md exists already, so I should re-run the interview to refresh it" | When `STRATEGY.md` is present and valid, Phase 3 step 5c skips the prompt silently. Refreshing strategy mid-init is out of scope — that is what the standalone `/harness:strategy` update flow is for. Surface a one-line detection note and continue.                                                                                                                |
-| "STRATEGY.md is present but invalid, so I should block init"                 | Phase 3 step 5c explicitly does NOT block. It surfaces the validation error, offers three repair paths (fix now / move-to-bak / ignore), and continues based on the user's choice. Init is the wrong place to gate on strategy correctness.                                                                                                                           |
-| "Validation passed, so the project is ready"                                 | Phase 5 captures baselines, configures telemetry identity, surfaces legacy warnings, and wires Tier-0 integrations. Validation alone is not sufficient.                                                                                                                                                                                                               |
-| "Plugin install means setup is done"                                         | The marketplace plugin ships skills, slash commands, subagents, hooks, and MCP — but it cannot mutate the user's project state. `harness.config.json`, baselines, telemetry identity, and Tier-0 integrations require running this skill once per project.                                                                                                            |
-| "Skip Phase 5 if we already ran `harness setup`"                             | Phase 5 is idempotent. It safely no-ops where setup already wired things and fills gaps where it didn't (common case: setup ran once, then a new Tier-0 integration was added or a layout was migrated). Re-running is the right behavior.                                                                                                                            |
-| "This is a test suite, we'll configure layers in this skill"                 | Phase 3 step 6 dispatches to `initialize-test-suite-project` for archetype selection, layer variants, and the rest. Do not inline test-suite-specific configuration here — the sub-skill owns it and carries the gotchas.                                                                                                                                             |
+| Rationalization                                                              | Why It Is Wrong                                                                                                                                                                                                                                                                                                                                                   |
+| ---------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "The generated AGENTS.md template looks fine -- no need to customize it"     | Phase 3 says do not blindly accept generated content. Without project-specific descriptions, agents receive generic instructions.                                                                                                                                                                                                                                 |
+| "We should start at the advanced level since we want full coverage"          | The skill recommends basic for new projects. Each level builds on the previous. Jumping to advanced creates misconfigured rules.                                                                                                                                                                                                                                  |
+| "I will skip the i18n question to keep setup fast"                           | Phase 3 requires asking about i18n and recording the decision. Skipping creates ambiguity about whether the omission was intentional.                                                                                                                                                                                                                             |
+| "I will skip the design-system question to keep setup fast"                  | Phase 3 step 5b requires asking about design and recording the answer in `design.enabled`. Skipping creates ambiguity about whether the omission was intentional and bypasses the linkage between init and the deferred `harness-design-system` invocation on `on_new_feature`.                                                                                   |
+| "I will skip the strategy question — it's just paperwork"                    | Phase 3 step 5c is the only point in the workflow where init asks the user to capture the strategic anchor. Skipping bypasses the grounding signal that brainstorming, ideate, and roadmap-pilot read; downstream skill output degrades silently. Even a "no" or "not sure" answer is better than no answer because it locks in the decision (or absence of one). |
+| "STRATEGY.md exists already, so I should re-run the interview to refresh it" | When `STRATEGY.md` is present and valid, Phase 3 step 5c skips the prompt silently. Refreshing strategy mid-init is out of scope — that is what the standalone `/harness:strategy` update flow is for. Surface a one-line detection note and continue.                                                                                                            |
+| "STRATEGY.md is present but invalid, so I should block init"                 | Phase 3 step 5c explicitly does NOT block. It surfaces the validation error, offers three repair paths (fix now / move-to-bak / ignore), and continues based on the user's choice. Init is the wrong place to gate on strategy correctness.                                                                                                                       |
+| "Validation passed, so the project is ready"                                 | Phase 5 captures baselines, configures telemetry identity, surfaces legacy warnings, and wires Tier-0 integrations. Validation alone is not sufficient.                                                                                                                                                                                                           |
+| "Plugin install means setup is done"                                         | The marketplace plugin ships skills, slash commands, subagents, hooks, and MCP — but it cannot mutate the user's project state. `harness.config.json`, baselines, telemetry identity, and Tier-0 integrations require running this skill once per project.                                                                                                        |
+| "Skip Phase 5 if we already ran `harness setup`"                             | Phase 5 is idempotent. It safely no-ops where setup already wired things and fills gaps where it didn't (common case: setup ran once, then a new Tier-0 integration was added or a layout was migrated). Re-running is the right behavior.                                                                                                                        |
+| "This is a test suite, we'll configure layers in this skill"                 | Phase 3 step 6 dispatches to `initialize-test-suite-project` for archetype selection, layer variants, and the rest. Do not inline test-suite-specific configuration here — the sub-skill owns it and carries the gotchas.                                                                                                                                         |
 
