Skip to content

Latest commit

 

History

History
897 lines (775 loc) · 45.2 KB

CHANGELOG.adoc

File metadata and controls

897 lines (775 loc) · 45.2 KB

Changelog

[Unreleased]

Added — the classification registry reads a deed first

  • load_user_classifications now reads <project_root>/audits/assail-classifications_praxis.deed first: a DEED praxis-deed with :profile assail-classifications and one (classification :file "…" :category <Symbol> :audit "…" :rationale "…") clause per audited finding. When the deed exists, the legacy audits/assail-classifications.a2ml and .panic-attack-classifications.a2ml are not read; otherwise they are, with their reader unchanged. Suppression semantics are identical: a finding is suppressed when its location and category match an entry. The name ends _praxis.deed because the standards deed grammar (deed.abnf v2.0.0) dispatches a deed’s head on its file name, and a bare assail-classifications.deed matches no pattern.

  • New module panic_attack::deed: a minimal reader of the deed grammar (header, form, fields, clauses, every value form, filename dispatch, the :schema-version and identity-field side conditions), with line-numbered errors.

Changed — a malformed deed registry is an error

  • load_user_classifications returns Result<Vec<UserClassification>> and apply_user_classifications returns Result<()> (both were infallible). A deed registry that breaks the grammar or the profile (unknown field or clause, a string :category, an unknown category, a missing :file, the wrong head, trailing content, over 4 MiB, not UTF-8) is an Err naming the file and line, and analyze() / analyze_verbose() fail with it instead of reading it as an empty registry or falling back to the A2ML files. Library callers must handle the Result.

Changed — the AI manifest reads a deed first; this repo is off A2ML

  • Manifest::load_default, panic-attack manifest and panic-attack diagnostics read 0-AI-MANIFEST.deed first (a pointer-only repo-deed, parsed with panic_attack::deed), then the legacy 0-AI-MANIFEST.a2ml or AI.a2ml as a read-only fallback. A deed manifest carries no report settings, so it gets the built-in defaults (formats json and nickel, storage filesystem); a legacy manifest’s (reports …) is still honoured. A malformed or non-repo-deed deed manifest is an error. New: MANIFEST_CANDIDATES, Manifest::default_path, Manifest::path, Manifest::is_deed.

  • diagnostics looks for panicbot directives as the (bot-directives (bot :name panicbot …)) clause of the repo deed, not at .machine_readable/bot_directives/panicbot.scm; none declared is reported OK (panicbot runs the full suite) rather than as a warning.

  • The repo’s own .a2ml files are gone (25 of 26), translated where a standards mapping spec covers them into panic-attack_chora.deed, .machine_readable/descriptiles/STATE.deed and 0-AI-MANIFEST.deed. docs/history/A2ML-REMOVED.adoc records each file and what was dropped. Adjustfile.a2ml stays while just adjust-check reads it.

[2.6.0] - 2026-11-14

Release of the work merged on main since 2.5.5 (143 commits), including PR #292 and PR #293. Contract change: panic-attack adjudicate now exits 1 for a complete fail (previously 0) and 2 for incomplete input; see the first entry below.

Fixed — adjudicate no longer passes when it processed nothing

  • panic-attack adjudicate printed verdict: pass (processed 0, failed 1) and exited 0 when a report was missing or unrecognised. A run that processed no report, or failed to load any report, now gets the new verdict error (JSON verdict; each failure is in notes and echoed to stderr) and exits with status 2, the same code as assail --fail-on-incomplete. Verdict precedence is fail > error > warn > pass: a real fail from a loaded report is still reported as fail. Exit contract: 2 for incomplete input (zero processed or any report failed to load), whatever the verdict, winning over everything; 1 for a complete fail (new: previously exit 0); 0 for a complete pass or warn. The Chapel MassPanic adjudicate pass now reads exit 1 as a valid fail verdict and exit 2 as error.

Fixed — the shell /tmp/ rule ignores comments and reports a line

  • analyze_shell raised "Hardcoded /tmp/ path without mktemp" on a whole-file substring match, so a /tmp/ mentioned only in a # comment fired, and the word mktemp in a comment silenced a real /tmp/ use. Both tests now look only at code: full-line comments and trailing + ` comments outside quotes are skipped. A `` starts a comment only at an unquoted word boundary (line start, after whitespace, or after an operator such as `;+), so # inside quotes (including ANSI-C $'...\'...'), $#, ${#var}, after an escaped space and in heredoc bodies is kept as code, and any doubt still fires the rule. Here-strings (<<<) are not heredocs, and heredoc delimiters are parsed whole ('---', EOF-TEXT). The finding carries the first offending line number; findings from embedded scripts carry none, because their lines are relative to the extracted block. Word state carries across line breaks inside a quote and over a backslash-newline continuation, so "a<newline>b"# /tmp/x is not mistaken for a comment.

Fixed — readiness tests find the binary under Cargo’s split build directory

  • tests/readiness.rs derived the binary path from current_exe, which fails when Cargo places test binaries under build/<pkg>/<hash>/out/ (15 of 21 tests failed in CI). It now uses CARGO_BIN_EXE_panic-attack.

Fixed — the Idris2 ABI proofs are type-checked, and the stripping proof is re-proved

  • No CI job had ever type-checked src/abi/, and as committed none of its five modules checked. A new idris-abi job (proven-tests.yml, using scripts/ci/check-idris-abi.sh) now checks all five on every PR with Idris2 0.7.0. It fails on a type error, on a missing %default total, or on believe_me, assert_total, assert_smaller, postulate or idris_crash.

