@opengroundplan/core 0.1.1

package/registry/commands/adr/command.md

@@ -0,0 +1,80 @@
+---
+description: Author a new ADR under docs/adr/ and add it to the ADR index in AGENTS.md
+argument-hint: <decision title>  e.g.  use the outbox pattern for domain events
+allowed-tools: Read, Write, Edit, Grep, Glob
+---
+
+# ADR: $ARGUMENTS
+
+Record an architectural decision and index it. Propose first, write after approval.
+
+## Phase 1 — Plan (read-only)
+
+1. If `$ARGUMENTS` is empty → ask what decision is being recorded. Never guess.
+
+2. Glob `docs/adr/*.md`. Take the highest existing `NNNN`, add 1, zero-pad to 4. None found → `0001`. Slugify `$ARGUMENTS` (lowercase, hyphens) for the filename.
+
+3. Grep `docs/adr/` for the same subject. An existing ADR on this topic means this is a **supersede**, not a fresh decision — say which one it replaces and ask whether to continue.
+
+4. Read `AGENTS.md` and locate the ADR index section. Missing → say so, and say exactly where it will be added.
+
+5. Glob the resolved target path. Exists → CONFLICT. Stop and offer overwrite / rename / cancel.
+
+6. Draft the ADR and show it in full:
+
+   ```markdown
+   # NNNN. <Title>
+
+   - Status: proposed
+   - Date: YYYY-MM-DD
+   - Supersedes: <NNNN or none>
+
+   ## Context
+   <the forces in play — what makes this decision necessary now>
+
+   ## Decision
+   <what we will do, active voice, present tense>
+
+   ## Consequences
+   <what gets easier, what gets harder, what we can no longer do>
+
+   ## Alternatives considered
+   - <option> — rejected because <specific reason>
+   ```
+
+7. If the decision changes an allowed dependency direction, say which `Boundaries` line in `AGENTS.md` it affects. Do not edit that line here — the ADR records the decision; changing the boundary is separate work.
+
+8. Print the manifest:
+   ```
+   Will create:
+     docs/adr/NNNN-<slug>.md              (NEW)
+   Will edit:
+     AGENTS.md                            (ADR index — one line appended)
+   Will NOT touch:
+     any existing ADR body                (a supersede changes only its Status line, on request)
+     AGENTS.md Boundaries/Conventions     (separate change)
+     any source file
+   Verification after apply:
+     the new ADR appears exactly once in the AGENTS.md index
+   ```
+
+9. STOP. Ask "Proceed? (yes / customize / cancel)".
+
+## Phase 2 — Apply (only after an explicit yes)
+
+10. Write `docs/adr/NNNN-<slug>.md` with the banner as its first line:
+    ```
+    <!-- GENERATED by /adr on YYYY-MM-DD — review before commit -->
+    ```
+
+11. Append **one line** to the ADR index in `AGENTS.md`: `- NNNN — <title> — <status>`. The index points at ADRs; it never restates their bodies. Duplicated content is how AGENTS.md drifts.
+
+12. Superseding an earlier ADR → change only its `Status:` line, and only after the user confirms. Never rewrite a past decision's reasoning; the record is why it exists.
+
+13. Report back:
+    ```
+    Created:    docs/adr/NNNN-<slug>.md
+    Indexed in: AGENTS.md
+    Supersedes: <NNNN | none>
+    Next:       /done <module> adr NNNN: <title>
+    ```

package/registry/commands/adr/command.yaml

@@ -0,0 +1,7 @@
+id: adr
+kind: command
+name: Architecture Decision Record
+description: Author docs/adr/NNNN-title.md and append it to the ADR index in AGENTS.md.
+tags: [architecture, decisions, documentation]
+targets: [claude-code, cursor, github-copilot, windsurf, cline, gemini-cli]
+requiresCapability: [slash-commands]

