Spec resolve

/spec-resolve <feature-description>

Build a resolved context bundle of relevant specifications for a feature. The bundle is a pre-filtered, token-budgeted input artifact for the implementation agent. You never navigate the .spec/ corpus manually — the bash resolver does all file discovery and filtering.

---

Pre-flight guard

Run corpus health check:

bash .claude/scripts/spec-stats.sh

If output says "Not initialized", tell the user to run /spec-init first. Stop.

---

Step 1 — Run the resolver script

bash .claude/scripts/spec-resolve.sh "$ARGUMENTS" 25000

Capture the full output as BUNDLE.

---

Step 2 — Domain completeness check (single inference step)

Read the Domains matched: line from the bundle header.

Then read ONLY .spec/CLAUDE.md — specifically the Domain Taxonomy table. No other files.

Ask yourself: given the feature description, could there be relevant domains not captured by keyword matching? Consider:

• Does the feature touch interfaces owned by an unmatched domain?
• Does the feature description use synonyms or abstractions that don't appear literally as domain names?

If you identify missing domains, run once more with override:

OVERRIDE_DOMAINS="storage,query,compaction" \
  bash .claude/scripts/spec-resolve.sh "$ARGUMENTS" 25000

Replace BUNDLE with this output.

Hard limit: maximum two script invocations. Do not loop.

If after two runs the domain set still seems incomplete, append a note to the bundle header:

Possible missing domains: <your assessment> — load manually if needed

---

Step 3 — Handle budget omissions

If Omitted (budget): is non-empty, append to the bundle:

## Budget Omissions
These specs were excluded due to token budget. If your feature directly
modifies their interfaces, load them individually:
<list omitted IDs>

---

Step 3.5 — Displacement resolution

If the bundle contains a ## Displacement section, the resolver detected that new spec requirements contradict existing APPROVED specs. Each displacement must be resolved before the bundle is complete.

Skip this step entirely if no ## Displacement section exists.

Parse and group

Parse each DISPLACED: line from the section. Group by existing spec ID so the user sees all displacements per spec together.

Resolve each displacement

For each displaced existing spec, display:

── Displacement detected ─────────────────────
  New: <new_id>.<req_id> — "<requirement text>"
  Existing: <existing_id>.<req_id> — "<requirement text>"
  Signal: <signal from DISPLACED line>

Read the full requirement text from both spec files to give the user context.

Use AskUserQuestion with options:

• "Accept" (description: invalidate the existing requirement — adds <existing_id>.<existing_req_id> to the new spec's invalidates array)
• "Narrow new spec" (description: revise the new spec to avoid the conflict — stop and re-run after editing)
• "Narrow old spec" (description: revise the existing spec to coexist — stop and re-run after editing)
• "Defer" (description: record as unresolved obligation and continue)

Process decisions

• Accept: Record <existing_id>.<existing_req_id> for the new spec's invalidates array. The spec-write step will include it when registering.
• Narrow new / Narrow old: Stop the resolve. Display which spec needs editing and suggest /spec-author to make the revision. The bundle is incomplete — the user must re-run /spec-resolve after editing.
• Defer: Append [UNRESOLVED: displacement — <new_id> vs <existing_id>] to the new spec's open_obligations. The spec will remain DRAFT and be excluded from downstream stages until resolved.

Append resolution summary

After all displacements are resolved (or deferred), append to the bundle:

## Displacement Resolution
| Existing Spec | Requirement | Decision | Action |
|---------------|-------------|----------|--------|
| <id> | <req_id> | Accept | → invalidates in new spec |
| <id> | <req_id> | Defer | → open_obligations in new spec |

If any decision was "Narrow new" or "Narrow old", the bundle output stops here — do not proceed to Step 4.

---

Step 4 — Output the bundle

Emit the complete bundle. Do not summarize, rewrite, or interpret it. The bundle is an input artifact for the implementation agent.

---

Hard constraints

• Maximum two script invocations per resolve call
• Never read spec files outside what the script returns
• Never invoke /spec-resolve recursively
• If NEEDS_DOMAIN_INFERENCE=true appears in script stderr, that is handled by step 2 — it is not an error