  • Stripping.idr: the 2026-06-02 proof of stripLineCommentsIdempotent never type-checked, so it is re-proved through two facts: stripping is the identity on text with no //, and the strippers' output has no //. The strippers are restated with explicit Bool tests. An exhaustive run over 29,524 inputs shows they behave exactly as before.

  • Types.idr: adds the missing Eq Severity that Ord needs, and fixes the LangId declaration syntax.

  • PROOF-NEEDS.adoc and EXPLAINME.adoc now say that the earlier claims had no receipt.

Added — assail --output-format kyaml writes a KYAML copy

  • With --output-format kyaml and --output report.json, assail writes report.kyaml beside the JSON report. The JSON stays the report of record: it is what --attest seals. --output must end in .json (any letter case), and no two of the files written (report, copy, and the --attest sidecar) may be the same file: a symbolic link at any of their paths, or two existing paths that are hard links to one file, is refused before anything is written. So is a path that exists but is not a regular file (a directory, say) or that cannot be looked up, such as a derived name over the filesystem’s length limit. Each output is also opened for writing before anything is written (an absent one is created and removed again), so a read-only copy or sidecar, or a directory the caller cannot create files in, is refused too, instead of failing after the report has been replaced. These checks run before the writes, so a link swapped in between a check and its write is not caught, and a crash can leave a partial file: the planned generation-directory publish is meant to close both. Other formats are unchanged for assail.

Fixed — KYAML keys holding a line separator

  • KYAML output now escapes U+2028 (LS) and U+2029 (PS) as \uXXXX, as it already did for NEL (U+0085). go-yaml, the parser behind yq, reads all three as line breaks, so a raw one in a key made the document unparseable by yq and by standards' kyaml-format.sh. crate::yaml (serde-saphyr) read these files correctly before and after.

Changed — attestation hashes are reproducible JCS

  • assail --attest now hashes the intent and evidence records as their JCS (RFC 8785) bytes instead of serde_json struct order, and writes the .attestation.json sidecar itself as I-JSON/JCS. Anyone holding the sidecar can recompute both hashes in any language, for example with jq -cjS .attestation.intent | sha256sum.

  • The seal gains a hash_scheme field (jcs-sha256). The scheme is prefixed to the chain hash, so it is covered by any Ed25519 signature: relabelling a seal as legacy to skip the content check breaks the chain hash.

  • attest verify now rehashes the embedded intent and evidence and fails on a mismatch. Before, it checked only that the chain hash matched the three stored digests, so the intent and evidence could be edited after sealing without detection. Sidecars written before this change (no hash_scheme) still verify, reported as not rehashed (legacy hash scheme).

  • attest verify rejects any digest field that is not exactly 64 lowercase hex characters. Fixed widths keep the legacy and scheme-prefixed chain inputs (192 and 202 bytes) apart; otherwise moving the scheme name into intent_hash would reproduce the signed bytes on the legacy path.

Added — KYAML report output

  • --output-format kyaml writes reports in KYAML, the estate YAML dialect (crate::kyaml). The output is a fixed point of yq -p yaml -o kyaml (v4.54.1), so it passes standards' scripts/kyaml-format.sh --check. On the reports checked it is also byte-identical to yq -p json -o kyaml run on the JCS report. (Reading JSON, yq reformats some floats, 1e-7 to 1e-07; the KYAML report keeps the JCS form, which yq preserves.)

  • Keys and numbers are in JCS order and form, so the KYAML and JSON reports of a scan hold the same data.

  • One deliberate difference from yq: the YAML non-printables U+007F to U+009F, U+FFFE and U+FFFF are escaped as \uXXXX. yq writes them raw, and U+0085 (NEL) then reads back as a line break. kyaml-format.sh refuses such a file rather than judging it, because `yq’s own rewrite would change the data.

  • README.adoc now lists the real --output-format values; it named markdown and text, which never existed.

Added — .kyaml input (estate YAML policy D280/D307)

  • Attack profiles, ambush timelines, amuck specs, diff reports and a2ml bundles now accept the .kyaml extension alongside .yaml and .yml. KYAML is a strict subset of YAML, so the same parser reads it. The extension list lives in one place, crate::yaml::EXTENSIONS.

  • assemblyline fingerprints .kyaml files, so editing one triggers a rescan.

  • The example attack profiles are in estate formats: profiles/attack-profile.example.kyaml (KYAML, renamed from .yaml) and profiles/attack-profile.example.json (I-JSON/JCS). A test keeps the two in agreement.

Fixed — analysis of files whose path is not valid UTF-8

  • Language detection now consults Path::extension() rather than a lossy to_str() conversion (Language::detect_path; Language::detect delegates to it). A file such as invalid-\xFF.php has an ASCII extension even when its stem is not UTF-8, so it is now analysed instead of being filed away as unsupported — or, worse, excluded, which is not a coverage gap and so let a scan report complete while skipping source it had been asked to scan. invalid-\xFF.rs was affected the same way.

Fixed — a bound on reading hexad ledgers

  • load_hexad_dir no longer allocates from the file’s own size. HEXAD_FILE_READ_LIMIT (64 MiB — two orders of magnitude above any real ledger) is checked from metadata before opening, and the read itself is capped, so a malformed or hostile file in the VeriSimDB directory costs a bounded amount of memory and is reported on stderr instead of being consumed whole.

