---
name: openspec-execute
description: Scope guardrail and deviation classifier consulted during /gsd-execute-phase when a phase has a linked OpenSpec change. Not an executor itself — GSD's own executor subagents write the code; this skill only classifies what to do when an executor's discovery doesn't match the locked spec.
allowed-tools: Read
license: MIT
metadata:
  author: openspec-gsd-integration
  version: "0.1.0"
---

## Purpose

Give GSD's executor subagents (and the human reviewing `SUMMARY.md`) a fixed, shared vocabulary for "the spec and the reality on the ground disagree" — so a deviation gets classified and routed, never silently reinterpreted mid-execution.

## Invocation conditions

- Consulted by an executor (or the user) whenever implementation reveals a mismatch with `openspec/changes/<change_id>/specs/` or `design.md`.
- Not wired to a loop extension point — execution-time guardrails are advisory text an executor reads, not a step this integration forces (the phase loop's `execute:pre/wave:pre/wave:post/post` points are reserved for GSD's own executor/verifier roles; this capability does not inject at those points to avoid interfering with wave dispatch).

## Required inputs

- The specific mismatch encountered (freeform).
- `.planning/phases/<NN>-<slug>/<NN>-OPENSPEC-LINK.md` for the change/requirement IDs in scope.

## Files read

- `openspec/changes/<change_id>/specs/**/spec.md`, `design.md`
- The current `PLAN.md` and any `<NN>-<PP>-SUMMARY.md` already written this phase

## Files this skill may write

- None. This is a classification aid, not a writer — the executor itself records the classification in its own `SUMMARY.md` (a field GSD's native summary template already supports for deviations).

## Algorithm — classification table

| Situation | Classification | Correct next action |
|---|---|---|
| A requirement's wording admits two readings, and the difference matters for implementation | `SPEC_AMBIGUITY` | Stop, ask the user (or the discussing agent on the next session) to disambiguate; do not guess and proceed. |
| The spec says one thing, `design.md` says another, and both can't be true | `SPEC_CONFLICT` | Stop; this is an OpenSpec-internal inconsistency — route to `/opsx:update <change-id>` to reconcile, not a code fix. |
| The design is internally consistent but doesn't fit the actual codebase (e.g. assumes an API that doesn't exist) | `DESIGN_CONFLICT` | Stop; route to `/opsx:update <change-id>` for `design.md`, or a new change if the fix is itself a real design decision. |
| A dependency, permission, or external system blocks the task regardless of what the spec says | `IMPLEMENTATION_BLOCKER` | Record in `SUMMARY.md`, use GSD's own blocker mechanism (`gsd-tools state add-blocker`) — not an OpenSpec concern. |
| The executor finds work worth doing that no requirement asked for | `OUT_OF_SCOPE_DISCOVERY` | Do **not** implement it under this phase. Record it (GSD's own `<deferred>`/todo capture, or a note for a future `/opsx:propose`) and continue only the in-scope work. |

## Outputs

One classification tag per mismatch, surfaced in the executor's own `SUMMARY.md` deviations section (no new file format — reuses GSD's existing summary structure).

## Failure states

None of its own — every classification above already names its correct escalation path; there is no "unclassified" exit.

## Guardrails

- Never lets an executor silently reinterpret a requirement to make the easiest implementation "technically" satisfy it.
- Never lets `OUT_OF_SCOPE_DISCOVERY` work get implemented under the current phase's commits — scope creep here is exactly what `openspec-verify`'s Level 4 (design compliance) is built to catch, so it must not happen upstream of it.

## Idempotency

N/A — advisory classification, no persisted state of its own.

## Example

```
Executor: The spec says "the endpoint SHALL return 200 for a healthy dependency" but design.md
says "return 204 with no body." Which one?
[openspec-execute] SPEC_CONFLICT — stop, do not guess. Route to `/opsx:update add-health-check-endpoint`.
```