Skip to content

Latest commit

 

History

History
132 lines (93 loc) · 32.8 KB

STATE.md

File metadata and controls

132 lines (93 loc) · 32.8 KB

STATE.md — Holmes build state (loop §5)

Updated: 2026-07-22 (Session 10 — Alfred Stage-1 integration evidence landed (Alfred main, PRs #21–#23): the artifact-level guard test is green in Alfred CI — obligation 1, the RC embedding unblock, is satisfied; the cross-repo obligations below are flipped; provider-lockdown.ts retired. Remaining Alfred-side item = the signed-update rollback dry-run, which gates Alfred's beta.4 publish, not Holmes RC.) · Session 9 — Phase 4 build (Observability & hardening): new crates/holmes-core/src/observability/ — born-redacted opt-in telemetry (content-free by construction; cross-stack correlation), finalized deny-by-default permission manifest; telemetry threaded additively through the six-phase case; D-14 = Option A (analytical open beta) DECIDED, riders (a) investigative absent from beta / (b) plain beta-scope copy wired; lock 4b security audit 6/6 HELD, zero criticals; F-037 doc-overclaim fixed) · prior: Session 8 Phase 2.5 (PR #15; locks 2.5a–2.5d; F-034/F-035/F-036; D-14 PROPOSED; A-09); Session 7 housekeeping (PR #14; F-032/F-033); Session 6 Phase 2 (PR #11) + A-07 (PR #12) + F-031 (PR #13); Session 5 Phase 1 (PR #10); Phase 0 CLOSED (PRs #7–#9) · Maintainer: the loop; refreshed at every checkpoint.

Git state

  • Current branch (Session 9): claude/claude-code-git-bash-path-f30buf, restarted from main after PR #15 merged (Phase 2.5). Prior phase PRs: #6 (Phase 0 guard), #7 (0e), #8/#9 (Phase 0 close), #10 (Phase 1), #11 (Phase 2), #12 (A-07 + F-029/F-030), #13 (F-031), #14 (README/F-032/F-033), #15 (Phase 2.5 + F-034/F-035/F-036).
  • origin/main head: ccd78e6 (merge of PR #15 — Phase 2.5); this session builds on it.
  • Ancestry: FF-OK
  • Landing mechanics: PR-only to main, explicit go-ahead (D-08); connector re-sync after every canon merge (F-012)

Gate Zero (loop §3.0) — CLEARED 2026-07-18

Call Decision Record
(a) Charter verdict 2026-07-13 second-pass audit accepted (SHIP WITH FIXES) D-06 · report on disk docs/audit/
(b) D-01 license AGPL-3.0-or-later ratified; LICENSE byte-identical to Alfred's D-01, F-001
(c) Git reconciliation Branch restart + one approved --force-with-lease; nothing lost D-07
(d) Discipline Confirmed D-08

Task 0 status (loop §3) — COMPLETE

3.0 ✅ · 3.1 ✅ (landed via PR #5) · 3.2 ✅ (spec v2.1, E1–E14, derived files re-seeded; F-014 resolved) · 3.3 ✅ (alias audit; primary re-verification still egress-blocked — carry) · 3.4 ✅ (roadmap reconciled; A-03). Standing spec amendments: A-02…A-05 (+ A-05 note: GOOSE_DISABLE_KEYRING now primary-cited from source @ 8e78960e, crates/goose/src/config/base.rs).

Phase 0 lock inventory (loop §6)

