owflow 0.5.0 → 0.7.0

package/dist/skills/orchestrator-framework/references/delegation-rules.md

@@ -1,54 +1,54 @@
-# Delegation Rules
-
-Shared rules for how orchestrators and subskills delegate work. Extracted from `orchestrator-patterns.md` (Section 1) so skills can link to it directly.
-
----
-
-**Always use Skill/Task tools to delegate. Never execute delegated work inline.**
-
-When a phase requires delegation:
-
-1. Use the **Skill tool** for **skills** — loads SKILL.md instructions into the main agent's context; the main agent executes the skill's instructions and continues with the orchestrator workflow afterward
-2. Use the **Task tool** for **subagents/agents** — spawns an isolated subprocess that returns results when complete
-3. Wait for completion before continuing
-
-**Skills and agents are NOT interchangeable.** Skills always use Skill tool; agents always use Task tool. Never invoke a skill via Task tool (`subagent_type`) — it will fail with "Agent type not found."
-
-**Why skills MUST use Skill tool**: Skills like `codebase-analyzer`, `implementation-plan-executor`, and `implementation-verifier` spawn their own subagents (Explore agents, reporters, planners). Subagents cannot spawn other subagents — so these skills must run in the main agent context via Skill tool.
-
-**Companion agent pattern** (e.g., `docs-operator`): Only works for skills that do NOT spawn subagents (like `docs-manager` which only does file operations). A companion agent preloads the skill via the `skills` frontmatter field and is invoked via Task tool. This pattern fails for any skill that needs to spawn subagents.
-
-## Anti-Patterns
-
-| Anti-Pattern                            | Why It's Wrong                              | Correct Approach                                      |
-| --------------------------------------- | ------------------------------------------- | ----------------------------------------------------- |
-| "I'll analyze the codebase..."          | Bypasses codebase-analyzer skill            | Use `Skill` tool with `codebase-analyzer`             |
-| "Let me create the specification..."    | Bypasses specification-creator              | Use `Task` tool with `specification-creator` subagent |
-| "Looking at the gaps between..."        | Bypasses gap-analyzer subagent              | Use `Task` tool with `gap-analyzer`                   |
-| "I'll implement this by..."             | Bypasses implementation-plan-executor skill | Use `Skill` tool with `implementation-plan-executor`  |
-| Reading a SKILL.md then doing the work  | Skill files are instructions FOR skills     | Use Skill tool to invoke                              |
-| Spawning Explore agents in orchestrator | Codebase-analyzer manages its own agents    | Invoke skill, let IT spawn agents                     |
-
-## When Inline Execution is Acceptable
-
-These do NOT require delegation:
-
-1. **Clarifying questions phases** — question is direct
-2. **State updates** — Reading/writing orchestrator-state.yml
-3. **Phase announcements** — Outputting status messages
-4. **Simple decisions** — Enabling/disabling optional phases
-5. **Finalization** — Creating summary, updating metadata
-
-For all analysis, planning, implementation, and verification phases: **ALWAYS DELEGATE**.
-
-**Never acceptable inline** (regardless of perceived task simplicity):
-
-- Specification creation → always delegate to `specification-creator` subagent
-- Implementation planning → always delegate to `implementation-planner` subagent
-- Gap analysis → always delegate to `gap-analyzer` subagent
-- Codebase analysis → always delegate to `codebase-analyzer` skill
-- Code review → always delegate to `code-reviewer` subagent
-- Test execution → always delegate to `test-suite-runner` subagent
-- Implementation completeness → always delegate to `implementation-completeness-checker` subagent
-
-"The task is simple" is NOT a valid reason to skip delegation.
+# Delegation Rules
+
+Shared rules for how orchestrators and subskills delegate work. Extracted from `orchestrator-patterns.md` (Section 1) so skills can link to it directly.
+
+---
+
+**Always use Skill/Task tools to delegate. Never execute delegated work inline.**
+
+When a phase requires delegation:
+
+1. Use the **Skill tool** for **skills** — loads SKILL.md instructions into the main agent's context; the main agent executes the skill's instructions and continues with the orchestrator workflow afterward
+2. Use the **Task tool** for **subagents/agents** — spawns an isolated subprocess that returns results when complete
+3. Wait for completion before continuing
+
+**Skills and agents are NOT interchangeable.** Skills always use Skill tool; agents always use Task tool. Never invoke a skill via Task tool (`subagent_type`) — it will fail with "Agent type not found."
+
+**Why skills MUST use Skill tool**: Skills like `codebase-analyzer`, `implementation-plan-executor`, and `implementation-verifier` spawn their own subagents (Explore agents, reporters, planners). Subagents cannot spawn other subagents — so these skills must run in the main agent context via Skill tool.
+
+**Companion agent pattern** (e.g., `docs-operator`): Only works for skills that do NOT spawn subagents (like `docs-manager` which only does file operations). A companion agent preloads the skill via the `skills` frontmatter field and is invoked via Task tool. This pattern fails for any skill that needs to spawn subagents.
