dlt-auditor

name

dlt-auditor

description

Run DLT Auditor stable audit designs and blind audit suites against target codebases from this repository. Use when the user asks to run, scaffold, resume, configure, or explain dlt-auditor audit execution, including max, optimal, or custom prompt-pack tiers, Codex and Claude Code agent modes, blind-suite integrity rules, worker-limit recovery, design selection, or the runtime-focused dlt-auditor workflow.

DLT Auditor

Use this repository as the runtime for stable audit designs from designs/ against target codebases. Run commands from this skill/repository root.

Do not add or reintroduce learning-loop features, benchmark ground truth, candidate scoring, scorecards, miss analysis, leaderboards, promotion workflows, or refinement systems unless the user explicitly asks to rebuild that separate system.

Blind Integrity

During blind audit execution, keep audit workers separated from answer-key material.

Do not read known findings, benchmark ground truth, scorecards, miss analyses, result records, leaderboards, refinement plans, audit-output snapshots, candidate result archives, prior round folders, sibling suite outputs, or stable designs/*/runs/** output.

Preserve generated suites when worker limits or external interruptions occur. Resume instead of recreating unless the user explicitly asks to replace the suite.

Agent Choice

Use Codex by default:

--agent codex --service-tier standard --reasoning-effort high --deep-reasoning-effort xhigh --deep-phases canonicalize,validations,aggregate,final

Use Claude Code when the user asks for Claude, Claude Code, or --agent claude:

--agent claude

For Claude Code, do not pass Codex reasoning or service-tier overrides. Use --claude-path or --claude-add-dir only when the user supplies those needs or local execution requires them.

Tiers

Treat "prompt packs" and "design packs" as the stable design folders under designs/.

Support these tier invocations:

$dlt-auditor max ...
$dlt-auditor optimal ...
$dlt-auditor custom ...

If the target repository or suite name is missing, infer a reasonable suite name from the target basename, tier, and timestamp when possible. Ask only when the target repository path is missing or ambiguous.

Max

Use every runnable design pack under designs/.

Discover packs with:

bin/run-blind-suite --list-designs

or, if needed:

find designs -mindepth 1 -maxdepth 1 -type d -printf '%f\n' | sort

Then run one blind suite with every discovered pack by repeating --design <design-name>.

Optimal

Analyze the target project and available design packs, choose only the packs that best fit the project, and run them. Select at most 5 packs.

Use this selection process:

1. Inspect target repo metadata such as README*, package manifests, lockfiles, language/framework files, top-level directories, and dependency names.
2. Inspect design pack names and high-level pack files such as designs/<name>/00_protocol_mapper.md, 01_base_hunter.md, 02_validation_and_impact.md, and 05_corpus_pattern_search.md.
3. Do not inspect generated runs or answer-key material while selecting.
4. Prefer packs whose protocol, ecosystem, implementation language, or audit venue matches the target.
5. If fewer than 5 packs clearly fit, run only the clear fits. If none clearly fit, choose the closest generally applicable packs and state that the selection is a best-effort match.

After choosing, immediately run one blind suite with the selected packs by repeating --design <design-name>.

Custom

Use exactly the design packs named by the user. Validate that each named pack exists under designs/; if a name is misspelled, list the available packs and ask for correction.

One Design

Scaffold a single design run:

bin/run-design <design-name> /path/to/target-repo --run-name <run-name> --parallel-jobs 8

Then execute the generated design run with that design's runner.

Codex:

designs/<design-name>/bin/run-parallel-codex designs/<design-name>/runs/<run-name> --jobs 8 --agent codex --service-tier standard --reasoning-effort high --deep-reasoning-effort xhigh --deep-phases canonicalize,validations,aggregate,final

Claude Code:

designs/<design-name>/bin/run-parallel-codex designs/<design-name>/runs/<run-name> --jobs 8 --agent claude

If worker limits are exhausted, preserve the run and resume with --resume if supported by the runner.

Blind Suites

Prefer blind suites when running one or more designs as a benchmark-style audit. Outputs live under runs/<suite-name>/.

List available designs when the user has not specified one:

bin/run-blind-suite --list-designs

Start a Codex blind suite:

bin/run-blind-suite --repo /path/to/target-repo --suite-name <suite-name> --design <design-name> --parallel-jobs 8 --agent codex --service-tier standard --reasoning-effort high --deep-reasoning-effort xhigh --deep-phases canonicalize,validations,aggregate,final

Start a Claude Code blind suite:

bin/run-blind-suite --repo /path/to/target-repo --suite-name <suite-name> --design <design-name> --parallel-jobs 8 --agent claude

Repeat --design <design-name> for multiple designs if requested.

Resume an interrupted or worker-limited suite:

bin/run-blind-suite --suite-name <suite-name> --resume

Use --scaffold-only when the user asks to prepare isolated runs without launching workers. Use --force only when the user explicitly wants to replace an existing suite or run.