Lock Status Evidence (executed 2026-07-18, this container)
0a — AC-DL-1 §§1–5, §7 VERIFIED (local, release, hermetic) — v3-conformant cargo test --release --locked -p holmes-guard → 40 passed / 0 failed (resolution §2/§5, proxy §2a/§4/§5 with planted proxy-honoring server + planted upstream forward proxy + named-endpoint denial, spawn §3, structural §1, unit; incl. regression + v3-delta tests). §6 SCHEDULED to lock 2b, recorded visibly in CI. CI-run leg PROVEN 2026-07-19: acdl-gate run #9 green on main head 3af25d9 (first real CI execution), then green on PR #7's fix commit 6595499 (both push- and PR-triggered runs) after correctly catching F-025 on run #10 — the gate demonstrably fires and passes in CI, not just locally
0b — AC-DL-2 all seven VERIFIED (local, release) — v3-conformant v3 control convention: positive control (planted lockfile) → FAIL / exit 1; negative control (real tree, lockfile discovery) → CLEAN / exit 0. Multi-ecosystem lockfile walk (§1), documented seed table with rationale (§2), dependency-path in failure output (§3 — pulled in via holmes-app -> middleware-lib -> async-openai). Criteria 1–7 each covered by named tests; c7 = joint workflow, same run as 0a
0c — ACP round-trip COMPLETE (2026-07-19, account funded) Funded re-run, same container/command: verdict ROUND-TRIP COMPLETE, exit 0 — anthropic/claude-sonnet-5 L1b-permitted pre+post handshake, L2 sanitized spawn, 14 ACP frames, egress 1/1 api.anthropic.com:443 allowed via L1a, streamed model completion "pong" (the smoke prompt's exact requested reply; 4 bytes; verdict from the F-026-fixed harness, which cannot count error relays as completions). Born-redacted transcript: docs/audit/evidence/0c-transcript-2026-07-19-complete.json. Prior blocked-run evidence retained below, superseded not deleted
0c prior evidence (superseded 2026-07-19) PARTIAL — guard/protocol legs proven, completion credits-blocked 2026-07-19, this container: goose rebuilt from the pin via scripts/build-goose.sh (aaif-goose/goose @ 8e78960e, v1.43.0, trimmed features, binary sha256 439a282e…7056, provenance file beside the binary). holmes-smoke --provider anthropic --model claude-sonnet-5 --credential-env MY_ANTHROPIC_API_KEY: pre-handshake L1b permitted → L2 sanitized spawn → real goose acp handshake + session/new + prompt (12 frames) → post-handshake configOptions pair (anthropic/claude-sonnet-5) L1b-permitted → egress events 1/1 api.anthropic.com:443 allowed through L1a. The endpoint authenticated the key (HTTP 400 billing — "credit balance is too low" — not 401 auth); goose relayed the error as agent text; harness verdict PROVIDER ERROR RELAYED, exit 6 (honest verdict added this session — F-026, fixed + re-run confirmed). Negative control: openrouter → typed pre-handshake denial, exit 3. Transcript born-redacted (verified: no key material) at docs/audit/evidence/0c-transcript-2026-07-19.json. Remaining to COMPLETE: fund the API account, re-run the same command, expect exit 0 with a model completion. Never faked
0d — embedding contract COMPLETE (Session 4, 2026-07-19) crates/holmes-core: §6.2 artifact types as committed canon (ResearchBrief in; EvidencePack/CaseFile out — types and validation only), invariant 5 enforced at construction (empty provenance / out-of-range confidence unrepresentable), invalidation-not-deletion (supersede flags valid_until, no removal API), handoff-only terminal state (journalist/lawyer/community/interim human reviewer; no execute/apply API), read-only provider seam re-exporting guard rosters + resolution (holmes_core::provider). Zero third-party deps (workspace-internal holmes-guard only). Evidence: 9 new tests green in release --locked (4 unit + 3 Alfred-shaped consumer embedding_contract.rs + 2 structural no_blueprint_exports.rs incl. F-020-style fires-check); workspace suite 49/49; clippy clean, fmt clean; CI runs the workspace suite in the acdl-gate job ("Lock 0d" step). Alfred's artifact-level guard-test obligation: already ledgered in the cross-repo table below. Phase 1 schema amendment (knowability/limits_of_this_finding) deliberately not pre-implemented
0e — full CI (SBOM/scanners) CVE gate PROVEN in CI (2026-07-19) .github/workflows/supply-chain.yml: Syft SBOM (SPDX+CycloneDX) + OSV-Scanner (primary, exit 1 on any vuln) + Grype (cross-check, --fail-on high, exit 2). No Trivy (CVE-2026-33634). Action-free ⇒ SHA-pinning satisfied by construction; scanners pinned (syft v1.48.0 / osv-scanner v2.4.0 / grype v0.116.0), checksum-verified fail-closed. Executed CI evidence: run #1 on 107f0a5 → success in 51s (2026-07-19 07:30:49Z — pinned installs verified against real release checksums, SBOM generated, OSV clean, Grype clean); re-run green ×2 on fix commit 6595499; landed on main via PR #7 (b6b19dc), all four checks green at merge. Still open in 0e: provenance/attestation (spec §6.6) — carry; human: mark acdl-gate + supply-chain required checks on main