Changed — MSRV 1.89 → 1.90

  • The exact pin on tree-sitter 0.27 sets the floor: it declares rust-version = "1.90". The msrv CI job builds with exactly the declared version on every pull request, so this cannot drift silently again.

Fixed — YAML parser replaced with safe Rust (RUSTSEC-2024-0449)

  • serde_yaml → serde-saphyr 2.0. serde_yaml 0.9 is deprecated and pulled in unsafe-libyaml, an archived c2rust translation of libyaml flagged unmaintained by RUSTSEC-2024-0449. serde-saphyr is #![forbid(unsafe_code)]. All YAML now goes through src/yaml.rs: hand-written configs (attack profiles, timelines, amuck specs) keep the parser’s default resource budget; machine-written reports get a node/event budget scaled to input length, so large-scan reports still load while alias expansion stays bounded.

  • MSRV raised from 1.85.0 to 1.89. No maintained safe-Rust serde YAML crate builds on 1.85 (serde-saphyr needs 1.89, noyalib 1.86+), and the committed lockfile already required 1.88.

  • YAML output is equivalent but not byte-identical. Quoting switches from '…' to "…" and multi-line strings use | blocks. Reports written by earlier versions still load. A bare ~/null in a string field is now rejected (YAML 1.2 null); quote it.

Fixed — assail detector precision (false-positive reduction, 2026-06-24)

Three assail analyzer fixes, all conservative (no new false negatives), found while triaging hyperpolymath/proven#68 and JoshuaJewell/paint-type#86:

  • UncheckedAllocation © is now NULL-check aware. The detector previously flagged every malloc(...) and emitted a line-less, file-level finding. It now scans per line, skips a malloc whose result is NULL-checked within a short window (if (p == NULL), if (!p), nullptr), and attaches a line number — which also lets an inline // panic-attack: accepted marker suppress a reviewed site (marker suppression is line-gated). A genuinely-unchecked malloc still fires. This is why a real null-check fix (proven stubs.c) previously failed to clear.

  • DynamicCodeExecution (JS/Python) is word-boundary aware. contains("eval(") matched FFI symbol names like proven_calculator_eval(. Now \beval\s*\( (and \b(?:eval|exec)\s*\( for Python); a genuine eval( still fires.

  • CommandInjection (Shell) no longer matches the --eval CLI flag. contains("eval ") matched --eval/-eval. Now the eval builtin is matched only in statement position ((?m)(?:^|[\s;&|(])eval[ \t]).

Verified end-to-end: proven 1→0 active Critical/High (stubs.c clears), paint-type 36→35 (gossamer --eval benchmark FP clears; genuinely-unsafe vendored FFI + the irreducible believe_me axiom correctly remain). 4 new tests in tests/analyzer_tests.rs; full analyzer suite green; zero warnings. PR #134. Refs #32.

Added — attestation unforgeability proof (Idris2, PROOF-PROGRAMME §3.2)

  • src/abi/AttestationUnforgeability.idr: Idris2 proof that the intent→evidence→seal attestation chain is unforgeable. Models chain_hash = H(intent‖evidence‖report) + the Ed25519 signature with the cryptographic facts (chain-hash collision-resistance, Ed25519 EUF-CMA message- and signer-binding, signature correctness) as a parameters block — hypotheses, not postulate (PA021 bans escape hatches), so it is an honest conditional theorem. Under %default total it Qed-closes integrity (tampering any phase invalidates the seal), authenticity (a verifying seal comes from the matching key), and nonRepudiation (a genuine seal verifies), plus two corollaries. Typechecks under Idris2 0.8.0. Closes #123.

Added — contractile registry (INDEX.a2ml)

  • .machine_readable/contractiles/INDEX.a2ml: the previously-missing contractile registry, modelled on echidna’s canonical INDEX. Catalogues all six verbs (must / trust / intend / adjust / bust / dust) with their actual current locations across the three pre-consolidation trees, flags the duplicate trust Trustfile, and records the canonical trident target. The physical consolidation of the three trees stays in #124 — it couples to the contractile gen-just generator (which reads the root contractiles/ tree) and needs the standards CONTRACTILE-SPEC to do safely.

Added — assay / assimilate / aggregate proof-integration subcommands

Three new a-themed subcommands that wire panic-attack into the PROOF-PROGRAMME loop (survey → swap → fold-in-proofs):

  • panic-attack assay [TARGET] [--proven DIR]… (src/assay/mod.rs): surveys a target for code that has a formally proven drop-in equivalent in a proven / proven-servers library and reports each candidate with the proof artifact that backs it — operationalising the “Proven cross-fit” table in PROOF-PROGRAMME.md mechanically instead of by hand. Built-in catalogue: SafePath (canonicalize/unwrap pattern) and SafeUrl (VERISIMDB_URL). On this repo: safe-path Offered (port present in src/safe_path.rs, call sites still to rewire), safe-url NoReplacementSource (not yet ported).

  • panic-attack assimilate [TARGET] --candidate ID [--proven DIR] [--from FILE] [--all] [--dry-run]: performs a swap — stages the proven module into the tree, backs up the original (*.orig), and writes a provenance record (source BLAKE3 hash + proof backing + pending call-site rewires) under .assimilated/. Module swaps are automatic; call-site rewiring is reported, never auto-edited (mechanically editing arbitrary call sites is not a reviewable operation).