+
+## Anti-Patterns
+
+| Anti-Pattern                            | Why It's Wrong                              | Correct Approach                                      |
+| --------------------------------------- | ------------------------------------------- | ----------------------------------------------------- |
+| "I'll analyze the codebase..."          | Bypasses codebase-analyzer skill            | Use `Skill` tool with `codebase-analyzer`             |
+| "Let me create the specification..."    | Bypasses specification-creator              | Use `Task` tool with `specification-creator` subagent |
+| "Looking at the gaps between..."        | Bypasses gap-analyzer subagent              | Use `Task` tool with `gap-analyzer`                   |
+| "I'll implement this by..."             | Bypasses implementation-plan-executor skill | Use `Skill` tool with `implementation-plan-executor`  |
+| Reading a SKILL.md then doing the work  | Skill files are instructions FOR skills     | Use Skill tool to invoke                              |
+| Spawning Explore agents in orchestrator | Codebase-analyzer manages its own agents    | Invoke skill, let IT spawn agents                     |
+
+## When Inline Execution is Acceptable
+
+These do NOT require delegation:
+
+1. **Clarifying questions phases** — question is direct
+2. **State updates** — Reading/writing orchestrator-state.yml
+3. **Phase announcements** — Outputting status messages
+4. **Simple decisions** — Enabling/disabling optional phases
+5. **Finalization** — Creating summary, updating metadata
+
+For all analysis, planning, implementation, and verification phases: **ALWAYS DELEGATE**.
+
+**Never acceptable inline** (regardless of perceived task simplicity):
+
+- Specification creation → always delegate to `specification-creator` subagent
+- Implementation planning → always delegate to `implementation-planner` subagent
+- Gap analysis → always delegate to `gap-analyzer` subagent
+- Codebase analysis → always delegate to `codebase-analyzer` skill
+- Code review → always delegate to `code-reviewer` subagent
+- Test execution → always delegate to `test-suite-runner` subagent
+- Implementation completeness → always delegate to `implementation-completeness-checker` subagent
+
+"The task is simple" is NOT a valid reason to skip delegation.

package/dist/skills/orchestrator-framework/references/dispatcher-handoff.md

@@ -1,27 +1,27 @@
-# Dispatcher & Handoff Pattern (Subskill Loop)
-
-How orchestrators delegate phase bodies to user-invocable **subskills**. Extracted from `orchestrator-patterns.md` (Section 7) so skills can link to it directly.
-
-Orchestrators MAY delegate phase bodies to user-invocable **subskills** instead of executing them inline. Two modes share one state file and one set of subskills:
-
-| Mode              | Entry point              | Behavior                                                                                                |
-| ----------------- | ------------------------ | ------------------------------------------------------------------------------------------------------- |
-| Orchestrated mode | wrapper skill/command    | Invokes subskills back-to-back via Skill tool in sequence; pauses at `question` gates between subskills |
-| Assisted mode      | dispatcher skill/command | Derives next phase from state, prints the suggested command, and **STOPS**; user invokes each subskill  |
-
-## Rules
-
-1. **State file is canonical.** `orchestrator-state.yml` remains the single source of truth. Subskills read it on entry and write step results on exit. `completed_phases` values MUST stay stable so both modes intermix freely on the same task.
-   - **Value convention**: `completed_phases` / `failed_phases` / `auto_fix_attempts` keys are **descriptive step slugs**, not `phase-N` numbers. Canonical slugs per workflow:
-
-   | Workflow     | Slugs                                                                                                                                                                      | Routing table                    | Template                                       |
-   | ------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------- | ---------------------------------------------- |
-   | development  | `codebase-analysed`, `gap-analysed`, `tdd-red-proven`, `spec-written`, `spec-audited`, `plan-created`, `implementation-done`, `tdd-green-proven`, `options-chosen`, `verification-done`, `e2e-run`, `docs-generated`, `task-completed` | `skills/development/SKILL.md`  | `templates/orchestrator-state-development.yml` |
-   | research     | `brief-written`, `plan-created`, `findings-gathered`, `synthesis-complete`, `alternatives-generated`, `approaches-chosen`, `design-generated`, `research-completed`        | `skills/research/SKILL.md`       | `templates/orchestrator-state-research.yml`    |
-   | performance  | `phase-N` (legacy holdout — kept internally consistent)                                                                                                                     | `skills/performance/SKILL.md`    | `templates/orchestrator-state-performance.yml` |
-   | migration    | `state-analysed`, `target-planned`, `strategy-specified`, `plan-created`, `migration-executed`, `options-chosen`, `verification-done`, `issues-resolved`, `docs-generated`, `task-completed` | `skills/migration/SKILL.md`      | `templates/orchestrator-state-migration.yml`   |
-2. **Subskills are self-contained.** Each has: an entry check (validate state + prerequisite artifacts), an execute section (delegation via Skill/Task tools per the [Delegation Rules](delegation-rules.md)), and an exit (state update + closing ritual). **State updates are per-step**: a subskill writes `orchestrator-state.yml` immediately after each of its steps completes — appending only the step slug actually performed, bumping `orchestrator.updated`, recording failures in `failed_phases`/`auto_fix_attempts`, and validating with `verify_template` — never as a single end-of-skill write.
-3. **Closing ritual (handoff) → Exit Gate.** Every subskill ends with the **Exit Gate** contract, see [Gate Contract](gate-contract.md) (results box → results-acceptance question → next-step hint on accept → STOP).
-4. **Entry checks replace phase gates in assisted mode.** A subskill invoked directly must validate the same prerequisites a phase gate would (e.g., spec exists before audit) and route to the prerequisite's command if missing — via its **Entry Gate** ([Gate Contract](gate-contract.md)).
-5. **Never chain automatically.** The orchestrated wrapper invokes; subskills only SUGGEST the next command and stop. Auto-chaining from a subskill skips user review.