Phase 1 lock inventory (loop §6, Session 5 — 2026-07-19, this container)

Lock Status Evidence
1a — ACH + corroboration gate at emission VERIFIED (local, release, hermetic) holmes-core::analysis: emission gate rejects any current finding with < 2 independent source roots (documented heuristic: distinct normalized roots — a deterministic floor, not proof of true independence), plus A-07 requirements (knowability assigned, non-empty limits_of_this_finding); non-empty provenance + confidence ∈ [0,1] already unrepresentable at Finding construction. Tests: lock1a_uncorroborated_findings_cannot_leave_the_case, lock1a_corroborated_case_emits_and_hands_off, emission unit suite. ACH: complete-matrix-or-refuse, fewest-inconsistencies verdict, tie-at-top reported honestly, diagnosticity computed
1b — full analytical case end-to-end on the smoke model COMPLETE (live, 2026-07-19) holmes-case (new bin) drove all six phases over one guarded ACP session (anthropic/claude-sonnet-5, MY_-prefixed credential seam): model supplied 4 hypotheses (la lluvia), 12/12 ACH cells, one likelihood pair, and an el-diablo challenge; deterministic side owned intake, fixtures, verdict, KAC, knowability (rule-assigned), emission, handoff. Result: verdict CASE COMPLETE exit 0, 64 frames (F-030 correction: the earlier "54" was the first run's count; the committed evidence is the re-run on the F-026-fixed harness), egress 1/1 api.anthropic.com:443 allowed, ACH honestly reported a 3-way tie at top (flag flowed to risk_flags), only E1 diagnostic. Born-redacted transcript: docs/audit/evidence/1b-case-transcript-2026-07-19.json
1c — blueprint-type CI check GREEN no_blueprint_exports structural test covers the whole holmes-core source tree including the new analysis module; passing in the workspace suite (CI step "Lock 0d")
1d — recipe safety scan live from the first recipe VERIFIED (local); CI steps landed First recipes committed (recipes/la-lluvia.yaml, recipes/el-diablo.yaml); recipe-scan bin (holmes-guard) detects zero-width/bidi/tag-block/variation-selector smuggling, fails closed, never auto-strips. Positive control: planted fixture (planted_recipe_smuggled.yaml, real U+200B/U+202E/tag-block payload spelling a hidden instruction) → 21 hits, exit 1. Negative: real recipes CLEAN exit 0; Spanish text does not false-positive. CI: both controls added to acdl-gate.yml

Phase 1 engines shipped (deterministic legs): hypothesis objects (prior, predicted-present/absent, running log-likelihood with per-update trace), LR scorer (both conditionals required — prosecutor's fallacy unrepresentable; CalibrationStatus::Uncalibrated carried, gating → 2.5), ACH matrix, KAC (unexamined/failed always visible), first-principles quarantine (unprovenanced input has no type-level path into a rebuild), six-phase state machine (forward-only, typed prerequisites, intake can decline terminally). Canon §5 firewall: the brief's self-reported-certainty field is referenced nowhere in analysis/ — enforced by structural regression test (which caught its own doc comment during the build; reworded per the F-025 precedent). Model-side legs live in recipes/ and the holmes-case prompts.

Phase 2 lock inventory (loop §6, Session 6 — 2026-07-19, this container) — The Wall

Backend/architecture (D-09/D-12): Graphiti dropped (F-027: base-deps openai + posthog — vendor + phone-home gates). The Wall is the owned subset in the new crates/holmes-wall, on Neo4j Community Edition (GPLv3) via neo4rs 0.8.0 (MIT, denylist-clean tree — rider (d), verified via crates.io sparse index). No Graphiti, no openai, no phone-home; zero Python surface.