package/registry/commands/breaking/command.md

@@ -0,0 +1,72 @@
+---
+description: Find callers of a published symbol, report the blast radius, and record the decision
+argument-hint: <symbol-or-file>  e.g.  InvoiceRepository  or  src/billing/invoice-repository.ts
+allowed-tools: Read, Grep, Glob, Write, Edit
+---
+
+# Breaking Change: $ARGUMENTS
+
+Detect every consumer of a published symbol and record why the break is worth taking.
+
+## Phase 1 — Plan (read-only)
+
+1. If `$ARGUMENTS` is empty → ask for the symbol or file. Never guess.
+
+2. Read `AGENTS.md` → `Boundaries`. A module's **published interface** is whatever those lines mark as exported or depended on by another module. Everything else is internal and cannot break a consumer. Nothing published → say so and stop; there is no blast radius to compute.
+
+3. Resolve the input:
+   - A file path → read it, list the symbols it publishes, and ask which one is changing
+   - A symbol name → use it directly
+
+4. Grep for consumers across the modules the boundaries allow to depend on the owning module:
+   ```
+   Grep pattern: <symbol>  output_mode: files_with_matches
+   ```
+   Then read ±5 lines around each hit. A name match is not a usage.
+
+5. Classify every consumer:
+   - **HIGH** — implements, extends, or destructures the symbol's shape
+   - **MEDIUM** — calls it with a dependency on the signature or return shape
+   - **LOW** — type-only or name-only reference
+
+   List test consumers separately. They break too, and they are the cheapest early warning.
+
+6. Grep `.agent/DECISIONS.md` for the symbol — a prior decision may already record its migration path.
+
+7. Show the blast radius and the draft entry:
+   ```
+   ## YYYY-MM-DD — Breaking: <what changed>
+   Symbol:    <name> (<kind>) in <module>
+   Why:       <the reason the break is worth it>
+   Migration: <what consumers must do>
+   Consumers: <count> — <high> high / <medium> medium / <low> low
+   Status:    accepted
+   ```
+
+8. Print the manifest:
+   ```
+   Will append to:
+     .agent/DECISIONS.md                  (append-only)
+   Will NOT touch:
+     any consumer file                    (migration is real work, done deliberately)
+     the symbol itself
+     AGENTS.md
+   Verification after apply:
+     every HIGH consumer named above has a migration line
+   ```
+
+9. STOP. Ask "Proceed? (yes / customize / cancel)".
+
+## Phase 2 — Apply (only after an explicit yes)
+
+10. Append the entry to `.agent/DECISIONS.md`. Append-only — never rewrite or delete an existing entry, even a wrong one. Supersede it with a new entry instead.
+
+11. Report back:
+    ```
+    Symbol:    <name>
+    Consumers: <count> across <n> modules
+    Recorded:  .agent/DECISIONS.md
+    Next:      migrate HIGH consumers first, then /done <module> breaking: <symbol>
+    ```
+
+12. Log with `/done` only once consumers are actually updated — not when the decision is recorded.

package/registry/commands/breaking/command.yaml

@@ -0,0 +1,7 @@
+id: breaking
+kind: command
+name: Breaking Change Recorder
+description: Scan consumers of a published symbol, report the blast radius, and record it in .agent/DECISIONS.md.
+tags: [architecture, boundaries, decisions]
+targets: [claude-code, cursor, github-copilot, windsurf, cline, gemini-cli]
+requiresCapability: [slash-commands]

