dummyindex-audit

name

dummyindex-audit

description

Run an on-demand adversarial audit over real source from a free-text request. Scaffolds `.context/audits/{slug}/`, selects a task-dependent panel from the persona catalog rather than a fixed roster, collects independent evidence, then makes auditors read and challenge one another through up to three rebuttal rounds with early convergence. Synthesis resolves verdicts, preserves unresolved disputes, deduplicates findings, and ranks them by severity in `report.md`. The CLI only manages deterministic workspace and resumption state; the active host orchestrates judgment. A full context index is optional. Use for `/dummyindex-audit`, `$dummyindex-audit`, audit X, argue and audit, security or correctness review, or have agents critique X and report findings.

/dummyindex-audit / $dummyindex-audit — argue-and-audit panel

Installed from dummyindex __VERSION__. Run dummyindex --version to confirm the CLI matches. If they diverge, diagnose with dummyindex context check --versions (it reports which layer is stale), then run /dummyindex-update on Claude or $dummyindex-update on Codex to bring the CLI, skills, and this repo's wiring back into sync — the update skill is non-destructive on a curated .context/. Don't reach for a blunt dummyindex install to "fix" a version skew.

You are the audit conductor. The user hands you a free-text description of what to audit. The dummyindex context audit CLI is deterministic plumbing — it scaffolds the workspace, emits the persona catalog, and keeps the debate resumption log. It never runs an agent, never picks the panel, and never decides convergence. That judgment is yours.

The point of this skill is adversarial review: not one pass of parallel opinions, but a panel that argues — auditors read each other's findings and concede, dispute, or defend until they reach agreement (or the round cap). What survives the debate is what you report.

Inputs

• A description of what to audit (required). If the user didn't give one, ask for it in one line.
• An optional scope — paths to focus on (--scope dummyindex/cli --scope dummyindex/context/domains/audit). Without it the audit considers the whole repo; always prefer a scope when the description implies one.
• A host-valid model selection. On Claude, reuse the configured Claude model or ask for opus-4.8|sonnet-4.6|haiku-4.5; never silently choose one. On Codex, always pass --model current so the audit uses the running Codex model. Do not offer Claude labels, and do not inherit one from a shared repo config.

The loop (run it literally)

0. Scaffold the workspace + read the catalog

Claude Code:

dummyindex context audit start --describe "<the user's request>" \
  [--scope <path>]... [--mode light|standard|deep] \
  [--model opus-4.8|sonnet-4.6|haiku-4.5] --json

Codex:

dummyindex context audit start --describe "<the user's request>" \
  [--scope <path>]... [--mode light|standard|deep] --model current --json

This creates .context/audits/<slug>/ (audit.json, description.md, catalog.json, findings/) and prints a JSON object: {slug, dir, mode, model, max_rounds, scope, catalog:[...]}. Read it. The catalog is your menu of available auditors — each entry has persona_id, role, subagent_type (the real Task-tool agent to launch), triggers, and description.

If the CLI exits with a "model is required" error, ask for a Claude model on Claude Code. On Codex, re-run with --model current; do not turn it into a model choice question.

1. Pick the panel — task-dependent

This is the step that makes the audit fit the task. Read the description and the catalog, then choose the auditors whose role/triggers/description actually bear on what was asked. Examples:

• "audit the auth flow for security holes" → security, correctness, maybe tests.
• "is this cache layer correct and fast?" → correctness, performance, architecture.
• "review this migration" → data-integrity, correctness, security.

Pick 2–5 auditors. State your panel and a one-line rationale to the user before dispatching. Don't run the whole catalog by reflex — an irrelevant auditor adds noise, not signal. If the description is broad ("audit this module"), a wider panel is fine.

2. Round 0 — independent findings (parallel)

Dispatch the chosen panel as parallel subagents through the active host. For each auditor:

• On Claude, set subagent_type from the persona card and fall back to general-purpose if unavailable. On the portable host path (e.g. Codex), use built-in worker because each auditor must persist a findings artifact, falling back to default. The audit remains read-only with respect to source code; writing its report workspace is expected. Never try to dispatch the Claude type name from the portable host path.
• Inline the persona's mandate into the prompt — don't hand the subagent a path. Each auditor's real instructions live in its markdown body at agents/<persona_id>.md (a companion file next to this SKILL.md), not in catalog.json (which carries only the one-line description, for your selection). You — the conductor — are running from the skill directory, so read agents/<persona_id>.md yourself and paste its body into the delegated prompt verbatim. A fresh subagent cannot resolve the skill path, so "tell it to read its persona file" will silently fail — inline it.
• Then add to the prompt: the description, the scope paths, an instruction to ground in .context/conventions/* and any relevant feature docs if they exist, to read the real source (not docs), and to write its findings to .context/audits/<slug>/findings/<persona_id>.md using the finding contract below.
• Log start/finish:

  dummyindex context audit-log --slug <slug> --round 0 --persona <persona_id> --status started
  dummyindex context audit-log --slug <slug> --round 0 --persona <persona_id> --status complete
  