Lock Status Evidence
2a — invalidation-not-deletion VERIFIED (contract, hermetic) + live leg env-gated Bi-temporal model (graph.rs): occurred-at/valid-until valid-time + ingested-at transaction-time; supersede sets valid_until+invalidated_by and appends — the Wall trait exposes no delete/remove capability. InMemoryWall reference store proves the contract (nothing deleted, superseded record preserved+flagged, facts_as_of returns history). Cypher builders (cypher.rs) asserted to carry no DELETE/DETACH DELETE/REMOVE (in-module test + tests/no_delete_audit.rs). Live leg (Neo4jWall, tests/wall_locks::lock2a) authored + env-gated on HOLMES_NEO4J_URI; in-container Neo4j is org-egress-blocked (Docker image blobs on production.cloudfront.docker.com → 403; dist.neo4j.org blocked) — same wall as 0c/0e; runs on any host with Neo4j
2b — AC-DL-1 §6 lands VERIFIED (local, release) — the CI §6 marker flips SCHEDULED→EXECUTING The deferral recorded in every CI run since Phase 0 is closed. memory.rs: the default (no-explicit-config) memory layer is local-only (loopback Ollama) — reaches no cloud endpoint at all; an excluded provider is unrepresentable (validated via L1b + the L1a allowlist). Live boundary (tests/wall_locks::lock2b): a forced CONNECT to the excluded canary → HTTP 403 at L1a + a Denied egress event. Structural inversion of Graphiti's excluded-vendor default. acdl-gate.yml §6 step now EXECUTING
2c — Tier-2 ingestion quality VERIFIED (scorer, hermetic) + live extraction run (Tier-1) Deterministic scorer (ingest.rs): grounded / ungrounded-citation / claim-exceeds-citation / malformed, as a failure-rate report — no model call in scoring; hermetic tests with good+bad fixtures. Live run (holmes-ingest bin) on the smoke model: 6/6 episodes grounded, 0 failures, egress 1/1 api.anthropic.com allowed; transcript docs/audit/evidence/2c-ingest-transcript-2026-07-19.json. Tier-2-on-Ollama failure-rate numbers are environment-gated (Ollama model egress blocked in-container, established Phase 0); the scorer is identical for both tiers — carried
2d — supervised backend VERIFIED (local, hermetic) supervise.rs: SupervisedBackend with kill-on-drop / no-orphan (Drop kills+reaps), path-confined data dir (..-collapsed confinement check; dir outside root refused), absolute program path required (no system-install assumption). Proven deterministically with a stand-in child: spawn → drop → the PID is reaped (signal-0 probe); confinement + absolute-path refusals tested
2e — AC-WP weight provenance VERIFIED (local, hermetic) provenance.rs: verify_weights checks SHA-256 (+ optional size + optional required signature) before load, failing closed on any mismatch; "before load" is structural — a loader accepts only a VerifiedWeights token minted solely by a passing verify. Tests: valid verifies; equal-length tamper → checksum fail-closed; size mismatch, malformed digest, and required-signature-with-no-verifier all fail closed

Sweeps (release, --locked): 101 tests pass (0 fail); fmt clean; clippy 0; AC-DL-2 on the new neo4rs lock → CLEAN (positive control still fires exit 1); recipe-scan CLEAN; A-01 0 hits; openai/posthog/graphiti absent from Cargo.lock, neo4rs present. F-028: the AC-DL-2 self-scan caught the new crate naming excluded vendors (enforcement code + comments) — resolved by the F-025 reword precedent (canary sourced from the exempted policy.rs), zero new exemptions.

Phase 2.5 lock inventory (loop §6, Session 8 — 2026-07-20, this container) — Safety before surface (HARD GATE)

Where it lives: new crates/holmes-core/src/safety/ (reader, approval, subjects) + the calibration-aware emission gate + crates/holmes-core/tests/safety_locks.rs and hostile fixtures under tests/fixtures/hostile/. Deterministic and model-free throughout (canon §5); the model-shaped seam is the ReaderBackend trait whose signature is the confinement. No new third-party dependency (holmes-core stays workspace-internal → Cargo.lock unchanged → CVE gate not re-triggered).