package/registry/commands/done/command.md

@@ -0,0 +1,50 @@
+---
+description: Append a compact one-line entry to .agent/DONE.md for completed work
+argument-hint: <module> <title>  e.g.  billing fix invoice rounding on refunds
+allowed-tools: Read, Write, Edit, Glob, Bash
+---
+
+# DONE Entry: $ARGUMENTS
+
+Append one line to the execution log. Propose the line, write after approval.
+
+## Phase 1 — Plan (read-only)
+
+1. If `$ARGUMENTS` is empty → ask for the module and the title. Never guess either.
+
+2. Parse `$ARGUMENTS`:
+   - First token = `<module>` — must be a module declared in the `Boundaries` section of `AGENTS.md`. Not declared → ask; do not invent a scope vocabulary this repo never agreed to.
+   - Rest = `<title>` — imperative, ≤10 words, describing the outcome not the activity.
+
+3. Resolve today's date as `YYYY-MM-DD` from the environment. Bash `date +%Y-%m-%d` only if the environment does not supply it.
+
+4. Infer `<files-or-area>` from the edits actually made in this session, not from the title.
+
+5. Glob `.agent/DONE.md`. Missing → it will be created with a `# DONE` heading. Existing → the line is appended; an append is never a CONFLICT.
+
+6. Print the proposed line and the manifest:
+   ```
+   Will append to:
+     .agent/DONE.md                       (under the ## YYYY-MM heading)
+   Line:
+     YYYY-MM-DD | <module> | <title> | <files-or-area>
+   Will NOT touch:
+     .agent/ACTIVE.md, .agent/DECISIONS.md, .agent/HANDOFF.md
+     any source file
+   Verification after apply:
+     re-read the tail of .agent/DONE.md
+   ```
+
+7. STOP. Ask "Proceed? (yes / customize / cancel)".
+
+## Phase 2 — Apply (only after an explicit yes)
+
+8. Append the line under the matching `## YYYY-MM` heading. Create the heading if this is the month's first entry.
+
+9. **Cap check:** count the entry lines (those starting with `20`). Over the log cap declared in `AGENTS.md` → propose moving the oldest entries verbatim to `.agent/DONE_ARCHIVE.md`. No cap declared → propose one once the log outgrows a single screen, and let the user record it in `AGENTS.md`. Propose only — never archive automatically.
+
+10. Do not write a second descriptive line unless the title is genuinely ambiguous. One line per entry is the default; the log is an index, not a narrative.
+
+11. The entry carries no GENERATED banner — it is a single line in a shared log, and a banner per entry would drown it.
+
+12. Report back the exact line that was added.

package/registry/commands/done/command.yaml

@@ -0,0 +1,7 @@
+id: done
+kind: command
+name: DONE Log Entry
+description: Append a compact one-line entry to .agent/DONE.md, with an archive check when the log grows.
+tags: [workflow, logging, handoff]
+targets: [claude-code, cursor, github-copilot, windsurf, cline, gemini-cli]
+requiresCapability: [slash-commands]

package/registry/commands/handoff/command.md

@@ -0,0 +1,65 @@
+---
+description: Create or update .agent/HANDOFF.md so a different AI tool can resume this work
+argument-hint: (no args)
+allowed-tools: Read, Write, Grep, Glob, Bash
+---
+
+# Session Handoff
+
+Capture the minimum state another tool needs to pick this up cold. Propose it, write after approval.
+
+## Phase 1 — Plan (read-only)
+
+1. `$ARGUMENTS` is ignored — this command reads the session, not an argument.
+
+2. Determine the current state:
+   - **Goal** — from `.agent/ACTIVE.md` if it exists, else from the user's first message this session
+   - **Progress** — which boxes in `specs/NNN-slug/tasks.md` are checked, and which task is next
+   - **Files edited** — `git status --short`. Bash is used here to read git state only; this command never runs a build and never changes the tree.
+   - **Build state** — green or red, plus the exact command that produced it, taken from the `commands.*` values in `AGENTS.md`. Not known → write `unknown`. Never guess a command and never claim a state you did not observe.
+
+3. If there is no next step, the work is complete → propose **deleting** `.agent/HANDOFF.md` instead of writing one. A stale handoff is worse than none. Stop here and ask.
+
+4. Glob `.agent/HANDOFF.md`. Existing → this is an update, not a CONFLICT; show what changes.
+
+5. Draft the file (≤30 lines) and show it in full:
+
+   ```markdown
+   # HANDOFF — YYYY-MM-DD HH:MM
+
+   ## Goal
+   <one sentence>
+
+   ## Status
+   - Spec:  <specs/NNN-slug/ or none>
+   - Done:  <tasks completed>
+   - Next:  <the single next step>
+   - Build: <green|red|unknown — the command that produced it>
+
+   ## Resume hint
+   <one paragraph: what to open first, what to verify before continuing, what already failed>
+
+   ## Uncommitted
+   <git status summary, or "clean">
+   ```
+
+6. Print the manifest:
+   ```
+   Will create/update:
+     .agent/HANDOFF.md                    (<NEW|UPDATE> — <n> lines)
+   Will NOT touch:
+     .agent/ACTIVE.md, .agent/DONE.md, .agent/DECISIONS.md
+     any source file
+   Verification after apply:
+     the file is ≤30 lines and names exactly one next step
+   ```
+
+7. STOP. Ask "Proceed? (yes / customize / cancel)".
+
+## Phase 2 — Apply (only after an explicit yes)
+
+8. Write `.agent/HANDOFF.md`. No GENERATED banner here — a handoff is live state, not scaffolding, and the banner would spend one of its 30 lines.
+
+9. Do not include exploration timelines, narrative, or session chatter. The next tool needs state, not a story.
+
+10. Report back: the path written and a 2-line preview.