-6. **Assisted mode is the gate.** In assisted mode, the user explicitly invoking the next command IS the phase gate — no additional `question` confirmation before a subskill starts.
+# Dispatcher & Handoff Pattern (Subskill Loop)
+
+How orchestrators delegate phase bodies to user-invocable **subskills**. Extracted from `orchestrator-patterns.md` (Section 7) so skills can link to it directly.
+
+Orchestrators MAY delegate phase bodies to user-invocable **subskills** instead of executing them inline. Two modes share one state file and one set of subskills:
+
+| Mode              | Entry point              | Behavior                                                                                                |
+| ----------------- | ------------------------ | ------------------------------------------------------------------------------------------------------- |
+| Orchestrated mode | wrapper skill/command    | Invokes subskills back-to-back via Skill tool in sequence; pauses at `question` gates between subskills |
+| Assisted mode      | dispatcher skill/command | Derives next phase from state, prints the suggested command, and **STOPS**; user invokes each subskill  |
+
+## Rules
+
+1. **State file is canonical.** `orchestrator-state.yml` remains the single source of truth. Subskills read it on entry and write step results on exit. `completed_phases` values MUST stay stable so both modes intermix freely on the same task.
+   - **Value convention**: `completed_phases` / `failed_phases` / `auto_fix_attempts` keys are **descriptive step slugs**, not `phase-N` numbers. Canonical slugs per workflow:
+
+   | Workflow     | Slugs                                                                                                                                                                      | Routing table                    | Template                                       |
+   | ------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------- | ---------------------------------------------- |
+   | development  | `codebase-analysed`, `gap-analysed`, `tdd-red-proven`, `spec-written`, `spec-audited`, `plan-created`, `implementation-done`, `tdd-green-proven`, `options-chosen`, `verification-done`, `e2e-run`, `docs-generated`, `task-completed` | `skills/development/SKILL.md`  | `templates/orchestrator-state-development.yml` |
+   | research     | `brief-written`, `plan-created`, `findings-gathered`, `synthesis-complete`, `alternatives-generated`, `approaches-chosen`, `design-generated`, `research-completed`        | `skills/research/SKILL.md`       | `templates/orchestrator-state-research.yml`    |
+   | performance  | `codebase-analysed`, `bottlenecks-identified`, `spec-written`, `spec-audited`, `plan-created`, `implementation-done`, `options-chosen`, `verification-done`, `task-completed`                                                                                                                     | `skills/performance/SKILL.md`    | `templates/orchestrator-state-performance.yml` |
+   | migration    | `state-analysed`, `target-planned`, `strategy-specified`, `plan-created`, `migration-executed`, `options-chosen`, `verification-done`, `issues-resolved`, `docs-generated`, `task-completed` | `skills/migration/SKILL.md`      | `templates/orchestrator-state-migration.yml`   |
+2. **Subskills are self-contained.** Each has: an entry check (validate state + prerequisite artifacts), an execute section (delegation via Skill/Task tools per the [Delegation Rules](delegation-rules.md)), and an exit (state update + closing ritual). **State updates are per-step**: a subskill writes `orchestrator-state.yml` immediately after each of its steps completes — appending only the step slug actually performed, bumping `orchestrator.updated`, recording failures in `failed_phases`/`auto_fix_attempts`, and validating with `verify_template` — never as a single end-of-skill write.
+3. **Closing ritual (handoff) → Exit Gate.** Every subskill ends with the **Exit Gate** contract, see [Gate Contract](gate-contract.md) (results box → results-acceptance question → next-step hint on accept → STOP).
+4. **Entry checks replace phase gates in assisted mode.** A subskill invoked directly must validate the same prerequisites a phase gate would (e.g., spec exists before audit) and route to the prerequisite's command if missing — via its **Entry Gate** ([Gate Contract](gate-contract.md)).
+5. **Never chain automatically.** The orchestrated wrapper invokes; subskills only SUGGEST the next command and stop. Auto-chaining from a subskill skips user review.
+6. **Assisted mode is the gate.** In assisted mode, the user explicitly invoking the next command IS the phase gate — no additional `question` confirmation before a subskill starts.

package/dist/skills/orchestrator-framework/references/fix-loop-contract.md

@@ -1,65 +1,65 @@
-# Fix Loop Contract — User-Driven Fix-Then-Reverify
-
-Shared contract for skills that resolve verification issues together with the user. It pins the option wording, iteration budget, state-write cadence, and exit conditions so each applying skill describes only its workflow-specific deltas instead of restating the loop.
-
-**Applies to**:
-
-- `dev-verify` — the `verification-done` step after `implementation-verifier` returns `passed_with_issues` or `failed`
-- `migration-fix` — when the verification state facts say fixable issues remain (`issues-resolved`)
-
-Skills that use the loop MUST reference this file and document only their deltas in the "Workflow-Specific Deltas" section.
-
----
-
-## Loop (max 3 iterations)
-
-1. **Display the detailed issue breakdown** grouped by category and severity, with location, description, and fixability.
-2. **Present critical + warning issues as a numbered list.** Info issues are listed for awareness, not actionable.
-3. **Ask via `question`** — "Which issues should I fix?" with exactly three options:
-   - **"Fix all fixable issues"**
-   - **"Let me choose specific issues"**
-   - **"Skip fixes, proceed as-is"**
-4. **Fix the selected issues directly.** Immediately after each fix, append it to `verification_context.fixes_applied` in state — never defer to the end of the loop. Bump `orchestrator.updated`; re-read the state file + run `verify_template`.