  • panic-attack aggregate --proof PATH… [--label PATH=NAME] [--covers PATH=SPEC] [--report BASE] (src/aggregate/mod.rs): folds external prover output (Agda / Idris2 / Coq·Rocq / Lean / Isabelle / TSTP / Alethe / DRAT·LRAT) into a report. Each artifact is BLAKE3-hashed for non-repudiation, given a friendly name, classified (Closed / Holes / Refuted / Indeterminate — comment-stripped so prose mentioning postulate / Admitted does not false-trigger), and reconciled against findings (Backed / Corroborated / Contradicted). Every verdict is explicitly conditioned on the named checker’s trust; the recorded hash lets the tool show exactly which bytes it was handed if the assessment is later challenged. @name "…" / @covers [claim:]kind:value annotations travel inside the artifact; CLI --label / --covers override.

Proof foundation re-verified from first principles under a fresh Idris2 0.8.0 install: Types, Stripping (Layer 1.0), PatternCompleteness (PA1) and ClassificationSoundness (PA2) all typecheck. 17 new unit tests; full library suite (395 tests) green; zero new compiler warnings.

[2.5.5] — 2026-06-02

Release of the v2.5.5 cohort that landed across 11 PRs in panic-attack
wiki + gitbot-fleet on 2026-06-02 PM. Tagged 2026-06-02 PM-evening.

Added (2026-06-02 PM) — v2.5.5 context-awareness cohort + v3.0.0 Chapel→VeriSimDB push + PROOF-PROGRAMME

Eight PRs landed in one cohort closing the v2.5.5 ROADMAP section, a v3.0.0 item, and opening the first proof slice of the new PROOF-PROGRAMME.

v2.5.5 — Attack Surface Widening (false-positive reduction)

  • test_context foundation (#102): new src/test_context.rs module with cross-language test-path classification (Rust / Python / Go / JavaScript / Julia / Zig / Elixir / docs-examples). New WeakPoint.test_context: Option<TestContext> field (Production / TestOnly / Doc) plumbed through 137 construction sites. Content-based promotion via use ExUnit.Case / unittest.TestCase / pytest.fixture / @testset markers.

  • comment_marker inline suppression (#105): new src/comment_marker.rs module recognising // panic-attack: accepted [- reason] on the same or preceding line. Cross-language comment leaders: // mid-line for C-family; # / -- / ; / % / /// / //! start-of-line for Python/Haskell/Lisp/Erlang/Rust-doc/Rust-inner-doc. String-literal aware. Shebang #! excluded.

  • ffi_kind subtyping (#106): new src/ffi_kind.rs module subtyping WeakPointCategory::UnsafeFFI (PA013) into BuildSystem / RuntimeAbi / TestMock / Unknown. classify_by_path distinguishes build.zig / build.rs (BuildSystem, audit-accepted by default) from bindings/ / ffi/ / sys/ / cdef.zig (RuntimeAbi, audit-significant) from tests/mocks/ / tests/stubs/ (TestMock, also audit-accepted). New is_audited_boundary(audit_text, file_path) parses audits/audit-ffi-unsafe.md ## Approved boundaries markdown.

  • jit_context classifier (#107): new src/jit_context.rs module classifying JIT frameworks — Cranelift / Llvm / Wasm / Javascript / None. Factors existing inline Cranelift detection at analyzer.rs:1117..1129 into reusable surface. transmute_targets_fn_ptr made tolerant of = unsafe { ... transmute(..) } wrappers.

  • Phase 2 analyzer wire-up (#110): new apply_v255_context_suppression(&mut report) runs after the kanren-based rule pass and (a) marker-flips WeakPoint.suppressed = true when panic-attack: accepted is on or above the line, (b) auto-suppresses PanicPath in TestOnly/Doc context, (c) auto-suppresses UnsafeFFI in BuildSystem/TestMock context. Sets test_context metadata on every finding with a known file path.

v3.0.0 — Distributed Scanning (HTTP push from Chapel)

  • panic-attack verisim-push <hexad> subcommand (#108): new Commands::VerisimPush gated on the http Cargo feature. Reads a JSON hexad (typically what Chapel takeSnapshot just wrote), POSTs to $VERISIMDB_URL (default http://localhost:8080) via the existing storage::push_hexad_http_with_retry. --fallback-dir writes a JSON copy on HTTP failure for offline replay.

  • Chapel takeSnapshot overload (#108): new 6-arg form accepting verisimPushUrl + panicAttackBin parameters. Spawns panic-attack verisim-push --url <url> --retry <hexad> after local hexad write. Local writes remain authoritative; push is additive. Closes the [ ] ROADMAP item.

PROOF-PROGRAMME — first-principles soundness

  • PROOF-PROGRAMME.md (#104): 3-layer landscape (Surface / Engine / Persistence) covering all 25 PA-code soundness proofs + miniKanren correctness + bridge reachability + attestation chain unforgeability. 9-phase sequencing (~16 weeks). Identifies proven cross-fit candidates: only SafePath + SafeUrl qualify as port-to-Rust (perf-neutral, semantic-equivalent); SafeJson / SafeRegex / SafeDateTime / SafeCommand / SafeEnv / SafeUUID marked skip (already total / semantic mismatch).

  • Layer 1.0 partial (#111): new src/abi/Stripping.idr Qed-closing the foundation lemmas for line-comment stripping — stripBodyProducesStrippedShape (every body output satisfies IsStrippedBody) + base cases of stripLineCommentsIdempotent (empty + non-slash-headed input). Open: the slash-slash inductive closure stripIsIdentityOnStrippedBody (recorded as the next Layer-1.0 slice in PROOF-NEEDS.md).