package/registry/commands/handoff/command.yaml

@@ -0,0 +1,7 @@
+id: handoff
+kind: command
+name: Session Handoff
+description: Write a ≤30-line .agent/HANDOFF.md so a different AI tool can resume the work.
+tags: [workflow, handoff, cross-tool]
+targets: [claude-code, cursor, github-copilot, windsurf, cline, gemini-cli]
+requiresCapability: [slash-commands]

package/registry/commands/plan/command.md

@@ -0,0 +1,81 @@
+---
+description: Turn an approved spec into specs/NNN-slug/plan.md — the HOW
+argument-hint: <NNN-slug>  e.g.  003-invoice-export
+allowed-tools: Read, Write, Grep, Glob
+---
+
+# Plan: $ARGUMENTS
+
+Translate an approved spec into a technical plan. Propose first, write only after approval.
+
+## Phase 1 — Plan (read-only)
+
+1. If `$ARGUMENTS` is empty → list the directories under `specs/` and ask which one. Never guess.
+
+2. Resolve `specs/$ARGUMENTS/spec.md`. Missing → stop. The WHAT must be settled first (`/spec`).
+
+3. Read the spec. If it still lists open questions, stop and show them — a plan built on an unsettled spec is rework.
+
+4. Read `AGENTS.md`: boundaries (where code may live), conventions (how it must be written), protected paths (what needs approval), and the agent-kit `commands.*` values (how it gets verified).
+
+5. Glob `specs/$ARGUMENTS/plan.md`. Exists → CONFLICT. Stop and offer overwrite / rename / cancel.
+
+6. Draft the plan and show it in full:
+
+   ```markdown
+   # Plan — NNN <Title>
+
+   ## Approach
+   <the design in ≤10 lines>
+
+   ## Rejected alternative
+   <one option, and the specific reason it lost>
+
+   ## Modules touched
+   | Module | What changes | Dependency direction relied on |
+   |---|---|---|
+
+   ## Data & contracts
+   <schema, API, or type changes — each marked additive or breaking>
+
+   ## Risks
+   <what could go wrong, and the cheapest way to find out early>
+
+   ## Verification
+   <the commands.* values from AGENTS.md, verbatim>
+   ```
+
+7. Every module in the plan must be one the boundaries declare, and every dependency it relies on must already be allowed. A plan that needs a new edge is an ADR first (`/adr`), not a plan.
+
+8. Never invent a build or test command for the Verification section. Take the declared `commands.*` strings; none declared → write `not declared in AGENTS.md` and flag it.
+
+9. Print the manifest:
+   ```
+   Will create:
+     specs/<NNN-slug>/plan.md             (NEW)
+   Will NOT touch:
+     specs/<NNN-slug>/spec.md             (the WHAT is settled)
+     AGENTS.md                            (boundaries are input, not output)
+     any source file                      (implementation comes after /tasks)
+   Verification after apply:
+     <build/test/lint commands read from AGENTS.md>
+   ```
+
+10. STOP. Ask "Proceed? (yes / customize / cancel)".
+
+## Phase 2 — Apply (only after an explicit yes)
+
+11. Write `specs/$ARGUMENTS/plan.md` with the banner as its first line:
+    ```
+    <!-- GENERATED by /plan on YYYY-MM-DD — review before commit -->
+    ```
+
+12. Report back:
+    ```
+    Created:  specs/<NNN-slug>/plan.md
+    Modules:  <list>
+    Breaking: <yes — what | no>
+    Next:     /tasks <NNN-slug>
+    ```
+
+13. Do not add a `.agent/DONE.md` entry — planning is not shipped work.

