| name | conch-review-deep-review |
| description | Run an evidence-heavy deep review of a specific Conch subsystem, operation, or failure class. Use when the user wants a focused investigation such as lifecycle correctness, snapshot behavior, cleanup bugs, parity drift, docs drift, or trust-boundary analysis. |
Conch Review Deep Review
Use this skill when the question is no longer "where should we look?" but "what is actually wrong in this specific area?"
Inputs
target_repo: path to the Conch repository
focus: required subsystem, operation, or failure class
- optional
scope: narrower file or directory set
- optional
reviewers: explicit reviewer subset if already known
Default target repo: /home/stopire/Conch
Load These Assets
policies/severity.md
policies/report-template.md
Load reviewers based on focus:
- lifecycle ->
reviewers/lifecycle-reviewer.md
- image or snapshot ->
reviewers/image-snapshot-reviewer.md
- cleanup ->
reviewers/resource-leak-reviewer.md
- concurrency ->
reviewers/concurrency-reviewer.md
- parity ->
reviewers/api-sdk-parity-reviewer.md
- docs ->
reviewers/docs-implementation-auditor.md
- safety ->
reviewers/security-boundary-reviewer.md
- runtime validation ->
reviewers/sandbox-e2e-validator.md
Use these scripts:
scripts/discover_conch.py
scripts/collect_entrypoints.py
scripts/run_smoke.py when runtime evidence is needed
scripts/normalize_findings.py when you need a unified findings envelope
Workflow
1. Require a clear focus
If the user did not provide a real focus, stop and ask for one concise target such as:
- sandbox create/delete lifecycle
- build -> unpack -> snapshot chain
- cleanup after failed create
- SDK/API parity
- docs versus implementation
Do not silently degrade into a broad health review.
2. Discover the repo shape
Run:
python3 scripts/discover_conch.py <target_repo>
3. Collect likely entrypoints
Run:
python3 scripts/collect_entrypoints.py <target_repo>
Use the result to narrow the evidence set before dispatching reviewers.
4. Select and dispatch reviewers
Pick the reviewer subset that matches the focus. Use reviewer boundaries strictly. Deep review means tracing the path with enough surrounding code and documentation to support a strong conclusion.
5. Add runtime evidence when needed
If static evidence is not enough, run:
python3 scripts/run_smoke.py <target_repo> --mode probe-only
or, if appropriate:
python3 scripts/run_smoke.py <target_repo> --mode execute
Use runtime evidence to support the review, not to replace static reasoning.
6. Synthesize a focused report
Use policies/report-template.md as the output skeleton.
Produce:
- clear focus definition
- participating reviewers
- evidence-backed findings grouped by concern
- open questions
- recommended next action: fix, investigate further, or validate more deeply
Constraints
- Preserve reviewer boundaries.
- Call out uncertainty explicitly.
- Do not broaden into a general dashboard.
- If the task turns into pure runtime validation, switch to
conch-review-smoke.