ferroxlabs / ijfw-cross-audit

Execution

1. Detect artifact. Accept: diff, file path, brief text, or HEAD~1..HEAD.
   If none provided, ask once: What should I audit? (diff, file, or paste text)

2. Detect auditors. Roster covers six independent training lineages:

   • codex (openai family) -- OpenAI CLI
   • gemini (google family) -- Gemini CLI
   • claude (anthropic family) -- Claude Code (fresh session)
   • deepseek (oss / cn lineage) -- DeepSeek API
   • qwen (oss / cn lineage) -- Qwen Code CLI
   • kimi (oss / cn lineage) -- Moonshot Kimi
   • opencode, aider, copilot (additional oss / openai-family auditors)

   Default selection: diversity strategy picks one openai-family + one
   google-family + caller, excluding the caller's own family. Reachability =
   CLI on PATH OR API key in env. The Donahoe Trident principle says: at least
   three lenses, three lineages.

3. Override roster with --with. Pass --with <id>[,<id>] to pin the
   exact auditors fired. Self-audit is rejected (same-lineage = single source).
   Example: cross-audit --with codex,gemini,deepseek <target>.

4. Dispatch in parallel. Every lens runs concurrently; the orchestrator
   short-circuits stragglers once minResponses productive results land. Per-
   provider timeouts (codex 120s, gemini 90s, default 90s) plus
   IJFW_AUDIT_TIMEOUT_SEC override.

5. Reconcile. Lexical-signature clustering (issue + location):

   • CONSENSUS -- flagged by ≥2 lenses (consensus: true, consensusCount, consensusLenses)
   • CONTESTED / single-lens -- flagged by exactly 1 lens
   • PASS -- no findings

   Output ranks CONSENSUS findings first, then single-lens findings, each
   ordered by canonical severity (critical → low).

6. Emit report using the report format rule below.

Budget controls

• IJFW_AUDIT_BUDGET_USD (default $2.00) -- per-session aggregate cap.
  Post-flight accumulation; first call is always allowed, subsequent calls are
  rejected when accumulated + estimate > budget. Raise to continue.

• IJFW_AUDIT_BUDGET_USD_PER_LENS -- granular cap on a single lens's spend
  across one convergence cycle. Useful for "spend at most $0.50 per gemini
  call" guardrails on long-running converge loops.

• IJFW_AUDIT_TIMEOUT_SEC -- per-auditor budget (clamped to [1, 3600]).
• IJFW_AUDIT_CONCURRENCY -- parallel-fire cap (default 3).

Report format rule (all cross-audit outputs)

Any reconciliation report presenting multiple paths MUST use this structure:

VERDICT
  <one-line converged recommendation>

OPTIONS
  A -- <short-name>: <one-line what it is>
  B -- <short-name>: <one-line what it is>

REVIEWER CONVERGENCE
  <reviewer>:  <letter>  <score>  "<their call>"  => Option <letter> -- <name>

RECOMMENDATION
  Option <X> -- <name> because <one-sentence why>.

NEXT ACTION
  <exact command or step>

Never write "Option A" or "Option B" without "-- name" immediately after it on the same line.

Severity / disposition vocabulary

Two axes, never mixed:

• Finding severity (sortable, per-item): critical | high | medium | low.
• Audit disposition (status of the whole audit): PASS | CONDITIONAL | WARN | FLAG | FAIL.

The merger coerces stray disposition values (warn, flag, etc.) on a per-
finding severity field into the canonical severity scale (warn→medium,
flag→high, fail→critical) so nothing sinks to the unranked bottom.

Output contract

Emit a gate-result block as the LAST content of your output. Nothing
after it. Use gate="cross-audit". Statuses: PASS | CONDITIONAL | WARN | FLAG | FAIL.

Format:

{
  "schema_version": "1.0",
  "gate": "cross-audit",
  "status": "<STATUS>",
  "project_type": "<from project-type-detector>",
  "lenses": [],
  "affected_artifacts": [],
  "accounting": {"duration_ms": 0, "lenses_invoked": 0, "cost_usd": null},
  "remediation": [],
  "receipts_ref": null,
  "supersedes": null,
  "gate_id": "<gate-with-colons-replaced-by-dashes>-<ts>-<rand4>",
  "emitted_at": "<ISO-8601>"
}