package/registry/commands/plan/command.yaml

@@ -0,0 +1,7 @@
+id: plan
+kind: command
+name: Spec Plan
+description: Turn an approved spec into specs/NNN-slug/plan.md via a plan → approve → apply gate.
+tags: [workflow, specs, planning]
+targets: [claude-code, cursor, github-copilot, windsurf, cline, gemini-cli]
+requiresCapability: [slash-commands]

package/registry/commands/review/command.md

@@ -0,0 +1,54 @@
+---
+description: Run the read-only reviewer agents over the current diff and aggregate one verdict
+argument-hint: [path-or-range]  e.g.  src/billing/  or  main..HEAD  (defaults to the working diff)
+allowed-tools: Read, Grep, Glob, Bash, Task
+---
+
+# Review: $ARGUMENTS
+
+Fan the reviewer subagents out over a diff and merge their verdicts into one report.
+
+## Steps
+
+1. Resolve the diff. `$ARGUMENTS` empty → the uncommitted working diff. A path → the diff limited to that path. A range → that range. Bash is used for `git diff` and `git diff --name-only` only; this command never builds, never tests, never writes.
+
+2. Empty diff → say so and stop. There is nothing to review.
+
+3. Read `AGENTS.md` once and pass its boundaries, conventions, and security lines to every agent, so all of them judge against the same declared rules rather than their own priors.
+
+4. Dispatch only the agents the diff actually warrants:
+
+   | Condition in the diff | Agent |
+   |---|---|
+   | any source file changed | `solid-reviewer` |
+   | an import crosses a module edge, or a module is added | `architecture-reviewer` |
+   | auth, input handling, secrets, serialization, paths, or outbound calls | `security-reviewer` |
+   | test files added or changed | `test-coverage-reviewer` |
+   | a published symbol changed shape | `breaking-change-detector` |
+   | a `specs/NNN-slug/` exists for this work | `spec-consistency-checker` |
+
+   Run the selected agents in parallel. Do not run one whose condition is false — an agent with nothing to judge returns noise that buries the real findings.
+
+5. Merge the results. De-duplicate by `file:line + issue`; when two agents flag the same line, keep the higher severity and name both.
+
+6. Report back:
+   ```
+   DIFF:    <n> files, +<n>/-<n>
+   AGENTS:  <ran: ...>  <skipped: ... — why>
+
+   CRITICAL (<n>)
+     <path>:<line> — <summary>   [<agent>]
+   WARNING (<n>)
+     <path>:<line> — <summary>   [<agent>]
+   NIT (<n>)
+     <path>:<line> — <summary>   [<agent>]
+
+   AGENT VERDICTS
+     <agent>: PASS | FAIL
+
+   VERDICT: PASS | FAIL
+   ```
+
+   The aggregate is `FAIL` if any agent returned `FAIL`. Never soften an agent's verdict when merging.
+
+7. Do **not** auto-edit. Report findings only.

package/registry/commands/review/command.yaml

@@ -0,0 +1,8 @@
+id: review
+kind: command
+name: Aggregate Review
+description: Run the read-only reviewer subagents over the current diff and merge their verdicts into one report.
+tags: [review, quality, read-only]
+# Narrower than the other commands: this one dispatches reviewers, so it needs subagents too.
+targets: [claude-code, github-copilot]
+requiresCapability: [slash-commands, subagents]