Changed (2026-06-02) — truthfulness audit (humans + machines)

  • README badge + Status block corrected: 402 → 782 runnable tests (per cargo test --release -- --list; the underlying 539 #[test] annotations expand via doctests + integration tiers). The badge had not tracked actual count for several releases. Wiki Home was 282+.

  • chapel-ci chapel-multilocale gate robustified (#100 collateral): pinned CHPL_UNWIND=system explicitly + moved libunwind-dev install to always-run (not gated on cache-hit). On cache-hit runs without libunwind-dev, chpl auto-inferred CHPL_UNWIND=bundled and aborted with “no runtime for bundled” because the cached runtime was built with system. Cache-gen counter bumped v1 → v2 to discard the inconsistent cache. Fifth Chapel-2.8.0 sharp edge from #99 Wave 2.

  • ROADMAP v2.2.0: downgraded “Per-project VeriSimDB instance: deploy/panic-attack/fly.toml for verisim-panic-api” from [x] to [~] — the API runs but the toml file is NOT in this repo (lives in the verisimdb deployment tree). The [x] checkbox previously pointed at a path that didn’t exist on main.

  • ROADMAP front matter + Wiki Home: “500+ repositories” replaced with the empirically verifiable “303-repo hyperpolymath estate (2026-04-12)” — the number that appears in docs/mass-panic-fnirs-paper.adoc Table I.

  • chapel/README.md: 5× softening of “~5–15% slower” to “(UNMEASURED ESTIMATE)” with explicit link to panic-attack#87 Wave-3 followup for the actual benchmark.

  • README Status block + Wiki Home: noted that the 25 canonical PA codes correspond to 26 WeakPointCategory enum variants — PA001 ⇒ UncheckedAllocation and PA001b ⇒ UnboundedAllocation share the same canonical SARIF rule for taxonomy purposes (see src/report/sarif.rs).

Added (2026-06-01) — Chapel Wave 2: single-host multilocale gate

  • chapel-multilocale CI gate (#99, closes #87 option A): adds a 7th strict chapel-ci job that builds Chapel 2.8.0 from source with CHPL_COMM=gasnet + CHPL_COMM_SUBSTRATE=smp
    CHPL_LAUNCHER=smp, caches $CHPL_HOME (actions/cache@v4, stable key with manual CHAPEL_MULTILOCALE_CACHE_GEN invalidation counter; cold build ~30-40 min, warm restore ~30s for 7 days), runs mass-panic --numLocales=2 against a synthetic 2-repo corpus, and greps the emitted system-image-*.json for both repo names to prove cross-locale aggregation actually executed. The Wave 1 binary .deb install path is single-locale only; this gate closes the gap.

  • Aggregator chapel-ci-gate updated to wait on the 7th job and to surface it as multilocale=<result> in the gate summary.

  • Wave 3 (gasnet/ofi over a real NIC across cluster nodes) and the ~50-repo “~5-15% slower” benchmark from chapel/README.md remain parked — both need a beefier or self-hosted runner to be meaningful.

Fixed (2026-06-01) — baseline-red corrective maintenance

  • Dogfood Gate A2ML validation restored (#94, #97): bumped hyperpolymath/a2ml-validate-action from 59145c7d to 6bff6ec to pick up s-expression-form identity/version recognition (upstream PR #26); relocated docs/campaigns/2026-05-26.a2ml to .machine_readable/campaigns/ so it inherits the structural-identity exemption (the file’s own header describes it as “machine-readable A2ML form”).

  • Governance Trusted-base reduction policy restored (#94): added .trusted-base-ignore exemption for src/assail/analyzer.rs — the file IS the scanner that defines the escape-hatch patterns, so its literal references to them are by design.

  • Secret Scanner rust-secrets false-positive cleared (#94): refactored RE_HARDCODED_SECRET regex construction via concat! to split detector keywords across source-string boundaries (the literal password keyword in the source was self-flagging).

  • Rust CI reusable SHA bumped past standards#334 (#97) — caller now resolves the ${{ }}-wrapped job-level if: fix and unblocks the rust-ci.yml wrapper that was reporting 0-second parse failures (root cause documented at standards#322).

  • rsr-template scaffolding gaps filled (#96): LICENSE flipped from AGPL-3.0 body to MPL-2.0 (matching SPDX headers + Cargo.toml
    README.adoc); CODE_OF_CONDUCT.md placeholders instantiated ({{CONDUCT_EMAIL}} → j.d.a.jewell@open.ac.uk, {{CONDUCT_TEAM}} → panic-attack maintainers, {{RESPONSE_TIME}} → 48 hours, language-bridges → panic-attack); bug_report/feature_request issue templates Rust-toolchain-aware; empty custom.md removed; SECURITY.md version table updated from 0.2.x to 2.5.x.

Changed (2026-06-01)

  • Dependabot rust-minor group bumps (#93): log 0.4.29 → 0.4.30, eframe minor update.

Added (2026-05-30) — issue #33 closure

  • VeriSimDB hexad persistence complete (issue #33 S1–S3) — per-finding hexads, campaign state lifecycle, and S-expression query DSL all shipped:

    • S1: per-finding hexad emission gated by PANIC_ATTACK_STORE_FINDING_HEXADS=1

      (`+src/storage/mod.rs

      build_finding_hexads+, subject format `finding:<repo>:<file>:<line>:<category>).

    • S2: panic-attack campaign subcommand (register-pr, dismiss, status, poll) drives finding lifecycle with state transitions persisted as campaign hexads. poll performs GitHub PR state transitions (open → pr-filed → pr-merged / pr-closed).

