Qedgen Auditor

QEDGen Auditor

Audit Solana programs with explicit target resolution, bounded execution, and evidence separate from impact. Never present an unverified pattern match as a confirmed vulnerability.

Required workflow

1. Resolve the target and run preflight

Require a concrete program root. If the repository contains multiple plausible programs and the user did not select one, ask which program to audit.

Resolve paths relative to this skill directory (<skill-root>), not the user's current working directory. Run:

<skill-root>/scripts/preflight.sh --root <program-root> [--spec <path>] [--qedgen <binary>]

Read runtime detection if detection is ambiguous or the target contains assembly. Report the resolved root, runtime, mode, spec, QEDGen version, and available repro lanes before making findings.

Run compilation separately and retain its exit status. Do not say the target compiles until the command has completed successfully:

<skill-root>/scripts/preflight.sh --root <program-root> [--spec <path>] --compile

Missing or stale QEDGen reduces the audit to read-only; it does not prevent source review. A failed program build prevents fired Mollusk reproducers but does not erase source-established evidence.

Record each lane independently: source review, ordinary probe, compilation, Mollusk, Miri, and Crucible. Failure or an empty result in one lane must not cancel invariant extraction or be reported as evidence that no bugs exist.

2. Select an audit profile

Use standard unless the user requested a quick look or high assurance. Select the model by capability rather than venue name. Read model selection before dispatching an independent audit worker or benchmark run.

Do not start an unbounded convergence loop. Parallel or independent workers are an optional venue capability, never a requirement. Use them only when the venue permits them and the user or governing instructions authorize delegation. See orchestration profiles.

3. Build the work list

Spec-aware:

qedgen probe --spec <path>

Spec-less:

qedgen probe --bootstrap --root <program-root>

Probe output prioritizes investigation; it is not a verdict. Independently walk the program's handlers, authority graph, state transitions, account identities, arithmetic, CPIs, and documented invariants.

Read the probe envelope by evidence tier (schema v3, version: 3):

• findings[] — issues backed by a replayable reproducer. Treat as confirmed subject to your own review.
• candidates[] — predicate hits and static patterns that warrant investigation but carry no severity and no reproducer. These are a work list, never results: each is a lead to confirm or dismiss by reading the impl, and each reason says why it isn't yet a finding (usually "no reproducer constructor yet"). Never file a candidate as a finding without independently confirming it and attaching a reproducer.
• engine_runs[] — per-engine passed | partial | blocked | failed | skipped. A partial run lists skipped_files it could not read; a blocked run (e.g. budget-0 fuzz) did not execute. An empty findings[] next to a partial/blocked engine is weak evidence, not a clean pass.
• outcome — passed_with_coverage | no_findings_low_coverage | blocked_incomplete_harness | engine_failed | dry_run. Only passed_with_coverage licenses "the probe found nothing here"; the rest mean the probe under-ran and you must lean on manual review.

Always extract a domain dossier from source, tests, comments, documentation, and paired operations, even when QEDGen is missing or stale, the ordinary probe fails or returns no sites, the runtime adapter is unsupported, or compilation fails. Capture asset flows, quantities and units, lifecycle transitions, authority capabilities, candidate economic equations, external assumptions, source anchors, and confidence. Use the dossier for manual review and intent ratification immediately; preserve it for spec drafting and later executable verification when blocked lanes recover. Read domain invariant extraction before interviewing the user or drafting domain properties.

Write both .qed/audit/<timestamp>/domain-dossier.json and its human-readable .md rendering. Validate the JSON against <skill-root>/schemas/domain-dossier.schema.json; use its stable candidate IDs in interviews, specs, findings, and fuzz results. Also maintain .qed/audit/<timestamp>/run-manifest.json against <skill-root>/schemas/audit-run-manifest.schema.json. Update each lane as it starts or finishes; blocked lanes require a concrete reason and exact resume command. After ratification, require spec-handoff.json against <skill-root>/schemas/spec-handoff.schema.json. It separates structural, domain, and regression layers, links clauses to stable provenance IDs, and records language gaps instead of flattening unsupported domain semantics into comments. For each needs_authoring domain clause, follow its authoring.constructs, parser-shaped authoring.template, and authoring.notes; each language gap states current_language_support so finite sums, nominal dimensions, floor/ceiling/half-up rounding, typed external fields, lifecycle transitions, and authority guards are not mistaken for missing syntax. Also require domain-sequences.json against <skill-root>/schemas/domain-sequences.schema.json. Use its ratified setup/forward/reverse/teardown plans as stateful coverage targets, but do not claim exact-sequence coverage while any unresolved_parameters remain or the runner has not replayed that plan. Validate the artifacts with:

For exact replay, author domain-sequence-bindings.json against domain-sequence-bindings.schema.json. Every plan/action/parameter key must be explicit; never infer accounts or argument values. Domain-mode Crucible writes resolved-domain-sequences.json, validates it against resolved-domain-sequences.schema.json, replays each byte-exact seed before exploratory fuzzing, and feeds the same corpus into the subsequent run. Each exact replay appends durable evidence at <harness>/.qedgen/domain-replay-report.json, validated against domain-replay-report.schema.json; the report pins the resolved document, account overlay, generated harness, seed content, native replay command, exit status, and plan ID. A seed merely present in the exploratory corpus is not exact-sequence coverage. Account targets use the explicit fixture:<account> namespace. QEDGen verifies them against non-PDA, non-default spec accounts, collapses them into account-binding-overlay.json, rejects conflicts across plans, and compiles that overlay into action account literals and signer selection before encoding handler arguments. The collapse step also rejects same-action fixture aliasing and fixture targets whose signer, writable, program, account-type, authority/owner, or imported namespace constraints cannot satisfy the source account. Default-address and PDA accounts remain generator-managed and must not appear in the overlay. For PDAs, QEDGen retains Anchor IDL seed tuples and derives each action's address from literal, account, numeric argument, and supported state-field seeds. Dependent PDAs are ordered before their consumers; unresolved or cyclic derivations stop codegen instead of falling back to an empty-seed address. Non-init PDAs are materialized on demand, while init targets remain absent for the instruction to create.

<skill-root>/scripts/check-domain-artifacts.sh \
  --dossier .qed/audit/<timestamp>/domain-dossier.json \
  --manifest .qed/audit/<timestamp>/run-manifest.json \
  --handoff .qed/audit/<timestamp>/spec-handoff.json \
  --sequences .qed/audit/<timestamp>/domain-sequences.json \
  --bindings .qed/audit/<timestamp>/domain-sequence-bindings.json \
  --resolved-sequences .qed/audit/<timestamp>/resolved-domain-sequences.json \
  --account-overlay .qed/audit/<timestamp>/account-binding-overlay.json \
  --replay-report fuzz/<program>/.qedgen/domain-replay-report.json

Read only the relevant detailed references:

• category catalog: per-category predicates, runtime-specific patterns, and the composition cookbook.
• manual review passes: the detailed investigation workflow when the compact steps here need expansion.
• report and grading: classification rules, the severity grading procedure, and the report format.
• audit handbook: adversarial mindset, tool surface, and the reproducer-only contract.
• known non-findings: always before filing a reentrancy, closed-account-discriminator, float-determinism, token self-transfer, instruction-introspection, partial-state-commitment, or unchecked-CPI-return finding.
• trust-surface primitives: only when a small dependency supplies a security-critical primitive.
• data-structure dependency invariants: only when a niche collection or zero-copy dependency controls fund movement.
• domain invariant extraction: always for spec-less audits and whenever the existing spec is partial or only structural.
• probe orchestration: only when using Mollusk, Miri, or Crucible repro lanes.
• model selection: before dispatching audit, benchmark, or reconciliation workers.

Framework wrappers are evidence, not syntax to ignore. Confirm what the active framework and version enforce before filing missing signer, owner, rent, discriminator, initialization, or realloc findings.

3a. Ratify intent and schedule fuzzing

Do not ask the user to rediscover facts already visible in code. The intent interview's questions are the probe envelope's hypotheses[] records rendered conversationally — ask in-harness, in the conversation, one answer per hypothesis (accept / reject / bug), each presented with its claim, evidence anchors, and payoff — plus agent-derived candidates (domain invariants, lifecycle, authority capabilities, threats, intentional exceptions) the binary abstained on. Time it after the first MED+ finding, or after the autonomous pass finishes dry or probe-blocked; when there is no finding to win trust with, the ranked hypotheses themselves become the trust-winning first deliverable. Write the answers to answers.json in the audit dir and run qedgen ratify --audit-dir <dir> to apply confirmed hypotheses as executable spec clauses. Keep unratified semantic claims as hypotheses. An in-harness confirmed invariant is checking until a backend runs; only impl-bound evidence supports #[qed(verified)].

Phase numbers refer to the orchestration in manual review passes: Phase 1 autonomous discovery, Phase 2 intent interview, Phase 3 refined second wave.

