| name | speckit-multi-model-review-apply-review |
| description | Ingest the reviewer's review-report.md and drive remediation, optionally using configured subagent routing for accepted fixes. |
| compatibility | Requires spec-kit project structure with .specify/ directory |
| metadata | {"author":"github-spec-kit","source":"multi-model-review:commands/apply-review.md"} |
speckit.multi-model-review.apply-review
Spec Kit extension command: /speckit.multi-model-review.apply-review
Legacy Claude plugin command: /multi-model-review:apply-review
Consume a reviewer agent's review-report.md and walk through the findings.
Arguments: $ARGUMENTS
Steps
When the work-assist orchestration layer is available, use it to plan and apply accepted fixes: Superpowers systematic-debugging and subagent-driven-development to plan, /ulw to batch independent findings, and omo to implement-and-verify. Fall back to the mmr-* subagents otherwise, keep every user confirmation gate intact, and never pass secrets into their prompts. See skills/multi-model-review/SKILL.md -> "Work-assist orchestration".
-
Locate the package.
- If
$ARGUMENTS contains a path, use it.
- Else pick the latest
.cross-review/packages/*/ that contains review-report.md.
- If none exist, abort with: "No review-report.md found. Did the reviewer save its output?"
-
Parse review-report.md using .specify/templates/review-report.md.
Context sufficiency
Verdict
Summary
- findings with:
severity
confidence
location
summary
detail
suggested_fix
- Read
metadata.json when present for:
subagent_routing
- selected implementation model/options
- package scope and path filters
-
Handle context sufficiency first.
- If
Context sufficiency = needs-full-package:
- present the current findings
- do not start broad edits yet
- recommend
/multi-model-review:review-package --full
- if the findings are clearly path-local, recommend
/multi-model-review:review-package --paths <subset>
- stop unless the user explicitly wants to continue with the current report
-
Filter findings.
- Default: drop findings with
confidence < 70
--min-confidence N overrides
- Group remaining findings by severity, then file
-
Present a numbered checklist.
- Quote the reviewer's wording instead of paraphrasing it
- Do not edit files yet
-
For each finding the user accepts:
- read the target file at the cited line
- propose one edit for that finding
- if
subagent_routing.mode = auto and the accepted finding is path-local, route the fix to mmr-implementation-worker
- if the finding is cross-cutting, ambiguous, migration-heavy, or security-sensitive, ask
mmr-heavy-planner for a remediation plan before editing
- apply only after confirmation
- if two findings touch the same line, re-read between edits
- record status in
.cross-review/packages/<pkg>/review-state.json
-
After the pass:
- if subagent routing is enabled and changes were made, run
mmr-review-checker for a read-only local preflight when the scope is small enough
- summarize applied vs skipped findings
- remind the user that a re-review means:
/multi-model-review:review-package
- run the reviewer again
-
Finish with the Completion token report.
- Append the
**Token, Headroom & RTK** block from skills/multi-model-review/SKILL.md (section "Completion token report").
- The ingest pass itself rarely gathers large shell output, so RTK and Headroom are often
used=no here; read rtk gain and headroom_stats for any session savings, and read the package metadata.json compressed_blocks to credit what the earlier packaging compressed.
- Report the two layers separately from measured stats only; never estimate, and keep secrets out of the report.
**Token, Headroom & RTK**
- **RTK**: used=<yes|no> — saved ≈ <N> tok (<P>%) · via `rtk gain`
- **Headroom**: used=<yes|no> — saved ≈ <N> tok (<P>%) · via `headroom_stats` / package `compressed_blocks`
- **Combined saved**: ≈ <RTK+Headroom> tok (only when both layers are measured)
- **Work-assist** (orchestration, usage only): ulw=<used|n/a>, omo=<used|n/a>, lazycodex=<used|n/a>, superpowers=<used|n/a>
- **Subagent routing** (usage only): scout=<used|n/a>, worker=<used|n/a>, heavy-planner=<used|n/a>, review-checker=<used|n/a>
Guardrails
- Never auto-apply a
critical finding without explicit confirmation.
- Never merge multiple findings into one edit.
- If the compact package was declared insufficient, prefer re-packaging before broad remediation.
- Do not use subagents to bypass user confirmation for review fixes.