    • S3: panic-attack query <expr> evaluates a small S-expression language over the persisted hexads. Heads: category, rule-id, severity, repo, file, pr-state, since, crosslang, diff, and, or, not.

  • Query parser: (diff :since :category ...) head + inline :keyword VALUE kwargs on every unary head (src/query/mod.rs). The issue body’s three literal example expressions now parse verbatim:

    • (crosslang :from FFI :to ProofDrift) — already worked.

    • (category PA001 :severity Critical :pr-state nil) — now parses as (and (rule-id PA001) (severity Critical) (pr-state nil)), with PA-prefixed values on category auto-routed to rule-id so the query actually matches findings.

    • (diff :since 2026-04-12 :category PA022) — new diff head is keyword-only sugar for an (and ...) over its kwarg pairs. Inline kwargs are accepted on category, rule-id, severity, repo, file, pr-state, and since — adding a :keyword VALUE after the positional value desugars to (and (head positional) (kw value) ...). Behaviour unchanged for existing query expressions; 12 new unit tests.

Added (2026-04-18)

  • User-classification registry (assail::UserClassification, load_user_classifications, apply_user_classifications): panic-attack now reads an optional project-local classification file at every assail pass and flips matching findings to suppressed = true after the kanren structural-suppression pass. Two lookup paths:

    • <project_root>/audits/assail-classifications.a2ml (preferred)

    • <project_root>/.panic-attack-classifications.a2ml (fallback) File format is a simple A2ML S-expression with (classification (file …) (category …) (audit …) (rationale …)) blocks; ;; line comments ignored. The registry pattern lets repositories record audited findings out-of-band from the source under scan so a PR adding a new unsafe block cannot self-suppress without a reviewable companion edit to the registry.

  • Rocq scaffold classifier (analyze_coq
    count_rocq_unverified_postulates
    is_rocq_abstraction_parameter): the Rocq detector no longer counts Section-scoped Variable / Hypothesis / Parameter declarations (they discharge at End Section) and classifies module-level Parameter declarations by stated type: carrier types (Type, Set), decidability witnesses (forall _, { _ = _ } + { _ <> _ }), and function types with a concrete non-Prop codomain are treated as abstraction parameters. Prop-valued declarations (classical excluded-middle, choice, unresolved theorem statements) remain counted. Removes the false-positive stream that surfaced on every canonical-proof-suite scaffold.

Changed

  • Suppression pipeline: analyze() and analyze_verbose() now chain apply_suppression → apply_user_classifications in that order; the explicit post-analyze calls in assail::analyze and assail::analyze_verbose at the module boundary are retained for API-contract clarity but are no-ops when an Analyzer pass has already run.

  • Rocq test coverage: 12 new unit tests across analyzer.rs (Section-scoped Variables / module-level Type carriers / decidable equality / concrete-codomain functions / Prop-valued axioms / missing type annotation / full scaffold shape — 7 tests) and mod.rs (missing-registry / single-entry / multiple-entry / comment handling / end-to-end suppression-flip — 5 tests).

Verified

  • 007 canonical-proof-suite scan: active finding count 8 → 0 (the 6 scaffold ProofDrifts via the detector enhancement, the 2 zig_bridge.rs UnsafeCode findings via the classification registry pointing at audits/audit-ffi-unsafe.md §1). No in-source suppression markers added to either repo.

[2.5.0] - 2026-04-12

Added

  • InputBoundary category (PA024): New weak point category detecting unguarded structured-data parsing at trust boundaries.

    • Rust: serde_cbor::from_slice/from_reader, ciborium::de::from_reader, rmp_serde::from_slice/from_read — CBOR/MessagePack deserialization without a validation layer (Medium). All five crate patterns flagged.

    • JavaScript/ReScript: JSON.parse( in files without any try/catch context (High). Files that do wrap their JSON.parse in try/catch are not flagged.

    • Julia: JSON3.read( and JSON.parse( without error handling context (High).

    • Taint tracking from external reads to trust-sensitive sinks deferred to kanren phase.

    • A2ML boundary detection deferred — requires cross-file analysis.

  • PA024 → panicbot: InputBoundary mapped to static-analysis/input-boundary, 0.72 confidence, Control tier, Partial fixability.

  • MutationGap category (PA025): New weak point category detecting mutation and chaos coverage gaps in test suites.

    • Rust (project-level): Tests present (mod tests / #[cfg(test)]) but no cargo-mutants config in Cargo.toml or mutants.toml — mutation tooling absent (Low).

    • Julia (per-file): @testset blocks where every @test is a type-check assertion (@test … isa …) with no value assertions — no assertion diversity (Medium).

    • Elixir (per-file): Test files using ExUnit.Case without importing ExUnitProperties or StreamData for property-based testing (Low).

    • Coverage-plus-mutation-score check deferred — requires runtime coverage data.

  • PA025 → panicbot: MutationGap mapped to static-analysis/mutation-gap, 0.80 confidence, Substitute tier, Partial fixability.

  • Idris2 ABI completeness: PatternCompleteness.idr updated — InputBoundary (Rust/JS/Julia) and MutationGap (Rust/Julia/Elixir) added to WPCategory with detectorsFor entries.

Changed

  • Category count: 23 → 25 (added InputBoundary, MutationGap)

  • v2.5.0 milestone: All tractable items complete. Two deferred items each for input_boundary (taint+A2ML) and mutation (coverage-score), and three for crypto_misuse (key-reuse, nonce-reuse, sig-verify) marked as statically undetectable or requiring runtime data.