 ## Examples
 
@@ -458,7 +460,7 @@ Step 4 (roadmap): "Set up a project roadmap now?"
   design.enabled === true detected → manage_roadmap action: add
     feature: "Set up design system"
     status: planned
-    milestone: Current Work
+    milestone: Intake
     summary: "Run harness-design-system to define palette, typography, and generate W3C DTCG tokens..."
   Result: docs/roadmap.md contains the planned design item.
 
@@ -470,7 +472,7 @@ git add harness.config.json .harness/ AGENTS.md docs/roadmap.md
 git commit -m "feat: initialize harness project with design and roadmap"
 ```
 
-**Final state:** `harness.config.json` has `design.enabled: true` + `design.platforms: ["web"]`; `docs/roadmap.md` lists "Set up design system" as a `planned` item under `Current Work`; on the first feature touching UI, `on_new_feature` fires `harness-design-system` which reads `design.enabled` and runs the full discover/define/generate/validate flow.
+**Final state:** `harness.config.json` has `design.enabled: true` + `design.platforms: ["web"]`; `docs/roadmap.md` lists "Set up design system" as a `planned` item under `Intake`; on the first feature touching UI, `on_new_feature` fires `harness-design-system` which reads `design.enabled` and runs the full discover/define/generate/validate flow.
 
 ### Example: Migrating Existing Project from Basic to Intermediate
 
@@ -547,7 +549,7 @@ npx @harness-engineering/cli init  # auto-detects framework
 Customize AGENTS.md with project-specific content.
 Phase 3 step 5 (i18n): "No, English only." → i18n.enabled = false.
 Phase 3 step 5b (design): "No, this is a backend service." → design.enabled = false.
-Phase 3 step 5c (strategy): "Not sure yet." → no STRATEGY.md, no init.strategy.declined flag; user can run /harness:strategy later.
+Phase 3 step 5c (strategy): "not sure." → no STRATEGY.md, no init.strategy.declined flag; user can run /harness:strategy later.
 Phase 3 step 6 (test-suite dispatch): not a test suite, skipped.
 ```
 

package/dist/agents/skills/claude-code/initialize-harness-project/skill.yaml

@@ -35,3 +35,4 @@ state:
 depends_on:
   - initialize-test-suite-project
   - harness-design-system
+  - harness-roadmap