Crucible has three bounded entry points:

1. Start spec-less protocol mode during Phase 1 as soon as supported runtime metadata is available. Treat it as mechanical state-diff coverage only.
2. Start skeleton mode during Phase 1 after auto-ratifying only literal, source-anchored high-confidence clauses. Do not wait for a finding.
3. Re-run domain mode after user ratification against the enriched spec, with cross-handler sequences and domain properties enabled.

If QEDGen, compilation, runtime metadata, or a fuzz harness is unavailable, mark that entry point blocked and continue with source review and agent-authored tests where possible. Never weaken or reject an invariant because Crucible could not run. See probe orchestration.

4. Classify evidence before severity

Assign exactly one evidence state:

• confirmed: an executable reproducer fired against the audited program.
• structural: source establishes the vulnerable and reachable path, but the execution environment could not run it.
• hypothesis: intent, reachability, or a required precondition is unresolved.
• rejected: disproved, framework-protected, unreachable, or suppressed. Name the specific guarantee that rejects it; see known non-findings.

Then grade impact using severity and evidence. Hypotheses are not vulnerability findings and must appear in a separate open-questions section, if useful at all.

For every candidate, record:

1. Attacker capability and required preconditions.
2. Exact source path and line.
3. Framework guarantees already in force.
4. Standalone impact.
5. Reachable compositions with other primitives.
6. Evidence state and reproducer status.

5. Gate HIGH and CRITICAL findings with reproducers

Write Mollusk reproducers under:

target/qedgen-repros/audit/<finding-id>.rs

Run:

qedgen verify --probe-repros --json

Apply these outcomes:

• Fired: evidence is confirmed; surface the finding.
• Build or simulator limitation: retain only if source establishes reachability; mark structural and name the limitation.
• Did not fire: mark rejected and omit it from vulnerability totals.
• Ambiguous result: mark hypothesis, not structural.

MEDIUM and LOW findings may be structural without a repro, but still require a concrete reachable path. See reproducer contract.

6. Write artifacts and report

Write the full report to .qed/findings/audit-<timestamp>.md. Reproducers stay ephemeral under target/. Write suppression rules only for stable, machine-detectable false positives.

Give the report header a run status: completed, build-blocked (the program did not compile), tooling-blocked (QEDGen, metadata, simulator, or another verification lane was unavailable while source review continued), or policy-interfered (a venue or provider policy stopped, truncated, or redirected the analysis). If an audit worker halts mid-run — refusal, truncation, or silent stop — the run is policy-interfered: report what was covered before the stop, switch to a worker that completes audit workloads (see model selection), and never present a partial pass as complete coverage.

Order the digest:

1. Confirmed CRITICAL/HIGH/MEDIUM findings.
2. Structural findings.
3. Specification gaps.
4. Open hypotheses, clearly excluded from vulnerability totals.
5. Suppressed count and emitted artifacts.

Include the dossier JSON/Markdown and run manifest in the artifact footer. A blocked run must point to the manifest's resume commands.

Every surfaced finding must include:

• severity and evidence state;
• location;
• attacker action and resulting state or fund delta;
• preconditions and reachability;
• standalone impact and any kill-chain;
• minimal implementation fix;
• minimal .qedspec regression guard when applicable;
• repro path and result for HIGH/CRITICAL findings.

When handing findings into specification, author three layers in order: structural contracts, ratified domain properties, then finding-derived regression guards. An audit is not fully specified merely because every finding has a guard. See finding to spec.

Do not edit the audited program unless the user separately asks for fixes. Do not publish third-party HIGH/CRITICAL findings; recommend responsible disclosure.

Reliability rules

• Treat the packaged repository copy of this skill as canonical. Maintainers run scripts/check-auditor-skill.sh when changing it. Installation tooling must receive an explicit venue-owned destination; the skill does not assume a venue-specific home directory.
• Include the skill package version and, when available, its source commit in the audit header so results can be traced to the exact catalog and workflow.
• Never infer spec-aware mode from a literal file named .qedspec; resolve an explicit spec or a unique *.qedspec file.
• Never reject a mixed repository because an unrelated .s file exists.
• Assembly-only source-pattern audits are unsupported. Spec-aware analysis is allowed when the probe operates from the spec without claiming source-level assembly coverage.
• Never claim “clean.” State what was examined, which lanes ran, and remaining coverage limits. Recall is probabilistic: a single empty pass is under-sampling, not evidence of absence.