-5. **After fixes, set `options.skip_test_suite: false`** (code changed — the suite must re-run on re-verification).
-6. **Ask via `question`** — "Re-run verification to check fixes?" with options:
-   - **"Yes, re-run verification"** → re-invoke `implementation-verifier` (Skill tool), then **immediately** update `verification_context.reverify_count` and refresh `verification_context.last_status` + `issues_found` from the verifier's structured return; return to step 1.
-   - **"No, proceed to next phase"** → exit the loop to the Exit Gate.
-7. **Record each user decision** (specific selection, proceed-with-warnings, deliberate skip) in `verification_context.decisions_made`.
-
-The loop is capped at **3 iterations**; at the cap, the question in "Exit Conditions" decides the outcome.
-
----
-
-## Exit Conditions
-
-| Condition | Action |
-| --------- | ------ |
-| ✅ No critical issues remain | Append the skill's success slug and proceed. |
-| ⚠ Max 3 iterations reached | `question` — "Proceed with known issues?" / stop-or-rollback option. Proceed ⇒ append the success slug with the approval recorded in `decisions_made`; stop/rollback ⇒ do NOT append, record the decision. |
-| Deliberate skip ("Skip fixes, proceed as-is") | A user-chosen skip, not a failure: no success slug, no failure entry; record the reason in `phase_summaries.<fix>.summary` + `decisions_made`. |
-| ❌ Critical issues remain unresolved | **MUST NOT proceed without explicit user approval.** |
-| Verifier itself fails | Do NOT append the success slug; append it to `orchestrator.failed_phases` and increment `auto_fix_attempts["<slug>"]`. |
-
----
-
-## State Writes (per iteration, never batched)
-
-- `fixes_applied` — after every fix.
-- `reverify_count` + refreshed `last_status` / `issues_found` — right after each verifier run.
-- `decisions_made` — after each user decision.
-- The success slug — only when the loop settles (✅, or an approved ⚠); never on a deliberate skip or a failure.
-
-**Single-writer split**: the verification step owns the initial `last_status` / `issues_found`; the fix loop owns `fixes_applied`, `reverify_count`, `decisions_made`, plus the re-verify refresh of `last_status` / `issues_found`.
-
----
-
-## Workflow-Specific Deltas
-
-### Development (`dev-verify`)
-
-None beyond this contract.
-
-### Migration (`migration-fix`)
-
-- **Compatibility verdict invalidation**: after fixes, set `verification_context.compatibility_status: null` (field-level fail-closed exception — the compatibility checks ran against pre-fix code). Only `migration-verify` may set a new verdict; when it is not `passed`, the required next step is `/owflow:migration-verify <task-path>`, never finalize.
-- **Data-integrity HALT**: data-integrity issues are NEVER auto-fixed. HALT, present the rollback option with the issue, and append the failure entry; rollback happens only on explicit user confirmation — no automatic rollback or revert ever. A deliberate skip cannot waive a data-integrity failure or any non-passed compatibility check.
+# Fix Loop Contract — User-Driven Fix-Then-Reverify
+
+Shared contract for skills that resolve verification issues together with the user. It pins the option wording, iteration budget, state-write cadence, and exit conditions so each applying skill describes only its workflow-specific deltas instead of restating the loop.
+
+**Applies to**:
+
+- `dev-verify` — the `verification-done` step after `implementation-verifier` returns `passed_with_issues` or `failed`
+- `migration-fix` — when the verification state facts say fixable issues remain (`issues-resolved`)
+
+Skills that use the loop MUST reference this file and document only their deltas in the "Workflow-Specific Deltas" section.
+
+---
+
+## Loop (max 3 iterations)
+
+1. **Display the detailed issue breakdown** grouped by category and severity, with location, description, and fixability.
+2. **Present critical + warning issues as a numbered list.** Info issues are listed for awareness, not actionable.
+3. **Ask via `question`** — "Which issues should I fix?" with exactly three options:
+   - **"Fix all fixable issues"**
+   - **"Let me choose specific issues"**
+   - **"Skip fixes, proceed as-is"**
+4. **Fix the selected issues directly.** Immediately after each fix, append it to `verification_context.fixes_applied` in state — never defer to the end of the loop. Bump `orchestrator.updated`; re-read the state file + run `verify_template`.
+5. **After fixes, set `options.skip_test_suite: false`** (code changed — the suite must re-run on re-verification).
+6. **Ask via `question`** — "Re-run verification to check fixes?" with options:
+   - **"Yes, re-run verification"** → re-invoke `implementation-verifier` (Skill tool), then **immediately** update `verification_context.reverify_count` and refresh `verification_context.last_status` + `issues_found` from the verifier's structured return; return to step 1.
+   - **"No, proceed to next phase"** → exit the loop to the Exit Gate.
+7. **Record each user decision** (specific selection, proceed-with-warnings, deliberate skip) in `verification_context.decisions_made`.
+
+The loop is capped at **3 iterations**; at the cap, the question in "Exit Conditions" decides the outcome.