package/dist/agents/skills/claude-code/outcome-eval/SKILL.md

@@ -0,0 +1,97 @@
+# Outcome Eval
+
+> Post-execution LLM-judgment: did the implementation actually satisfy its spec? Reads the spec's acceptance section, the change diff, and test output, and emits a confidence-rated `OutcomeVerdict` (`SATISFIED | NOT_SATISFIED | INCONCLUSIVE`) with a rationale and unmet criteria. Ship authority is derived in TypeScript, never trusted from the LLM: a high-confidence `NOT_SATISFIED` blocks ship; every other verdict is advisory. The harness's first blocking post-execution spec-satisfaction gate (the roadmap's named #1 gap). Each verdict persists as an `execution_outcome` node, compounding into skill-effectiveness baselines.
+
+## When to Use
+
+- At orchestrator step 6.5 — after Code Review, before Ship — on every change with a spec.
+- When you need a durable, structured answer to "did this code do what the spec said?"
+- NOT for pre-execution risk simulation (use PESL).
+- NOT for rule-based floors (lint/architecture/entropy) or craft ceilings (naming/spec/security) — those run elsewhere.
+- NOT for auto-remediation. outcome-eval judges; it does not fix.
+- NOT when no judgable spec section exists — the verdict degrades to INCONCLUSIVE/advisory and never blocks.
+
+## Process
+
+### Phase 1: GATHER — Collect inputs
+
+1. Capture the change under judgment as a unified diff: `git diff` (or `git diff <base>...HEAD` for a branch). Record it as `diff`.
+2. Capture test-runner output. If a test command is known, run it and capture stdout+stderr as `testOutput`; otherwise pass the most recent captured output. Empty/unparseable test output is tolerated (degrades to advisory).
+3. Resolve the spec path. Prefer the spec under `docs/changes/<feature>/proposal.md` for the current change. Record as `specPath`.
+
+### Phase 2: RESOLVE — Find the judgment section
+
+The evaluator resolves the section internally via the fallback chain `## Success Criteria` -> `## User-Visible Behavior` -> `## Overview`, recording the match in `judgedAgainst`. No manual action — pass `specPath` and let `OutcomeEvaluator` resolve. If no section is judgable, the verdict is INCONCLUSIVE/advisory.
+
+### Phase 3: JUDGE — Invoke the evaluator
+
+1. Invoke the MCP tool `mcp__harness__outcome_eval` with `{ specPath, diff, testOutput }` (optional `model`). The tool constructs `OutcomeEvaluator` cli-side and calls `evaluate({ specPath, diff, testOutput })`; the supported v1 provider is the anthropic analysis provider (`ANTHROPIC_API_KEY`).
+2. **`diff` and `testOutput` are required inputs and the agent MUST supply them** from the session (`git diff` + captured test-runner output). They are the evidence the judge reasons over — passing an empty `diff` or empty `testOutput` is the degradation path, not the normal path: the verdict degrades to INCONCLUSIVE/advisory (never blocking), which defeats the gate. Do not invoke the tool without real diff/test content.
+3. The LLM returns ONLY `verdict / confidence / rationale / unmetCriteria`. `authority` is computed in TypeScript from `(verdict, confidence)` and is never read from the LLM — do not attempt to override it. The tool returns the verdict exactly as the evaluator derives it.
+4. The call is degrade-safe: provider failure (incl. no `ANTHROPIC_API_KEY`), empty diff, empty test output, or missing judgable section yields INCONCLUSIVE/low/advisory. It never throws and never blocks.
+
+### Phase 4: GATE — Render and (conditionally) halt
+
+1. Render the verdict: `verdict`, `confidence`, `judgedAgainst`, `rationale`, and `unmetCriteria`.
+2. Authority rule (must match `deriveAuthority`): authority is `blocking` **iff** `verdict === 'NOT_SATISFIED' && confidence === 'high'`; every other combination — including all `INCONCLUSIVE` and `SATISFIED` cases, and all `medium`/`low` `NOT_SATISFIED` — is `advisory`.
+3. **On a blocking verdict: HALT before the Ship step.** Report the unmet criteria and stop; do not proceed to step 7. Resolution requires fixing the implementation (or the spec) and re-running outcome-eval.
+4. On an advisory verdict: report it and proceed. Advisory `NOT_SATISFIED` is surfaced for human attention but does not stop the workflow.
+
+## Harness Integration
+
+- **`mcp__harness__outcome_eval`** — MCP tool (the invocation surface). Inputs: `specPath` (required), `diff` (required), `testOutput` (required), `model` (optional), `path` (optional project root for graph persistence). The agent supplies `diff` and `testOutput` from the session; omitting them degrades the verdict to INCONCLUSIVE/advisory (never blocking). The handler builds the cli `AnalysisProvider` + a `GraphStore`, constructs `OutcomeEvaluator`, and returns the `OutcomeVerdict` with authority exactly as derived in TypeScript.
+- **Evaluator surface:** `OutcomeEvaluator`, `deriveAuthority`, `verdictSchema`, `OutcomeVerdict` are exported from `@harness-engineering/intelligence`.