Lock Status Evidence
2.5a — hostile-content suite green (injection defense) VERIFIED (local, release, hermetic) CaMeL-style dual-model quarantine (reader.rs): UntrustedContent holds raw bytes (private, born-redacted Debug, one pub(crate) name-firewalled accessor); the quarantined ReaderBackend receives raw text + request and nothing else; the deterministic validator rejects (never repairs) empties/oversizes/recipe-scan smuggling classes/control chars/unrequested kinds/floods; survivors are sealed Extraction values carrying no authority (no text→grant/phase/handoff path). Planted instruction.html (SYSTEM OVERRIDE + forged grant + exfil demand + raw-byte canary) drives a full case: handoff goes to the human reviewer at 0.7, the demanded exfiltrate.send stays deny-by-default refused, the raw-byte canary never reaches the emitted pack. Planted smuggled.txt (zero-width/bidi/tag-block) → ≥3 smuggling rejections, clean lines survive. Raw-bytes accessor name-firewalled across every crates/*/src
2.5b — knowability gate proven VERIFIED (local, release, hermetic) Emission gate now calibration-aware (emission::emit(pack, CalibrationStatus)): a confident finding (confidence ≥ CONFIDENT_FLOOR, ASSUMED 0.75) from an uncalibrated source cannot emit — and the core mints only Uncalibrated (no fake calibration machinery), so the cap is the calibration fallback; and bare high confidence in a LowValidity domain is blocked without the prominent uncertainty_statement. Named recovery downgrade_uncalibrated supersedes (never deletes) the confident finding with a capped copy and flags it. Proven on the full case path (lock2_5b_confident_finding_blocked_then_downgraded_then_emits) + emission unit suite
2.5c — approval round-trip headless VERIFIED (local, release, hermetic) approval.rs: deny-by-default per case; no tool fires without a ToolGrant, and the only mint is an operator Approved decision on a previewed request. Unanswered blocks (authorize refuses); Denied grants nothing; decisions cannot be flipped; ToolGrant sealed (unforgeable); log born-redacted (case/tool/decision/at — no purpose/args/content). Round-trip demonstrated headlessly: preview → approve → grants; deny path; immutability
2.5d — anti-doxxing refusals + handoff-only VERIFIED (local, release, hermetic) subjects.rs: private-individual targeting refused permanently AND declines the case terminally (no override API); Blacksky's doxxing definition adopted verbatim (spec §6.4) — disclosure of every InfoClass without a sealed ConsentRecord refused for any subject; person-naming findings carry the stricter bar (≥3 independent roots, verbatim quote on every provenance entry, uncertainty statement) at resolution, decorated duplicates cannot reach the floor; labeling non-destructive (superseded, never deleted). Structural test: no action API anywhere in src/safety — resolution stays handoff-only

Adversarial verification (loop §6 2.5(v), MANDATORY): 14 defense claims, one independent skeptic each + a second on every breach (17 agents, 0 errors). 12 HELD, 3 breaches — all fixed with locking regressions before this checkpoint: F-034 (MAJOR, cross-crate raw-bytes reach → pub(crate) + workspace-wide firewall), F-035 (MINOR, invisible-Unicode coverage gap → extended smuggling_class), F-036 (MAJOR, forgeable consent → sealed ConsentRecord). Ledgered in full.

Sweeps (release, --locked): 130 tests pass (0 fail); fmt clean; clippy 0; recipe-scan real CLEAN / planted fires exit 1 (21 hits) — re-run after the smuggling_class extension; AC-DL-2 real CLEAN / planted fires exit 1; A-01 0 hits (code/config); osv-scanner.toml intact at exactly 4; Cargo.lock unchanged (no new deps). stated_confidence firewall now covers src/safety as well as src/analysis (≥10 files scanned).

Carried forward: the reader's process-separation leg (a separate no-tools model session behind L2) is design posture until run on an open-egress host — recorded in docs/security.md as an Alfred/OS obligation; approval-request rendering is Alfred's surface (protocol + blocking tested headless here); CONFIDENT_FLOOR/CALIBRATION_CAP/PERSON_FINDING_ROOT_FLOOR are ASSUMED numbers, amendable only by D-item. A-09 (spec/canon uncertainty_statement schema delta) PROPOSED. → Beta Scope Decision = D-14 PROPOSED (brief in docs/beta-scope-decision-brief.md); agents must not resolve it.