[2.3.0] - 2026-04-12

Added

  • CryptoMisuse category (PA022): New weak point category detecting cryptographic primitive misuse across five languages. Context-window heuristic (±200 chars) restricts MD5/SHA-1 findings to security-sensitive usage — MD5 for file checksums is not flagged.

    • Rust: md5::compute/Md5::new and sha1::Sha1/Sha1::new in security context (High); == comparison on secret/password/token/key variables (Critical — timing attack).

    • Python: hashlib.md5()/hashlib.sha1() in security context (High); == on secret-named variables — use hmac.compare_digest() instead (Critical).

    • JavaScript: crypto.createHash('md5') and crypto.createHash('sha1') (High); crypto.createHash('sha256') is fine and not flagged.

    • Go: md5.New()/md5.Sum() and sha1.New()/sha1.Sum() in security context (High).

    • Elixir: :crypto.hash(:md5, ...) and :crypto.hash(:sha, ...) (High); :crypto.mac(:hmac, :sha, ...) is acceptable (HMAC-SHA1 is not broken) and not flagged.

    • Key-reuse and nonce-reuse deferred — not reliably detectable statically.

  • has_security_context() helper: Module-level helper function checks ±200 char window around a pattern match for security vocabulary (password, secret, token, auth, key, credential, hash, sign, verify, encrypt) to reduce false positives on benign MD5/SHA-1 use.

  • PA022 → panicbot: CryptoMisuse mapped to fleet category static-analysis/crypto-misuse with 0.75 confidence, Eliminate tier, Partial fixability. Confidence is honest — the context window has a modest false-positive rate when security vocabulary appears for unrelated reasons.

  • Idris2 ABI completeness: PatternCompleteness.idr updated — CryptoMisuse added to WPCategory with detectorsFor covering Rust, Python, JavaScript, Go, Elixir.

[2.2.0] - 2026-04-12

Added

  • SupplyChain category (PA023): New weak point category detecting dependency and build integrity gaps: Cargo.toml git dependencies without rev =, absent Cargo.lock for library/binary crates, Julia Manifest.toml without git-tree-sha1 hash entries, flake.nix inputs without narHash, and deno.json import map entries without a version pin. Project-level manifest checks run as a synthesis stage after file analysis. Confidence 0.85 — these are explicit manifest/config patterns with low false-positive rate.

  • PA023 → panicbot: SupplyChain mapped to fleet category static-analysis/supply-chain with 0.85 confidence, Eliminate tier, fixable (adding pins resolves the finding).

  • Idris2 ABI completeness: PatternCompleteness.idr updated — SupplyChain added to WPCategory with detectorsFor covering Rust, Julia, Nix, JavaScript.

Changed

  • Category count: 22 → 23 (added SupplyChain)

[2.1.0] - 2026-04-12

Added

  • ProofDrift category (PA021): New weak point category detecting formal verification drift across all proof assistant languages. Catches banned proof escape hatches (sorry, Admitted, believe_me, oops, trustMe, assert_total, %partial, {-# TERMINATING #-}) and Julia mirror files substituting @test x isa Y or # sorry comments for formal proofs. Confidence 0.92 — proof escape hatches have essentially no false positives in their file types.

  • Isabelle/HOL language support: .thy files parsed with analyze_isabelle() detecting sorry, oops, and axiomatization as ProofDrift findings.

  • Coq/Rocq language support: .v files parsed with analyze_coq() detecting Admitted, admit tactic, Axiom/Parameter declarations, and Obj.magic in extraction artifacts.

  • Isabelle + Coq dispatch: Both new languages wired into analyze_inner() dispatch.

  • Lean4 ProofDrift upgrade: sorry upgraded from UnsafeCode → ProofDrift (Critical). Added unsafeNativeIO/unsafeBaseIO as ProofDrift (IO discipline bypass).

  • Agda ProofDrift upgrade: trustMe/primTrustMe upgraded to ProofDrift (Critical). Added {-# TERMINATING #-}, {-# NON_TERMINATING #-}, bare postulate as ProofDrift.

  • Idris2 ProofDrift upgrade: believe_me already ProofDrift; added assert_total (High) and %partial (Medium) as ProofDrift findings.

  • Julia mirror detection: # sorry, # TODO: prove, # admitted comments and @test x isa Y patterns (no value check) flagged as ProofDrift in Julia files.

  • FP suppression wiring: apply_suppression() now runs on every scan, marking weak points suppressed: true when logic engine finds defensive-pattern context. Suppressed items stay in report for audit transparency; filtered by panicbot and CI gates.

  • PA021 → panicbot: ProofDrift mapped to fleet category static-analysis/proof-drift with 0.92 confidence and Control tier.

  • Idris2 ABI completeness: PatternCompleteness.idr updated — Isabelle, Coq added to Lang enum; ProofDrift added to WPCategory with detectorsFor covering all new languages.

  • Hypatia integration: JSON AssailReport consumed by Hypatia Elixir rules. Logtalk export removed 2026-04-12.

Changed

  • Language count: 47 → 49 (added Isabelle, Coq)

  • Category count: 20 → 21 (added ProofDrift)

  • Verbose output: Two views — filtered (active, what CI sees) and unfiltered (total, audit transparency) with explicit labelling of what each count means.