+- **Provider path (v1 supported):** the anthropic analysis provider (`ANTHROPIC_API_KEY`). When no provider is configured the call degrades to INCONCLUSIVE/advisory. The openai-compatible _strict_ structured-output path is a known follow-up (see Known Limitations).
+- **Orchestrator:** runs as step 6.5 between Code Review and Ship in `harness.orchestrator.md`.
+- **Persistence:** each `evaluate()` writes one `execution_outcome` node via `ExecutionOutcomeConnector`, consumable by `effectiveness/scorer.ts`.
+
+## Known Limitations
+
+- **INCONCLUSIVE persistence:** the persisted node maps `INCONCLUSIVE -> result: 'failure'` for type-validity, but it OMITS `agentPersona` and writes `affectedSystemNodeIds: []`. The effectiveness scorer (`gatherOutcomes`) ignores any node missing `agentPersona` or `outcome_of` edges, so outcome-eval nodes are **scorer-non-counting** in v1 — the INCONCLUSIVE-as-failure mapping is therefore harmless and does not punish any persona. If a future change attaches persona/affected-system attribution, it MUST first change INCONCLUSIVE modeling (do not persist INCONCLUSIVE, or use a distinct result value the scorer excludes) before the node becomes scorer-counted.
+- **openai-compatible strict mode:** `zodToJsonSchema` does not emit `additionalProperties: false`, which OpenAI strict structured output requires. The v1 supported path is claude-cli / anthropic. Follow-up tracked.
+- **CI required-check wiring:** deferred to roadmap #540 (unbuilt CI workflow template).
+
+## Success Criteria
+
+See `docs/changes/outcome-eval/proposal.md` for the full 9 criteria. This skill satisfies SC8 (orchestrator step 6.5 + blocking halt) and SC9 (introduces no new `harness validate` findings; layer rules respected).
+
+## Examples
+
+### Example: NOT_SATISFIED with high confidence (blocks)
+
+**Input:** spec Success Criteria require `GET /api/users/:id` to return 404 with `{ error: 'User not found' }`; the diff implements the happy path only, no 404 branch; test output shows the 404 test failing.
+
+**Verdict:**
+
+```
+verdict:        NOT_SATISFIED
+confidence:     high
+judgedAgainst:  success-criteria
+authority:      blocking
+unmetCriteria:
+  - "404 path for nonexistent user is unimplemented; the failing test asserts { error: 'User not found' }."
+rationale:      "The diff adds the lookup but returns 200 with an empty body when the user is missing."
+```
+
+**Action:** HALT before Ship. Report unmet criteria; do not open the PR.
+
+### Example: partial implementation (advisory)
+
+**Input:** the diff meets most criteria; one acceptance item is ambiguous in the diff.
+
+**Verdict:** `NOT_SATISFIED confidence: medium authority: advisory` — surfaced for review, workflow proceeds.
+
+## Gates
+
+- **Authority is never read from the LLM.** The verdict's `authority` is always `deriveAuthority(verdict, confidence)` computed in TypeScript. If you find yourself letting the model assert blocking/advisory, STOP — that defeats the entire purpose of this gate.
+- **Block only on high-confidence NOT_SATISFIED.** `authority === 'blocking'` iff `verdict === 'NOT_SATISFIED' && confidence === 'high'`. Every other combination — all `SATISFIED`, all `INCONCLUSIVE`, and every `medium`/`low` `NOT_SATISFIED` — is advisory. Do not halt the workflow on an advisory verdict.
+- **Always supply `diff` and `testOutput`.** Omitting them degrades the verdict to INCONCLUSIVE/advisory (a silent false-negative at the ship gate). Gather them from the session before invoking the tool.
+- **Never block on infrastructure noise.** A provider failure, an unparseable response, or a missing spec must resolve to INCONCLUSIVE/advisory, never a thrown error or a block. The evaluator enforces this; do not reintroduce a hard failure in the wrapper.
+- **Do not skip step 6.5.** The gate runs after Code Review and before Ship. A blocking verdict halts before step 7; it is not optional.
+
+## Escalation
+
+- **Blocking verdict the author disputes:** the resolution is to fix the implementation (or amend the spec's Success Criteria if they were wrong) and re-run outcome-eval — not to override the gate. If the spec itself is wrong, that is a spec change, reviewed on its own merits.
+- **Repeated INCONCLUSIVE on a real change:** usually means `diff`/`testOutput` were not supplied, or no judgable section exists in the spec. Confirm inputs and that the spec has a Success Criteria / User-Visible Behavior / Overview section.
+- **No `ANTHROPIC_API_KEY` configured:** every verdict degrades to INCONCLUSIVE/advisory and nothing blocks. Surface this to the human — the gate is effectively disabled until a provider is configured.
+- **Verdict seems wrong (false positive/negative):** capture the spec section, diff, and verdict, and route to the maintainers; do not loosen the conservative-confidence prompt ad hoc.