CVE gate on the neo4rs tree (D-13, 2026-07-19): the supply-chain gate fired as designed on 4 informational: unmaintained RUSTSEC advisories (paste/backoff/instant/rustls-pemfile — no CVSS, no fix; verified primary). Martin ruled allowlist exactly the 4 → osv-scanner.toml (each with rationale + ignoreUntil = 2026-10-17 90-day backstop; neo4rs 0.9 stable reopens the bump). CI now (a) passes --config osv-scanner.toml and (b) asserts the suppression set is exactly those 4 — a 5th suppression fails the gate until a new D-item ratifies it. Gate stays fail-closed for every scored/fixable advisory (rider d). First CI run of these steps lands on the PR head.

Phase 4 lock inventory (loop §6, Session 9 — 2026-07-20, this container) — Observability & hardening

Beta scope: D-14 = Option A — analytical open beta (DECIDED, Martin, 2026-07-20). Phase 4 hardens the analytical surface only (rider (a): investigative surface absent from the beta, not dark-flagged). New crates/holmes-core/src/observability/ (telemetry, permission).

Lock Status Evidence
4a — telemetry provably content-free on captured payloads VERIFIED (local, release, hermetic) observability::telemetry: born-redacted, opt-in (disabled default records nothing), local-only (no I/O, no phone-home — only std::sync::atomic imported). Content-free by construction: every TelemetryEvent payload field is a compiled &'static str class-name or an integer — no owned String/buffer/generic. CorrelationId (opaque u64 counter) is the cross-stack trace correlation — ties one case's events with no content. The case is instrumented additively (telemetry disabled by default → zero behaviour change; the pre-Phase-4 130 tests unchanged). Lock test drives a full observed case seeded with a content canary + an sk-ant secret canary and asserts no payload contains either, all under one correlation id; opt-in and declined-case variants covered. F-037 (audit nit): the "cannot" overclaim corrected to "closed vocabulary + no-leak", the no-leak half enforced by telemetry_feeding_source_contains_no_leak
permission manifest — deny-by-default, finalized VERIFIED (local, release) observability::permission: decide is total — unlisted capability → Denied (no third arm); read-only → RunFree; write/shell/network → AskFirst (routed to the 2.5c approval protocol). Private capabilities, single constructor, no runtime widening. Spec §7 Phase 4 "deny-by-default permission manifest" + line 166 read/write split
D-14 riders VERIFIED (local, release) (a) investigative cargo feature off by default — the reserved Phase-3 landing zone, gating nothing yet (absent, not dark-flagged); analytical_beta manifest declares no Investigative capability; locks beta_manifest_carries_no_investigative_capability + beta_build_compiles_with_investigative_off + compile-time const _: () = assert!(INVESTIGATIVE_ABSENT). (b) README Project Status states plainly that collection mode is not shipped and sits behind the safety gates
smuggling regression (from Phase 1) stays green VERIFIED recipe-scan real CLEAN / planted fires exit 1 (21 hits) — including the F-035 smuggling_class extension
4b — Skillsmith-style security audit passed VERIFIED (2026-07-20) 6 independent auditors (+ adversarial corroboration on any finding), 0 errors. 6/6 HELD, zero findings, zero criticals. Attacked: telemetry content-freedom (every emit site traced to a literal), opt-in/local-only (no forge-to-Enabled, no I/O), deny-by-default decide (total, no runtime widening), D-14(a) investigative absence, correlation-not-content, instrumentation-is-additive (side-effect-free w.r.t. control flow). Lone honesty nit (F-037) pre-empted before the audit closed

Sweeps (release, --locked): 143 tests pass (0 fail); fmt clean; clippy 0; AC-DL-2 real CLEAN / planted fires exit 1; recipe-scan real CLEAN / planted fires exit 1; A-01 0 hits (code/config); osv-scanner.toml intact at exactly 4; Cargo.lock unchanged (no new deps → no CVE-gate trigger). stated_confidence firewall covers src/safety; new no-leak firewall covers all src.

Carried forward: telemetry decide is declarative — wiring it to a runtime tool-dispatch layer (with a lock asserting every dispatched capability routes through decide) is a forward-looking item when Phase 3/6 tool code lands; emit_evidence_pack/hand_off_case are classed ReadOnly (no outward reach; revisit if Phase-6 handoff code ever writes/transmits). → Next: Phase RC (analytical beta release candidate); Phases 3 and 5 proceed after beta behind unchanged gates (D-14 rider (c)); the RC beta-safety re-run stays mandatory.