package/registry/commands/scope/command.md

@@ -0,0 +1,45 @@
+---
+description: Detect the scope of a path from the repo's declared boundaries and load only that context
+argument-hint: <file-or-directory-path>  e.g.  src/billing/invoice-service.ts
+allowed-tools: Read, Grep, Glob
+---
+
+# Scope Detection: $ARGUMENTS
+
+Determine which declared module owns the given path and load the minimum context needed to work in it.
+
+## Steps
+
+1. If `$ARGUMENTS` is empty → ask which path to scope. Never guess a default.
+
+2. Read `AGENTS.md` and extract the **Boundaries** section (the agent-kit `boundaries` values). Each line declares a module, the paths it owns, and the direction it may depend. This is the path→scope map — there is no hardcoded table.
+
+3. Match `$ARGUMENTS` against the declared boundaries. Longest path prefix wins. Pull from `AGENTS.md` only:
+
+   | Section | What to take |
+   |---|---|
+   | `Boundaries` | the owning module, its paths, its allowed dependency direction |
+   | `Conventions` | only the rules that name the matched module |
+   | `Protected paths` | whether this path needs explicit approval before a write |
+
+   No match → report `Scope: unmapped`, name the nearest declared module, and stop. Do not invent a scope the repo never declared.
+
+4. Load **only** what the user's intent needs:
+   - always: the matched boundary line and the conventions that name that module
+   - `.agent/ACTIVE.md` — if present, to see whether this path is already in flight
+   - `specs/NNN-slug/spec.md` — only when `ACTIVE.md` points at a spec covering this path
+   - `.agent/DECISIONS.md` — only when the user is asking *why* something is the way it is (grep, don't read whole)
+
+5. Report back:
+   ```
+   Scope:     <module declared in AGENTS.md>
+   Owns:      <declared paths>
+   Depends:   <allowed dependency direction>
+   Protected: <yes — why | no>
+   Loaded:    <files read>
+   Next:      <what to do given the user's intent>
+   ```
+
+6. Do **not** read files belonging to other modules — that is the context bloat this command exists to prevent.
+
+7. Do **not** auto-edit. Report findings only.

package/registry/commands/scope/command.yaml

@@ -0,0 +1,7 @@
+id: scope
+kind: command
+name: Scope Detection
+description: Detect which declared module owns a path and load only that module's context.
+tags: [context, boundaries, read-only]
+targets: [claude-code, cursor, github-copilot, windsurf, cline, gemini-cli]
+requiresCapability: [slash-commands]

package/registry/commands/spec/command.md

@@ -0,0 +1,81 @@
+---
+description: Create specs/NNN-slug/spec.md — the WHAT and WHY, with no HOW
+argument-hint: <short feature title>  e.g.  export invoices as CSV
+allowed-tools: Read, Write, Grep, Glob
+---
+
+# Spec: $ARGUMENTS
+
+Capture what is being built and why, before anyone decides how.
+
+## Phase 1 — Plan (read-only)
+
+1. If `$ARGUMENTS` is empty → ask what is being specified. Never guess a title.
+
+2. Glob `specs/*/` and pick the number: take the highest existing `NNN`, add 1, zero-pad to 3. No specs yet → `001`. Slugify `$ARGUMENTS` (lowercase, hyphens, ≤5 words) for the rest of the directory name.
+
+3. Grep `specs/` for the same subject. A near-duplicate → name the overlapping spec and ask whether to extend that one instead of opening a second front on the same problem.
+
+4. Read `AGENTS.md` for the declared boundaries. A spec names which **modules** are in scope; it does not name files.
+
+5. Glob `specs/<NNN-slug>/spec.md`. Exists → CONFLICT. Stop and offer overwrite / rename / cancel.
+
+6. Draft the spec and show it in full:
+
+   ```markdown
+   # NNN — <Title>
+
+   ## Why
+   <the problem, and what stays broken if nothing changes>
+
+   ## What
+   <the behavior from the outside — what a user or caller can now do>
+
+   ## Requirements
+   - R1. <a testable statement of behavior>
+   - R2. ...
+
+   ## Non-goals
+   - <what this explicitly does not do — the line that stops scope creep later>
+
+   ## Acceptance criteria
+   - [ ] <observable, checkable outcome>
+
+   ## Open questions
+   - <anything that must be answered before /plan>
+   ```
+
+7. Strip every HOW: no file paths, no class or table names, no library choices, no algorithms. Anything cut here belongs in `/plan`. Say what you removed and where it went.
+
+8. Every requirement must be falsifiable. "Fast", "clean", "robust" are not requirements — replace them or move them to Open questions.
+
+9. Print the manifest:
+   ```
+   Will create:
+     specs/<NNN-slug>/spec.md             (NEW)
+   Will NOT touch:
+     any existing spec                    (overlap is resolved by asking, not by editing)
+     AGENTS.md, .agent/*                  (a spec is not a decision or a log entry)
+     any source file
+   Verification after apply:
+     every requirement is falsifiable and no HOW survived the cut
+   ```
+
+10. STOP. Ask "Proceed? (yes / customize / cancel)".
+
+## Phase 2 — Apply (only after an explicit yes)
+
+11. Write `specs/<NNN-slug>/spec.md` with the banner as its first line:
+    ```
+    <!-- GENERATED by /spec on YYYY-MM-DD — review before commit -->
+    ```
+
+12. If `.agent/ACTIVE.md` exists, show the one-line update pointing at this spec — as a snippet for the user to paste. Do not write it; what is in flight is the user's call, not the generator's.
+
+13. Report back:
+    ```
+    Created:      specs/<NNN-slug>/spec.md
+    Requirements: <n>
+    Open questions: <n>
+    Next:         answer the open questions, then /plan <NNN-slug>
+    ```