package/dist/agents/skills/claude-code/outcome-eval/skill.yaml

@@ -0,0 +1,51 @@
+name: outcome-eval
+version: '0.1.0'
+description: >-
+  LLM-judgment skill that produces a structured, confidence-rated verdict on
+  whether an implementation satisfied its spec. Reads the spec's acceptance
+  section, the change diff, and test output; emits an OutcomeVerdict
+  (SATISFIED | NOT_SATISFIED | INCONCLUSIVE) with confidence, rationale, and
+  unmet criteria. Authority is derived in TypeScript, never from the LLM: a
+  high-confidence NOT_SATISFIED blocks ship; every other verdict is advisory.
+  The verdict persists as an execution_outcome node and feeds skill-effectiveness
+  baselines. The harness's first blocking post-execution spec-satisfaction gate.
+stability: draft
+cognitive_mode: constructive-architect
+triggers:
+  - manual
+  - on_pr
+platforms:
+  - claude-code
+tools:
+  - Bash
+  - Read
+  - Glob
+  - Grep
+mcp:
+  tool: outcome_eval
+  # specPath, diff, and testOutput are ALL required. diff/testOutput are the
+  # evidence the judge reasons over: omitting them degrades the verdict to
+  # INCONCLUSIVE/advisory (never blocking), so the agent MUST supply them.
+  input:
+    specPath: 'string (required) — absolute or repo-relative path to the spec markdown to judge against'
+    diff: 'string (required) — unified diff of the change under judgment (git diff); omission degrades the verdict to INCONCLUSIVE/advisory'
+    testOutput: 'string (required) — captured test-runner stdout+stderr; omission degrades the verdict to INCONCLUSIVE/advisory'
+    model: 'string (optional) — model override for the outcome-eval LLM call'
+    path: 'string (optional) — project root used to resolve the knowledge graph (default cwd)'
+type: rigid
+tier: 2
+phases:
+  - name: gather
+    description: Collect the unified diff (git diff), capture test output, resolve the spec path
+    required: true
+  - name: resolve
+    description: Resolve the judgment section (Success Criteria -> User-Visible Behavior -> Overview)
+    required: true
+  - name: judge
+    description: Invoke OutcomeEvaluator.evaluate(); LLM returns verdict/confidence/rationale/unmetCriteria; authority derived in TS
+    required: true
+  - name: gate
+    description: Render the verdict; on a blocking verdict (NOT_SATISFIED + high confidence) halt before Ship
+    required: true
+state:
+  persistent: false

