| name | deep-code-scan |
| description | Multi-lane, decomposition-first codebase scan campaign that writes markdown docs only. Use when the user asks for a deep code scan, comprehensive audit, whole-repo issue hunt, or organized remediation plan spanning multiple lanes such as correctness, tests, performance, security, quality, structure, or architecture without adopting agile-workflow. Interviews for goal, lanes, rigor, scanner tier, and altitude bands; maps the repo; fans out scoped scanners; runs a review gauntlet; and consolidates findings into a code-audit scan folder of markdown documents.
|
Deep Code Scan
Orchestrate a goal-driven scan campaign: decompose the codebase, scan from narrow components outward,
verify findings, and produce an organized remediation plan. This standalone variant writes markdown
documents under code-audit/scan-<goal>/. It never creates .work/ items, backlog entries, release
gates, or commits.
Use this when the scan spans multiple lanes or needs the decompose -> scan -> review -> consolidate
campaign shape. For a single-domain pass, route to the specialist skill: bug-scan, perf-scout,
repo-eval, or bold-refactor.
Deliverables
Default output root:
code-audit/scan-<goal>/
00-plan.md
01-component-map.md
02-findings-ledger.md
03-review-gauntlet.md
04-remediation-plan.md
Temporary scanner packets may live under code-audit/scan-<goal>/.artifacts/ while the campaign is
running. After final collation, remove that scratch directory unless the user asks to keep raw
packets.
Composition, Not Replacement
In-campaign scanners reuse local reference catalogs. Do not invoke specialist skills mid-campaign;
load their references as scanner knowledge:
| Lane | Reference source | Standalone alternative |
|---|
| correctness | ../bug-scan/references/* | bug-scan <path> |
| performance | ../perf-scout/references/* | perf-scout <path> |
| quality / holistic | ../repo-eval/references/* | repo-eval <path> |
| architecture / bold | ../bold-refactor/SKILL.md conceptual lenses | bold-refactor <path> |
| tests | lane-catalog test guidance | focused test-audit scanner |
| tests | lane-catalog test guidance | test-scan <path> |
| security | lane-catalog security guidance | security-scan <path> |
| structure / refactor | lane-catalog refactor guidance plus project pattern docs | focused structure scanner |
| custom | bespoke brief from the user's goal | none |
Dials
Set these with the user before scanning:
- Goal: the north star, such as
all-issues, correctness, performance, or a custom phrase.
- Lanes: selected scan domains.
all-issues selects correctness, tests, performance, security,
quality, and structure. Architecture/bold is opt-in.
- Rigor:
floor, standard, or full.
- Scanner tier:
mixed, codex-high, codex-xhigh, claude-opus, claude-sonnet, or the
nearest host-native equivalents.
- Altitude bands: default
leaf -> module -> subsystem -> system; trim for small repos.
Compute an agent budget before scanning. If the plan implies more than about 150 total scanner/review
calls or more than 20 scanners in one wave, surface that cost and narrow the plan unless the user
explicitly accepts it.
Phase 1: Kickoff Interview
Read enough of the repo to propose grounded dials: top-level layout, manifests, languages, entry
points, tests, and docs. Ask concise questions or use a structured question tool to settle the dials.
Do not proceed to fan-out until the user has approved the goal, lanes, bands, and budget.
Phase 2: Decompose The Codebase
Build the altitude component map. Read references/decomposition.md
for the method.
Output the approved map to 01-component-map.md: band -> components, where each component has a
stable slug, role, file set, and parent edges for roll-up.
Phase 3: Write The Plan
Create 00-plan.md before scanning. It records:
- goal and scope;
- selected lanes and why;
- altitude bands and component counts;
- scanner tier per lane;
- rigor;
- estimated scanner calls and review calls;
- output root and temporary artifact policy.
Stop and present the plan. Adjust if the user redirects.
Phase 4: Drive The Scan
Walk each lane and its bands from narrow to broad. At each band, fan out parallel scanner sub-agents
across components. Read references/scanner-brief.md for the dispatch
brief and finding schema.
Each scanner is read-only over source code and writes only its assigned temporary markdown packet.
The orchestrator spot-checks candidate findings, writes accepted findings to
02-findings-ledger.md, and rolls concise inherited findings upward to the next band.
Phase 5: Review Gauntlet
Run the review gauntlet before consolidation. Read
references/review-gauntlet.md. At standard rigor, run at least two
fresh-context review rounds; at full, run to convergence or cap at four rounds. Record the ledger
in 03-review-gauntlet.md.
Phase 6: Consolidate
Cluster surviving findings into an implementation-neutral remediation plan. Read
references/consolidation.md. Write 04-remediation-plan.md with:
- findings grouped by fix locality and theme;
- severity and confidence;
- recommended sequence;
- risks and intent constraints;
- validation checks to run before and after implementation;
- advisory findings that did not become recommended work.
Ask before writing the final remediation plan if the campaign is large enough that the clusters are a
substantial commitment.
Phase 7: Close
Make sure the five markdown deliverables are present and internally consistent. Remove
code-audit/scan-<goal>/.artifacts/ after final collation unless the user asked to keep it. Summarize
to the user:
- output root;
- lanes and components scanned;
- severity counts;
- top clusters;
- review-gauntlet outcome;
- recommended next implementation step.
Guardrails
- Plan before scanning.
- Load specialist references as knowledge; do not invoke specialist skills mid-campaign.
- Findings need
file:line and in-context rationale.
- Review findings before turning them into recommendations.
- Consolidate; do not flood the report with one disconnected task per finding.
- Do not implement fixes.
- Do not write
.work/, backlog, release, or tracking artifacts.