Finding contract (every auditor writes bullets, no essays):

## <persona_id> findings

- `path:Lstart-Lend` — **severity** (critical|high|medium|low|info) — claim in one sentence — evidence (what in the code shows it) — suggested fix (or "none").

3. Rebuttal rounds 1..N — argue to agreement (max_rounds, default 3)

light mode skips this step entirely — go to synthesis. For standard/deep, loop:

For round r (starting at 1, up to max_rounds from the JSON):

• Consolidate the current findings (all findings/*.md).
• Re-dispatch the same panel as parallel host-native subagents — again inlining each persona's body (from agents/<persona_id>.md) plus the consolidated findings into the prompt. Each auditor re-reads the consolidated findings — its own and its peers' — and, for each finding it has a view on, marks one of: concur, dispute, defend, or concede. It updates the finding's status (open → confirmed | disputed | refuted | withdrawn) and appends its rebuttal note inline. Log each persona for round r.
• Check convergence. Compare this round's finding statuses to the previous round's. If no status changed — the panel agrees — STOP. Do not run further rounds. Early agreement is the goal, not a full three rounds.
• Never exceed max_rounds. If you reach the cap with findings still disputed, that's fine — carry the open disagreement into synthesis and report it as unresolved.

deep mode only: before synthesis, run one adversarial refutation pass — for each still-confirmed finding, dispatch a skeptic (prompted to refute it) and downgrade any finding a majority of skeptics refute. Keep this lean; don't turn it into another full debate.

4. Synthesis — resolve, dedupe, rank → report.md

Write .context/audits/<slug>/report.md yourself (or via one Claude Software Architect / Codex default subagent with the synthesis mandate inlined). It must:

• Assign each surviving finding a verdict: confirmed (survived the debate), disputed (unresolved disagreement — show both sides), or refuted/withdrawn (dropped, listed briefly so the reader knows it was considered).
• Dedupe findings different auditors raised about the same path:range.
• Rank by severity (critical → info), then by verdict (confirmed first).
• Open with a 3–5 line executive summary: how many confirmed, the headline risks, and whether anything stayed unresolved.

5. Report to the user

Summarize in chat: the panel you ran, how many rounds it took to converge (or that it hit the cap), the confirmed findings by severity, any unresolved disputes, and the path to report.md.

Discipline (non-negotiable)

• Task-dependent panel. Pick auditors that fit the description. Justify the panel in one line. Don't run the whole catalog by default.
• They must actually argue. Round 0 is opinions; the value is in the rebuttals. An auditor that only re-states its finding in round 1 hasn't engaged — prompt it to read peers and take a position.
• Evidence over assertion. Every finding and every rebuttal cites a path:range. "Could be vulnerable" without an attack path, or "this is slow" without a hot path, is noise — drop it.
• Stop on agreement. The cap is 3; the target is consensus. The moment a round changes no status, you're done debating.
• The code wins. Conventions and feature docs are context, not authority — when a doc and the code disagree, cite the code.
• Read-only by default. An audit reports findings; it does not fix them. If the user wants fixes, that is a separate /dummyindex-plan → /dummyindex-build cycle on Claude or $dummyindex-plan → $dummyindex-build on Codex, seeded from the report.

CLI reference (deterministic state only)

dummyindex context audit start --describe "<text>" [--scope PATH]... \
    [--mode light|standard|deep] [--model current|opus-4.8|sonnet-4.6|haiku-4.5] \
    [--slug S] [--force] [--root DIR] [--json]
    → scaffold .context/audits/<slug>/ (audit.json, description.md, catalog.json,
      findings/) and emit {slug, dir, mode, model, max_rounds, scope, catalog:[...]}.
      Model is REQUIRED unless .context/config.json provides one. Codex passes
      current explicitly; Claude uses a configured or user-selected Claude model.
dummyindex context audit show --slug S [--root DIR] [--json]
    → current state: config + which rounds are complete + report path (if written).
dummyindex context audit-log --slug S --round N --persona P --status STATE [--note "..."] [--root DIR]
    → append a debate-log row. STATE: started|complete|failed|skipped. Drives resumption.