package/dist/agents/skills/codex/harness-audit-harness-strength/SKILL.md

@@ -0,0 +1,188 @@
+# Harness Audit: Harness Strength
+
+> Mechanically audit whether a project's harness is load-bearing. Orchestrates the deterministic check-harness-strength engine; interprets results. Not a deep/AI review.
+
+## When to Use
+
+- On a milestone gate, to confirm the harness still constrains rather than merely decorates the project
+- When adding the strength audit as a required CI check, so the seven STRENGTH patterns block the merge instead of being prose advice
+- When validating a harness-distribution (toolkit) repo before shipping templates downstream — to catch a default that would weaken every adopter
+- When a reviewer suspects a gate "passes" without doing anything (a snapshot that asserts `passed: true` for a check it never ran)
+- NOT for deep security review — use `harness-security-review` (semantic, AI-assisted)
+- NOT for design-system drift — use `detect-design-drift`
+- NOT for reimplementing pattern detection by hand — the engine (`HarnessStrengthAuditor`) owns detection. This skill runs it and interprets the output.
+
+## Process
+
+This skill ORCHESTRATES `harness check-harness-strength`. It never re-greps configs, never re-parses hooks, and never re-derives findings. The engine reads the project once and returns a structured `AuditResult`; the skill's job is to run it with the right mode/severity and turn the JSON into an actionable report.
+
+### Phase 1: SCAN — Run the engine
+
+1. **Resolve the project root.** Use the provided `path` argument, else the current working directory.
+
+2. **Resolve the mode.**
+   - If `--toolkit` or `--adopter` (or `--mode toolkit|adopter`) is supplied, honor it. Explicit mode wins.
+   - Otherwise auto-detect: treat the repo as **toolkit** when it is a harness distribution — both `templates/` and `agents/skills/` exist at the root. Otherwise **adopter**.
+   - Do not inspect config files to "double-check" the mode. The engine resolves and reports the mode it used.
+
+3. **Invoke the command** via Bash, requesting JSON:
+
+   ```bash
+   harness check-harness-strength [--mode <adopter|toolkit>] [--severity <error|warning|info>] [--report-only] --json
+   ```
+
+   The `--json` payload is the raw structured `AuditResult` (score, tier, mode, findings with `id`, `gearPiece`, `severity`, `message`, `remediation`, and file:line evidence). Parse this — do not re-derive it.
+
+4. **Do NOT inspect config files by hand.** The engine reads `harness.config.json`, hook profiles, baselines, snapshots, and tier defaults exactly once. Hand-grepping `.husky/pre-commit` or `harness.config.json` to "confirm" a finding reimplements detection and violates the core decision behind this skill (ADR 0039 / spec D1).
+
+### Phase 2: DETECT — Interpret findings
+
+1. **Map each finding to its STRENGTH pattern** using the JSON `id` and `gearPiece` fields. Every finding the engine emits corresponds to exactly one of the seven patterns below. Do not invent findings the engine did not emit, and do not silence findings it did.
+
+2. **The seven STRENGTH patterns** (the engine's rule registry — for interpretation only; the engine, not the skill, decides which fire):
+
+   | ID           | Gear piece              | Pattern                                                              | Default |
+   | ------------ | ----------------------- | -------------------------------------------------------------------- | ------- |
+   | STRENGTH-001 | blocking-gate           | Hook documented "never blocks"/"always exits 0" in an active profile | error   |
+   | STRENGTH-002 | regression-baseline     | Pre-commit auto-updates baselines/thresholds on regression           | error   |
+   | STRENGTH-003 | skip-discipline         | `--skip` list > 2 categories without inline justification            | warning |
+   | STRENGTH-004 | architecture-thresholds | `layers` defined but `architecture.thresholds` empty/absent          | error   |
+   | STRENGTH-005 | tier-default            | Init/config defaults to lowest tier (`basic`)                        | warning |
+   | STRENGTH-006 | review-gate             | Baseline-update PR auto-approved without independent review          | error   |
+   | STRENGTH-007 | snapshot-honesty        | `passed:true` in health snapshot whose `signals[]` names that check  | error   |
+
+3. **Attach evidence.** Each finding carries a `file:line` (or config-key) reference and a remediation string from the engine. Surface them verbatim — do not paraphrase the remediation.
+
+### Phase 3: SCORE/REPORT
+
+1. **Surface the score and tier.** The engine returns a 0-100 strength score and a tier label:
+   - `solid` — score >= 85
+   - `at-risk` — score 50-84
+   - `theatre` — score < 50
+
+2. **Report format** (mirror of the security-scan report block):
+
+   ```
+   Harness Strength: [PASS/FAIL] — tier: <solid|at-risk|theatre> (score N/100, mode <adopter|toolkit>)
+   Findings: <count>  (Errors: N | Warnings: N | Info: N)
+
+   [STRENGTH-00N] <gearPiece> <file:line> (severity)
+     <message>
+     Remediation: <engine remediation string>
+   ```
+
+3. **PASS/FAIL.** FAIL if any error-severity finding survives the severity threshold, unless `--report-only` was passed (then exit softens to PASS but findings are still listed). Otherwise PASS.
+
+## Gates
+
+- **Error-severity findings are blocking.** The report is FAIL and the exit is non-zero when any error-severity finding survives filtering — unless `--report-only` was explicitly requested. Do not report PASS over an unsuppressed error finding.
+- **No reimplementation.** The skill must run `harness check-harness-strength` and interpret its JSON. It must never hand-grep hooks, hand-parse `harness.config.json`, or re-derive a score. Reimplementing detection violates ADR 0039.
+- **"Not evaluable" is not a pass.** When an input the engine needs is absent (e.g., no hook profile, no snapshot), the engine reports that state. Surface it as-is. Do not convert "could not evaluate" into "passed."
+- **Mechanical only.** No AI judgment about whether a weakness "really matters." The patterns are deterministic; the engine decides what fires.
+
+## Escalation
+
+- **Disputed finding / false positive:** Do not suppress by editing files or skipping the rule in the skill. Adjust severity via config — `audit.harnessStrength.severities` in `harness.config.json` (e.g., downgrade STRENGTH-003 to `info` for a justified wide `--skip`). Document the rationale alongside the override.
+- **Engine misses a known weakness:** This skill cannot add detection. A missing pattern is engine work — file a new `StrengthRule` (`{ id, gearPiece, defaultSeverity, appliesIn(mode), evaluable?(ctx), detect(ctx) }`) in `packages/core/src/harness-strength/`. Out of scope for the skill.
+- **Mode mis-detected:** If auto-detection picks the wrong mode (e.g., a repo with `templates/` that is not a distribution), pass `--mode adopter` (or `--toolkit`) explicitly. Do not work around it by hand-editing detection.
+- **Engine throws or the build is stale:** If `harness check-harness-strength` errors (command not found / not in the built CLI), the dist may be stale. Rebuild the CLI and re-run. Do not substitute a manual config read for the engine.