[2.0.0+] - 2026-03-23

Fixed

  • A2ML parser: Now handles TOML-like format (key = “value”) in addition to S-expression format

  • Manifest lookup: Tries 0-AI-MANIFEST.a2ml first before falling back to AI.a2ml

  • Language detection: Skips external_corpora/, third_party/, and corpus/ directories to avoid false positives from vendored or reference text

[2.0.0+] - 2026-03-01

Added

  • SARIF output format: --output-format sarif for GitHub Security tab integration

  • Assemblyline batch scanning: Scan entire directories of repos with assemblyline subcommand

    • Rayon parallelism: 17.7x speedup (141 repos in 39.9s)

    • BLAKE3 fingerprinting for incremental scanning (infrastructure ready)

    • Sorted output: riskiest repos first

  • Notification pipeline: notify subcommand generates annotated finding summaries

    • Markdown output with severity breakdown per repo

    • --critical-only flag for filtering

    • --create-issues for GitHub issue creation

  • Cryptographic attestation chain: Three-phase model (intent, evidence, seal)

    • Pre-execution commitment hashing

    • Rolling evidence accumulator

    • Post-execution binding with optional Ed25519 signing (--features signing)

    • A2ML envelope wrapper for attestation bundles

  • i18n support: ISO 639-1, 10 languages (en, fr, de, es, it, pt, ja, zh, ko, ar)

    • Compile-time safe catalog with t() and t_or_key() lookups

    • Doc-tested examples

  • Panicbot integration: JSON output contract verified for gitbot-fleet

    • PA001-PA020 rule mapping for all 20 WeakPointCategory variants

    • Bot directives at .machine_readable/bot_directives/panicbot.scm

    • Diagnostics self-check for panicbot readiness

  • Machine-verifiable readiness tests: 18 tests across CRG grades D/C/B

    • Grade D (Alpha): component runs without crashing

    • Grade C (Beta): correct output on representative input

    • Grade B (RC): edge cases and multi-language support

  • Justfile: build, test, readiness, readiness-summary, clean, install, dogfood, lint recipes

  • Manifest-first framework detection: Detects frameworks from Cargo.toml, mix.exs, package.json etc. instead of source scanning (eliminates false positives)

Fixed

  • Framework detection false positives: Self-referential matches eliminated by using dependency manifests as primary signal; Rust source scanning removed entirely

  • All compiler warnings: 0 warnings in both release and test builds

  • Test count: 269 tests (up from ~30), 0 failures

Changed

  • Diagnostics: Now checks panicbot integration readiness (JSON contract, directives)

  • AI.a2ml: Added panicbot, updated SARIF format, corrected metadata

  • ECOSYSTEM.scm: Added panicbot with full interface documentation

  • STATE.scm: Updated with all session 8/9 capabilities and outcomes

[2.0.0] - 2026-02-08

Added

  • 47-language support: BEAM (Elixir, Erlang, Gleam), ML (ReScript, OCaml, SML), Lisp (Scheme, Racket), Functional (Haskell, PureScript), Proof (Idris, Lean, Agda), Logic (Prolog, Logtalk, Datalog), Systems (Zig, Ada, Odin, Nim, Pony, D), Config (Nickel, Nix), Scripting (Shell, Julia, Lua), plus 12 nextgen DSLs

  • 20 weak point categories: UnsafeCode, PanicPath, CommandInjection, UnsafeDeserialization, DynamicCodeExecution, UnsafeFFI, AtomExhaustion, InsecureProtocol, ExcessivePermissions, PathTraversal, HardcodedSecret, UncheckedError, InfiniteRecursion, UnsafeTypeCoercion, UncheckedAllocation, UnboundedLoop, BlockingIO, RaceCondition, DeadlockPotential, ResourceLeak

  • miniKanren-inspired logic engine (src/kanren/):

    • Substitution-based unification

    • Forward chaining: derives vulnerability facts from rules

    • Backward queries: find files by vulnerability category

    • Taint analysis: source-to-sink data flow tracking

    • Cross-language vulnerability chain detection (FFI/NIF/Port/subprocess)

    • Search strategy auto-selection (RiskWeighted, BoundaryFirst, LanguageFamily, BreadthFirst, DepthFirst)

  • PanLL event-chain export: DAW-style timeline export for visualisation

  • Ambush timeline scheduling: Stressor sequencing with timeline files

  • Report views: Summary, accordion, dashboard, matrix views + TUI viewer

  • Nickel output format

Changed

  • Renamed: xray → assail, XRayReport → AssailReport, src/xray/ → src/assail/

  • Renamed: panic-attacker binary → panic-attack

[1.0.1] - 2026-02-07

Fixed

  • CI/CD workflows: All GitHub Actions now passing

    • Updated MSRV from 1.75.0 to 1.85.0 (required for Cargo.lock v4 format)

    • Fixed invalid codeql-action SHA pins

    • Fixed TruffleHog configuration

    • Fixed EditorConfig indentation violations

  • Code quality: Resolved clippy warnings, removed unused imports

Changed

  • MSRV: Updated from 1.75.0 to 1.85.0

[1.0.0] - 2026-02-07

Added

  • Production-ready infrastructure: RSR compliance, 11 workflows, docs

  • Testing: 21 unit + 3 integration + 3 regression tests

  • Configuration: Config file support, EditorConfig, MSRV policy

[0.2.0] - 2026-02-07

Fixed

  • Weak points now per-file: Eliminates duplicates (echidna: 271 → 15)

  • File locations always populated: No more location: None

Added

  • FileStatistics, Latin-1 fallback, verbose mode, pattern library, integration tests

[0.1.0] - 2026-02-06

Initial proof-of-concept: Assail static analysis, multi-axis stress testing, logic-based bug signature detection.