+
+---
+
+## Exit Conditions
+
+| Condition | Action |
+| --------- | ------ |
+| ✅ No critical issues remain | Append the skill's success slug and proceed. |
+| ⚠ Max 3 iterations reached | `question` — "Proceed with known issues?" / stop-or-rollback option. Proceed ⇒ append the success slug with the approval recorded in `decisions_made`; stop/rollback ⇒ do NOT append, record the decision. |
+| Deliberate skip ("Skip fixes, proceed as-is") | A user-chosen skip, not a failure: no success slug, no failure entry; record the reason in `phase_summaries.<fix>.summary` + `decisions_made`. |
+| ❌ Critical issues remain unresolved | **MUST NOT proceed without explicit user approval.** |
+| Verifier itself fails | Do NOT append the success slug; append it to `orchestrator.failed_phases` and increment `auto_fix_attempts["<slug>"]`. |
+
+---
+
+## State Writes (per iteration, never batched)
+
+- `fixes_applied` — after every fix.
+- `reverify_count` + refreshed `last_status` / `issues_found` — right after each verifier run.
+- `decisions_made` — after each user decision.
+- The success slug — only when the loop settles (✅, or an approved ⚠); never on a deliberate skip or a failure.
+
+**Single-writer split**: the verification step owns the initial `last_status` / `issues_found`; the fix loop owns `fixes_applied`, `reverify_count`, `decisions_made`, plus the re-verify refresh of `last_status` / `issues_found`.
+
+---
+
+## Workflow-Specific Deltas
+
+### Development (`dev-verify`)
+
+None beyond this contract.
+
+### Migration (`migration-fix`)
+
+- **Compatibility verdict invalidation**: after fixes, set `verification_context.compatibility_status: null` (field-level fail-closed exception — the compatibility checks ran against pre-fix code). Only `migration-verify` may set a new verdict; when it is not `passed`, the required next step is `/owflow:migration-verify <task-path>`, never finalize.
+- **Data-integrity HALT**: data-integrity issues are NEVER auto-fixed. HALT, present the rollback option with the issue, and append the failure entry; rollback happens only on explicit user confirmation — no automatic rollback or revert ever. A deliberate skip cannot waive a data-integrity failure or any non-passed compatibility check.

package/dist/skills/orchestrator-framework/references/gate-contract.md

@@ -1,117 +1,122 @@
-# Gate Contract — Entry Gate & Exit Gate (Skill Contract)
-
-Shared contract for how every user-invocable skill opens and closes. Extracted from `orchestrator-patterns.md` (Section 9) so skills can link to it directly.
-
-Every **user-invocable** skill (orchestrators, subskills, dispatchers, utility commands) opens with an **Entry Gate** and closes with an **Exit Gate**. Internal skills (subagent-invoked, `user-invocable: false`) use lightweight input/structured-output contracts instead and never call `question`.
-
-## Entry Gate
-
-The Entry Gate runs BEFORE any phase work. It validates that this skill is allowed to run and routes to what is missing instead of guessing.
-
-1. **Argument resolution** — resolve the argument in priority order: full path → identifier (directory name under the workflow's task type) → fresh description. If missing, ambiguous, or unmatched:
-   - Print a structured ask: the exact inputs accepted, with format examples, then WAIT.
-   - NEVER guess, auto-pick a task, or proceed with a resolved-by-hope path.
-2. **Prerequisite check** — verify each required upstream artifact, expressed as a table the skill keeps at the top of its body:
-
-   | Required for this skill | Where verified                                                      | Produced by                  |
-   | ----------------------- | ------------------------------------------------------------------- | ---------------------------- |
-   | State file exists       | `<task-path>/orchestrator-state.yml`                                | `/owflow:development <desc>` |
-   | Analysis done           | `gap-analysed` in `completed_phases` + `analysis/gap-analysis.md` exists | `/owflow:dev-analyze`        |
-   | Spec approved           | `implementation/spec.md` exists                                     | `/owflow:dev-spec`           |
-
-   Verify **presence and, where cheap, content** (state field values like `has_reproducible_defect`, marker artifacts like `implementation/tdd-red-gate.md`), not just file existence.
-
-3. **Blocked output** — when a prerequisite is unmet, print the blocked block and STOP:
-   1. Numbered "Steps that must be completed first (in order)", each with its exact prefixed command (`/owflow:dev-analyze <task-path>`).
-   2. List of resumable task identifiers (directories under the workflow's task type), if any exist.
-   3. Fresh-start hint: `Run /owflow:<entry-command> <description> to start a task from scratch.`
-4. **Skip/resume semantics** — when the skill's steps are already in `completed_phases` (validate artifacts before trusting state): report the existing results briefly and route to the Exit Gate instead of re-executing.
-5. **Conditional activation** — skills that only apply under a state condition (e.g., TDD red gate requires `has_reproducible_defect: true`) check it as part of the Entry Gate and route to the correct alternative command when inactive.
-
-### Missing-state bootstrap (mid-pipeline entry)
-
-When a subskill is invoked with a task path/identifier whose `orchestrator-state.yml` exists nowhere (no task, no state file) and the ONLY unmet prerequisite is the state file itself, the skill does not blindly send the user back to the full pipeline — the user chose to enter mid-pipeline. Instead:
-
-1. **Ask** via `question`: "No task exists at this path. Create a fresh standard task and start at `<this skill's first step slug>`?"
-2. **On confirm (bootstrap)**:
-   - Create the task directory `.owflow/tasks/<workflow-type>/YYYY-MM-DD-task-name/` (name derived from the argument or the user's input) — mark `orchestrator.entry_point: "<skill> (mid-pipeline bootstrap)"` and `started_phase` with this skill's starting step slug.
-   - Create `orchestrator-state.yml` from the workflow's template, honoring any of the skill's command flags into their state fields; leave `completed_phases` empty (upstream steps stay unrecorded).
-   - Run `verify_template` against the workflow's template.