package/registry/commands/spec/command.yaml

@@ -0,0 +1,7 @@
+id: spec
+kind: command
+name: Feature Spec
+description: Create specs/NNN-slug/spec.md capturing the WHAT and WHY, with the HOW deferred to /plan.
+tags: [workflow, specs, requirements]
+targets: [claude-code, cursor, github-copilot, windsurf, cline, gemini-cli]
+requiresCapability: [slash-commands]

package/registry/commands/tasks/command.md

@@ -0,0 +1,70 @@
+---
+description: Turn an approved plan into specs/NNN-slug/tasks.md — an ordered, checkable list
+argument-hint: <NNN-slug>  e.g.  003-invoice-export
+allowed-tools: Read, Write, Grep, Glob
+---
+
+# Tasks: $ARGUMENTS
+
+Break an approved plan into ordered tasks, each naming the files it touches.
+
+## Phase 1 — Plan (read-only)
+
+1. If `$ARGUMENTS` is empty → list the directories under `specs/` and ask which one. Never guess.
+
+2. Read `specs/$ARGUMENTS/plan.md`. Missing → stop. The HOW must exist first (`/plan`).
+
+3. Read `specs/$ARGUMENTS/spec.md` for the acceptance criteria. Every criterion must be reachable by at least one task, or the list is incomplete.
+
+4. Read `AGENTS.md` for the declared boundaries, protected paths, and `commands.*` values.
+
+5. Glob `specs/$ARGUMENTS/tasks.md`. Exists → CONFLICT. Stop and offer overwrite / rename / cancel.
+
+6. Draft the list and show it in full. Every task obeys these rules:
+   - one line, imperative, starting with `- [ ]`
+   - **names the exact files it touches** — a task without file paths is a wish, not a task
+   - ordered so each task leaves the repo buildable; nothing depends on a later task
+   - small enough to finish in one sitting; too big → split it
+   - a task touching a declared protected path is marked `(needs approval)`
+
+   ```markdown
+   ## Tasks
+
+   - [ ] 1. <imperative> — `<path>`, `<path>`
+   - [ ] 2. <imperative> — `<path>`  (needs approval)
+
+   ## Verification
+   <the commands.* values from AGENTS.md, verbatim>
+   ```
+
+7. Do not invent a build or test command for the Verification section. Take the declared strings; none declared → write `not declared in AGENTS.md` and flag it.
+
+8. Print the manifest:
+   ```
+   Will create:
+     specs/<NNN-slug>/tasks.md            (NEW)
+   Will NOT touch:
+     specs/<NNN-slug>/spec.md, plan.md    (both settled)
+     any source file                      (this writes a list, not code)
+   Verification after apply:
+     every acceptance criterion in spec.md maps to ≥1 task
+   ```
+
+9. STOP. Ask "Proceed? (yes / customize / cancel)".
+
+## Phase 2 — Apply (only after an explicit yes)
+
+10. Write `specs/$ARGUMENTS/tasks.md` with the banner as its first line:
+    ```
+    <!-- GENERATED by /tasks on YYYY-MM-DD — review before commit -->
+    ```
+
+11. Report back:
+    ```
+    Created:   specs/<NNN-slug>/tasks.md
+    Tasks:     <n> (<n> need approval)
+    Uncovered: <acceptance criteria with no task, or "none">
+    Next:      start task 1
+    ```
+
+12. Do not check any box and do not start the work. A box checked for work that was not done is how a task list starts lying.