+
+## Rationalizations to Reject
+
+### Universal
+
+These reasoning patterns sound plausible but lead to bad outcomes. Reject them.
+
+- **"It's probably fine"** — "Probably" is not evidence. Run the engine and cite the result.
+- **"This is best practice"** — Best practice in what context? Cite the source and confirm it applies to this codebase.
+- **"We can fix it later"** — If it is worth flagging, it is worth documenting now with a concrete follow-up plan.
+
+### Domain-Specific
+
+| Rationalization                                            | Reality                                                                                                                                                       |
+| ---------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "The config looks fine when I read it"                     | The engine exists because manual reading misses these patterns — that is the whole point of D1. Reading is not running. Run `harness check-harness-strength`. |
+| "It only warns, not errors, so it's fine"                  | A gate that warns but does not stop IS STRENGTH-001 — this is the recursion item the audit was built to catch. Do not normalize "warns but doesn't block."    |
+| "I'll re-grep the hooks myself to double-check the engine" | Reimplementing detection violates ADR 0039 / spec D1. The engine is the single source of truth — trust and interpret its output, do not shadow it.            |
+| "Score is 70, that's a passing grade"                      | 70 is `at-risk`, not `solid`. Tier labels are thresholds, not letter grades. Below 85 means the harness has load-bearing gaps.                                |
+| "No error findings, so I can skip reading the warnings"    | Warnings (STRENGTH-003, -005) are erosion signals — a `basic` tier default or an unjustified wide `--skip` weakens every future run. Surface them.            |
+
+## Success Criteria
+
+- The engine (`harness check-harness-strength`) ran and produced a score, tier label, and findings (or a clean result).
+- Every finding is interpreted against the seven-pattern table, with the engine's file:line evidence and remediation surfaced verbatim.
+- The exit code / PASS-FAIL reflects the gate (error-severity findings fail unless `--report-only`).
+- No manual config inspection was substituted for the engine — the report is derived entirely from the command's `--json` output.
+- "Not evaluable" states are surfaced as-is, never converted to passes.
+
+## Evidence Requirements
+
+When this skill makes claims about existing code, architecture, or behavior, it MUST cite evidence using one of:
+
+1. **File reference:** `file:line` format (e.g., `.husky/pre-commit:12`) — taken from the engine's finding, not re-derived.
+2. **Code pattern reference:** `file` with description (e.g., `harness.config.json` — "architecture.layers defined, thresholds absent").
+3. **Test/command output:** Inline or referenced output from the `harness check-harness-strength --json` run.
+4. **Session evidence:** Write to the `evidence` session section via `manage_state`.
+
+**Uncited claims:** Technical assertions without citations MUST be prefixed with `[UNVERIFIED]`. Example: `[UNVERIFIED] The pre-commit hook auto-updates baselines`.
+
+## Harness Integration
+
+- **`harness check-harness-strength`** — The CLI command this skill runs and interprets. Options: `--mode <adopter|toolkit>`, `--toolkit`/`--adopter` shortcuts, `--severity <error|warning|info>` (default `warning`), `--report-only`, `--json`.
+- **`HarnessStrengthAuditor`** — Core class from `@harness-engineering/core` that builds a `ProjectContext` once, runs the applicable `StrengthRule`s, and aggregates the score/tier/findings. The skill never instantiates it directly; it goes through the command.
+- **`harness.config.json` → `audit.harnessStrength.severities`** — Per-pattern severity overrides. The supported escalation path for false positives — never suppress by editing source.
+- **`docs/standard/article-failure-patterns.md`** — Narrative companion documenting the seven patterns (forthcoming — downstream item; may not exist yet).
+
+## Examples
+
+### Example: Clean toolkit-mode run
+
+**Input:** A harness-distribution repo (`templates/` and `agents/skills/` present), run at a milestone gate. `harness check-harness-strength --toolkit --json`.
+
+**Output:**
+
+```
+Harness Strength: PASS — tier: solid (score 100/100, mode toolkit)
+Findings: 0  (Errors: 0 | Warnings: 0 | Info: 0)
+```
+
+Every gear piece is load-bearing: no "never blocks" hook, baselines are not auto-updated on regression, architecture thresholds are set, the init default is not `basic`, baseline-update PRs require independent review, and no snapshot claims a check passed that it never ran.
+
+### Example: Findings detected (adopter mode)
+
+**Input:** An adopter repo, run with default severity. `harness check-harness-strength --json`.
+
+**Output:**
+
+```
+Harness Strength: FAIL — tier: at-risk (score 62/100, mode adopter)
+Findings: 2  (Errors: 2 | Warnings: 0 | Info: 0)
+
+[STRENGTH-001] blocking-gate .husky/pre-commit:12 (error)
+  Active pre-commit profile documents "always exits 0" — the gate reports but never blocks.
+  Remediation: Remove the unconditional `exit 0`; let failing checks return a non-zero status.
+
+[STRENGTH-004] architecture-thresholds harness.config.json (error)
+  architecture.layers is defined but architecture.thresholds is empty — layer rules cannot be enforced.
+  Remediation: Set architecture.thresholds (e.g. maxModuleLines, maxDependencyDepth) so the layer
+               definitions become enforceable, or remove the unused layers block.
+```
+
+Two error-severity findings → FAIL. STRENGTH-001 is the recursion item: a hook that runs but cannot block. Run with `--report-only` only when you explicitly want to surface findings without failing the build (e.g., a first baseline-capture run) — and never to normalize the weakness.
+
+## Skill Test Scenarios
+
+### Scenario 1: Red Flag — "I'll re-grep the hooks myself to double-check the engine"
+
+Input: The agent has the `harness check-harness-strength --json` output in hand but is about to open `.husky/pre-commit` and grep it by hand to "confirm" the STRENGTH-001 finding.
+Expected: Agent stops, cites the no-reimplementation gate (ADR 0039 / spec D1), and interprets the engine's JSON instead of hand-grepping the hook.
+
+### Scenario 2: Rationalization — "it only warns, not errors, so it's fine"
+
+Input: The engine reports a documented "never blocks" pre-commit profile, and the agent is tempted to treat it as acceptable because nothing currently fails the build.
+Expected: Agent rejects the rationalization, recognizes that "warns but doesn't stop" IS STRENGTH-001 (the recursion item), and surfaces it as a finding rather than normalizing it.
+
+### Scenario 3: Gate — error-severity STRENGTH-001 finding present without `--report-only`
+
+Input: The `--json` output contains an error-severity STRENGTH-001 finding, `--report-only` was not passed, and the agent is about to summarize the run as PASS.
+Expected: Agent halts, reports FAIL with a non-zero exit, lists the finding with its file:line evidence and remediation, and does not proceed past the gate.