-   - Then continue into Execute. Because bootstrapping creates **state only — never upstream artifacts**, later artifact prerequisites (e.g., `implementation/spec.md` before planning) still block with the normal precondition loop, so an incorrectly early bootstrap cannot silently skip work.
-3. **On decline** — print the blocked block with the fresh-start hint and STOP.
-
-#### dev-* bootstrap
-
-- Task directory under `.owflow/tasks/development/`; template `orchestrator-state-development.yml`; fresh-start hint: `Run /owflow:development <description> to start a task from scratch.`
-- Skill starting points (used for the bootstrap question and `started_phase`): dev-analyze → `codebase-analysed`; dev-tdd-red → `tdd-red-proven`; dev-spec → `spec-written`; dev-plan → `plan-created`; dev-implement → `implementation-done`; dev-verify → `options-chosen`; dev-finalize → `e2e-run`.
-
-#### research-* bootstrap
-
-- Task directory under `.owflow/tasks/research/`; template `orchestrator-state-research.yml`; fresh-start hint: `Run /owflow:research <question> to start a task from scratch.`
-- Skill starting points (used for the bootstrap question and `started_phase`): research-plan → `brief-written`; research-gather → `findings-gathered`; research-synthesize → `synthesis-complete`; research-brainstorm → `alternatives-generated`; research-converge → `approaches-chosen`; research-design → `design-generated`; research-finalize → `research-completed`.
-
-#### migration-* bootstrap
-
-- Task directory under `.owflow/tasks/migrations/`; template `orchestrator-state-migration.yml`; fresh-start hint: `Run /owflow:migration <description> to start a migration task from scratch.`
-- Skill starting points (used for the bootstrap question and `started_phase`): migration-analyze → `state-analysed`; migration-target → `target-planned`; migration-spec → `strategy-specified`; migration-plan → `plan-created`; migration-implement → `migration-executed`; migration-verify → `options-chosen`; migration-fix → `issues-resolved`; migration-finalize → `task-completed`.
-
-### Deterministic entry checks (future)
-
-Entry-gate checks above are currently executed by the agent reading state. For more deterministic routing, future work considers (a) a small script that checks whether `orchestrator-state.yml` exists, parses it, and answers slug/next-step queries, and (b) an OpenCode lifecycle hook that automates these entry-gate checks for dev-* skills before they run. Until then, follow the rules above exactly as written.
-
-## Exit Gate
-
-The Exit Gate runs after all phase work and state updates are final. It presents results, asks the user to confirm them, and only then hands off.
-
-1. **Results box** — one-screen markdown summary (no ASCII borders, markdown headings/bold carry the emphasis):
-
-   ```markdown
-   ## ✅ <SKILL> COMPLETE — <task name>
-
-   **<Field 1>** — [key outcome]
-   **<Field 2>** — [key outcome]
-
-   **Artifacts**
-
-   - `relative/path/to/artifact.md`
-
-   **Next ▸** `/owflow:<next-skill> <task-path>`
-   ```
-
-   Conventions:
-   - Status glyphs follow the outcome: ✅ pass/success, ⚠ pass-with-concerns/partial, ❌ fail.
-   - Set the **Next ▸** line to the suggested next command for this skill; omit it when the next-step hint depends on the Accept response (the handoff block is printed after Accept anyway).
-
-2. **Results-acceptance question** (MANDATORY, fires before any handoff hint):
-
-   Use `question` — "Are these results correct?" with options:
-   - **Accept** — results are good; proceed to the next-step hint.
-   - **Adjust** — user specifies what to change; re-work ONLY the affected parts, update artifacts/state, then re-present the results box and re-ask.
-   - **Discuss** — walk through a specific result in more depth (evidence, reasoning, alternatives); after the discussion, re-ask.
-   - **Stop here** — artifacts persist; print the resume command (`/owflow:<this-skill> <task-path>` or the next phase command) and end.
-
-3. **Next-step hint** — only after Accept: print the suggested next command derived from state, plus an "Other options" block of alternative valid commands. Every suggested command is a **compact annotated entry**:
-   - **Purpose** — one line stating what the proposed skill does.
-   - **Requirement tag** — `required` (mandatory next phase) or `optional` (shortcut, standalone utility, or conditional step) with its activation condition, e.g. `optional — only when has_reproducible_defect: true`.
-   - **Remaining plan** (for optional/alternative paths) — one compact line listing the phases still ahead in execution order, marking which are optional.
-   - Standalone review/utility commands are annotated `optional — does not advance phases`.
-
-   Example shape (~5 lines):
-
-   ```
-   → /owflow:dev-verify <task-path>
-     What: runs the verification pipeline (completeness, code review, fixes) — required before commit. Remaining after: /owflow:dev-finalize (required).
-   Other options:
-   /owflow:reviews-code <task-path> — optional, does not advance phases (standalone re-run of code review)
-   /owflow:goal-development <task-path> — optional shortcut: runs all remaining phases in one loop (… → verify → finalize)
-   ```
-
-   Then STOP. Never auto-invoke the next skill.
-
-4. **Gate ordering** — per the Phase Gate Behavior state ordering rule in `orchestrator-patterns.md`: finish phase work → present results box → acceptance question → user responds → THEN the state is already consistent (per-step writes happened during Execute); the Exit Gate never mutates phase state on its own except recording the user's acceptance decision where the state schema has a field for it.
-
-## Exceptions
-
-- **Dispatchers** (`development`, `research`, `migration`): the Exit Gate's acceptance question is adapted — the results box is the handoff block, and the question asks how to proceed (hand off to the suggested subskill / switch to autonomous mode / adjust / stop).
-- **Utility skills with explicit no-follow-up contracts** (`agents-md-generator`, `rule-reviewer`): the acceptance question is confirm-or-revise only; follow-up suggestions stay prohibited unless the user asks. See [confirm-or-revise exception](confirm-or-revise-exception.md).
-- **Orchestrated mode** (`goal-development`, `goal-research`, `goal-migration`): the subskill's own Exit Gate acceptance question IS the loop gate — Accept means "continue to the next subskill". The wrapper MUST NOT add a second consecutive `question`.
+# Gate Contract — Entry Gate & Exit Gate (Skill Contract)
+
+Shared contract for how every user-invocable skill opens and closes. Extracted from `orchestrator-patterns.md` (Section 9) so skills can link to it directly.
+
+Every **user-invocable** skill (orchestrators, subskills, dispatchers, utility commands) opens with an **Entry Gate** and closes with an **Exit Gate**. Internal skills (subagent-invoked, `user-invocable: false`) use lightweight input/structured-output contracts instead and never call `question`.
+
+## Entry Gate
+
+The Entry Gate runs BEFORE any phase work. It validates that this skill is allowed to run and routes to what is missing instead of guessing.
+
+1. **Argument resolution** — resolve the argument in priority order: full path → identifier (directory name under the workflow's task type) → fresh description. If missing, ambiguous, or unmatched:
+   - Print a structured ask: the exact inputs accepted, with format examples, then WAIT.
+   - NEVER guess, auto-pick a task, or proceed with a resolved-by-hope path.
+2. **Prerequisite check** — verify each required upstream artifact, expressed as a table the skill keeps at the top of its body:
+
+   | Required for this skill | Where verified                                                      | Produced by                  |
+   | ----------------------- | ------------------------------------------------------------------- | ---------------------------- |
+   | State file exists       | `<task-path>/orchestrator-state.yml`                                | `/owflow:development <desc>` |
+   | Analysis done           | `gap-analysed` in `completed_phases` + `analysis/gap-analysis.md` exists | `/owflow:dev-analyze`        |
+   | Spec approved           | `implementation/spec.md` exists                                     | `/owflow:dev-spec`           |
+
+   Verify **presence and, where cheap, content** (state field values like `has_reproducible_defect`, marker artifacts like `implementation/tdd-red-gate.md`), not just file existence.
+
+3. **Blocked output** — when a prerequisite is unmet, print the blocked block and STOP:
+   1. Numbered "Steps that must be completed first (in order)", each with its exact prefixed command (`/owflow:dev-analyze <task-path>`).
+   2. List of resumable task identifiers (directories under the workflow's task type), if any exist.
+   3. Fresh-start hint: `Run /owflow:<entry-command> <description> to start a task from scratch.`
+4. **Skip/resume semantics** — when the skill's steps are already in `completed_phases` (validate artifacts before trusting state): report the existing results briefly and route to the Exit Gate instead of re-executing.
+5. **Conditional activation** — skills that only apply under a state condition (e.g., TDD red gate requires `has_reproducible_defect: true`) check it as part of the Entry Gate and route to the correct alternative command when inactive.
+
+### Missing-state bootstrap (mid-pipeline entry)
+
+When a subskill is invoked with a task path/identifier whose `orchestrator-state.yml` exists nowhere (no task, no state file) and the ONLY unmet prerequisite is the state file itself, the skill does not blindly send the user back to the full pipeline — the user chose to enter mid-pipeline. Instead:
+
+1. **Ask** via `question`: "No task exists at this path. Create a fresh standard task and start at `<this skill's first step slug>`?"
+2. **On confirm (bootstrap)**:
+   - Create the task directory `.owflow/tasks/<workflow-type>/YYYY-MM-DD-task-name/` (name derived from the argument or the user's input) — mark `orchestrator.entry_point: "<skill> (mid-pipeline bootstrap)"` and `started_phase` with this skill's starting step slug.
+   - Create `orchestrator-state.yml` from the workflow's template, honoring any of the skill's command flags into their state fields; leave `completed_phases` empty (upstream steps stay unrecorded).
+   - Run `verify_template` against the workflow's template.
+   - Then continue into Execute. Because bootstrapping creates **state only — never upstream artifacts**, later artifact prerequisites (e.g., `implementation/spec.md` before planning) still block with the normal precondition loop, so an incorrectly early bootstrap cannot silently skip work.
+3. **On decline** — print the blocked block with the fresh-start hint and STOP.
+
+#### dev-* bootstrap
+
+- Task directory under `.owflow/tasks/development/`; template `orchestrator-state-development.yml`; fresh-start hint: `Run /owflow:development <description> to start a task from scratch.`
+- Skill starting points (used for the bootstrap question and `started_phase`): dev-analyze → `codebase-analysed`; dev-tdd-red → `tdd-red-proven`; dev-spec → `spec-written`; dev-plan → `plan-created`; dev-implement → `implementation-done`; dev-verify → `options-chosen`; dev-finalize → `e2e-run`.
+
+#### research-* bootstrap
+
+- Task directory under `.owflow/tasks/research/`; template `orchestrator-state-research.yml`; fresh-start hint: `Run /owflow:research <question> to start a task from scratch.`
+- Skill starting points (used for the bootstrap question and `started_phase`): research-plan → `brief-written`; research-gather → `findings-gathered`; research-synthesize → `synthesis-complete`; research-brainstorm → `alternatives-generated`; research-converge → `approaches-chosen`; research-design → `design-generated`; research-finalize → `research-completed`.