package/dist/agents/skills/codex/harness-audit-harness-strength/skill.yaml

@@ -0,0 +1,54 @@
+name: harness-audit-harness-strength
+version: '0.1.0'
+description: Mechanically audit a project's own harness setup against the seven STRENGTH failure patterns; reports per-pattern findings, a 0-100 strength score, and a tier label (solid/at-risk/theatre). Orchestrates harness check-harness-strength; never reimplements detection.
+stability: draft
+cognitive_mode: constructive-architect
+triggers:
+  - manual
+  - on_milestone
+platforms:
+  - claude-code
+  - gemini-cli
+  - cursor
+  - codex
+tools:
+  - Bash
+  - Read
+  - Grep
+  - Glob
+cli:
+  command: harness skill run harness-audit-harness-strength
+  args:
+    - name: path
+      description: Project root path
+      required: false
+    - name: mode
+      description: 'adopter (default) or toolkit (auto-detected in a harness-distribution repo)'
+      required: false
+    - name: severity
+      description: Minimum severity threshold (error, warning, info)
+      required: false
+    - name: report-only
+      description: Soften exit to 0 even when error-severity findings exist
+      required: false
+mcp:
+  tool: run_skill
+  input:
+    skill: harness-audit-harness-strength
+    path: string
+type: rigid
+tier: 2
+phases:
+  - name: scan
+    description: Resolve mode and run harness check-harness-strength against the target repo
+    required: true
+  - name: detect
+    description: Interpret the 7 STRENGTH pattern findings with file-line evidence
+    required: true
+  - name: score_report
+    description: Surface the 0-100 score, tier label, and per-pattern remediation
+    required: true
+state:
+  persistent: false
+  files: []
+depends_on: []

package/dist/agents/skills/codex/harness-autopilot/SKILL.md

@@ -213,6 +213,8 @@ Persist findings to `{sessionDir}/phase-{N}-review.json`. No blocking → PHASE_
 
 ## Process
 
+**Prompt the human in plain text** — every phase-continue and human-decision interaction in this skill is plain text only. Do not elevate to `AskUserQuestion`: natural headers like "Continue phase" exceed its 12-char cap, rendering the call as ERR.
+
 1. **INIT** — Resolve spec, derive session slug, check for existing state, parse phases.
 2. **ASSESS** — Route by complexity: low/medium auto-plans, high pauses for interactive planning.
 3. **PLAN → APPROVE** — Dispatch harness-planner, check approval signals, auto-approve or pause.