+
+#### migration-* bootstrap
+
+- Task directory under `.owflow/tasks/migrations/`; template `orchestrator-state-migration.yml`; fresh-start hint: `Run /owflow:migration <description> to start a migration task from scratch.`
+- Skill starting points (used for the bootstrap question and `started_phase`): migration-analyze → `state-analysed`; migration-target → `target-planned`; migration-spec → `strategy-specified`; migration-plan → `plan-created`; migration-implement → `migration-executed`; migration-verify → `options-chosen`; migration-fix → `issues-resolved`; migration-finalize → `task-completed`.
+
+#### performance-* bootstrap
+
+- Task directory under `.owflow/tasks/performance/`; template `orchestrator-state-performance.yml`; fresh-start hint: `Run /owflow:performance <description> to start a task from scratch.`
+- Skill starting points (used for the bootstrap question and `started_phase`): performance-analyze → `codebase-analysed`; performance-spec → `spec-written`; performance-plan → `plan-created`; performance-implement → `implementation-done`; performance-verify → `options-chosen`; performance-finalize → `task-completed`.
+
+### Deterministic entry checks (future)
+
+Entry-gate checks above are currently executed by the agent reading state. For more deterministic routing, future work considers (a) a small script that checks whether `orchestrator-state.yml` exists, parses it, and answers slug/next-step queries, and (b) an OpenCode lifecycle hook that automates these entry-gate checks for dev-* skills before they run. Until then, follow the rules above exactly as written.
+
+## Exit Gate
+
+The Exit Gate runs after all phase work and state updates are final. It presents results, asks the user to confirm them, and only then hands off.
+
+1. **Results box** — one-screen markdown summary (no ASCII borders, markdown headings/bold carry the emphasis):
+
+   ```markdown
+   ## ✅ <SKILL> COMPLETE — <task name>
+
+   **<Field 1>** — [key outcome]
+   **<Field 2>** — [key outcome]
+
+   **Artifacts**
+
+   - `relative/path/to/artifact.md`
+
+   **Next ▸** `/owflow:<next-skill> <task-path>`
+   ```
+
+   Conventions:
+   - Status glyphs follow the outcome: ✅ pass/success, ⚠ pass-with-concerns/partial, ❌ fail.
+   - Set the **Next ▸** line to the suggested next command for this skill; omit it when the next-step hint depends on the Accept response (the handoff block is printed after Accept anyway).
+
+2. **Results-acceptance question** (MANDATORY, fires before any handoff hint):
+
+   Use `question` — "Are these results correct?" with options:
+   - **Accept** — results are good; proceed to the next-step hint.
+   - **Adjust** — user specifies what to change; re-work ONLY the affected parts, update artifacts/state, then re-present the results box and re-ask.
+   - **Discuss** — walk through a specific result in more depth (evidence, reasoning, alternatives); after the discussion, re-ask.
+   - **Stop here** — artifacts persist; print the resume command (`/owflow:<this-skill> <task-path>` or the next phase command) and end.
+
+3. **Next-step hint** — only after Accept: print the suggested next command derived from state, plus an "Other options" block of alternative valid commands. Every suggested command is a **compact annotated entry**:
+   - **Purpose** — one line stating what the proposed skill does.
+   - **Requirement tag** — `required` (mandatory next phase) or `optional` (shortcut, standalone utility, or conditional step) with its activation condition, e.g. `optional — only when has_reproducible_defect: true`.
+   - **Remaining plan** (for optional/alternative paths) — one compact line listing the phases still ahead in execution order, marking which are optional.
+   - Standalone review/utility commands are annotated `optional — does not advance phases`.
+
+   Example shape (~5 lines):
+
+   ```
+   → /owflow:dev-verify <task-path>
+     What: runs the verification pipeline (completeness, code review, fixes) — required before commit. Remaining after: /owflow:dev-finalize (required).
+   Other options:
+   /owflow:reviews-code <task-path> — optional, does not advance phases (standalone re-run of code review)
+   /owflow:goal-development <task-path> — optional shortcut: runs all remaining phases in one loop (… → verify → finalize)
+   ```
+
+   Then STOP. Never auto-invoke the next skill.
+
+4. **Gate ordering** — per the Phase Gate Behavior state ordering rule in `orchestrator-patterns.md`: finish phase work → present results box → acceptance question → user responds → THEN the state is already consistent (per-step writes happened during Execute); the Exit Gate never mutates phase state on its own except recording the user's acceptance decision where the state schema has a field for it.
+
+## Exceptions
+
+- **Dispatchers** (`development`, `research`, `performance`, `migration`): the Exit Gate's acceptance question is adapted — the results box is the handoff block, and the question asks how to proceed (hand off to the suggested subskill / switch to autonomous mode / adjust / stop).
+- **Utility skills with explicit no-follow-up contracts** (`agents-md-generator`, `rule-reviewer`): the acceptance question is confirm-or-revise only; follow-up suggestions stay prohibited unless the user asks. See [confirm-or-revise exception](confirm-or-revise-exception.md).
+- **Orchestrated mode** (`goal-development`, `goal-research`, `goal-performance`, `goal-migration`): the subskill's own Exit Gate acceptance question IS the loop gate — Accept means "continue to the next subskill". The wrapper MUST NOT add a second consecutive `question`.