Adversarial self-verification (2026-07-18)

7 skeptics attacked each guard property; every reported defect re-reproduced against source before acceptance. 3 claims held (L2 env-strip, BYOK invariant, born-redacted output). 4 confirmed defects fixed + regression-locked this session: F-017 (MAJOR, L1a forward-proxy re-dispatch), F-018 (MAJOR, AC-DL-2 router exact-match evasion), F-019 (MINOR, mid-token excluded-family hole), F-020 (MINOR, §1 structural test fidelity). See findings-ledger.

Environment (this container)

  • Provider key: ABSENT (ANTHROPIC_API_KEY unset; only ANTHROPIC_BASE_URL=https://api.anthropic.com, keyless probe → 401; no Google key; D-05 residual)
  • goose: INSTALLED from source — origin aaif-goose/goose @ 8e78960e535ab7f34630e7c5921a42f146cbc9f4 (Apache-2.0, verified on disk), v1.43.0, binary /home/user/goose-src/target/release/goose (trimmed: --no-default-features --features rustls-tls — no V8/code-mode, telemetry, aws-providers, system-keyring, updater). Container-ephemeral: rebuild per session or vendor.
  • Do not cargo install goose-cli — the crates.io name is a squatter (F-016).
  • Ollama: install blocked (ollama.com + GitHub release egress 403) — Tier-2 local smoke unavailable here.
  • Siblings: Alfred / wecanjustbuildthings.dev ABSENT from this container (fresh clone, no symlinks). D-01 evidence already ledgered; no Alfred-touching work in scope.

Cross-repo obligations — Alfred

Obligation Status
Artifact-level guard CI test (runs in Alfred's CI) VERIFIED — the RC unblock (obligation 1). artifact-guard job green on Alfred main (CI #100 @ 056edcf; run github.com/MartinMontero/Alfred/actions/runs/29883979445): tauri build → the built product refuses an excluded provider and a planted config, permits a good one. Exercises the Linux artifact (stated in the job summary); the required-check click on Alfred main is Martin's.
OS/artifact-level egress enforcement (L1a residual: hostile binary ignoring proxy env) VERIFIED-with-honest-residual. L1a proxy env pinned on the shipped artifact (probe: HTTPS_PROXY set, NO_PROXY absent); the documented residual (a hostile binary ignoring proxy env escapes the library boundary) is ledgered as OS-level future work in Alfred (Track 3(2)), not declared closed.
Signed update channel with rollback VERIFIED (channel) · PENDING-Martin (rollback exercise). Channel live + minisign-signed (beta.2 + beta.3 published, feed verified live). Rollback checklist authored in Alfred (docs/beta/rollback-checklist.md; journey J6 — updater dry-run vs. a local latest.json). The rollback dry-run + the forward beta.3→beta.4 update proof are Martin's; they gate Alfred's beta.4 publish — an Alfred-side gate, not a Holmes RC blocker.
Memory/resurfacing channel VERIFIED. Born-redacted by construction in Alfred (telemetry.rs typed allowlist — counts/durations/enums/ids only, no field for note body/prompt/tool arg/key); canary born_redacted_canary_no_secret_or_note_body_on_disk green; live tap reads only bounded scalars; memory-review is proposal-only. Holmes's Loop E rides this channel and is granted no new one.
First-run rendering (tool-approval UX surface) VERIFIED (surface + structural gates) · live approval routed to Windows. Alfred renders only EmittedEvidencePack (sole DTO constructor — a raw pack has no render path), surfaces gate denials verbatim, and carries the ApprovalRequest-preview + consent operator patterns (grants and ConsentRecord/SubjectScope mint only from explicit UI actions, never from case content). A live approval on a real goose session is part of the routed Windows walk.
holmes-guard adoption retiring provider-lockdown.ts DONE. provider-lockdown.ts deleted (git rm; grep -r provider-lockdown src/ clean); provider selection routes through L1b, goose spawns through sanitized_spawn, sessions ride the L1a proxy; ADR-0008. 70 removed TS refusal tests → 50 Rust tests + the artifact test (mapping: Alfred docs/audit/holmes-stage1-track1-mapping.md); goose shell-sidecar capability perms removed. Crates pinned at Holmes 63f877a (re-pin to the RC tag when cut). Live-goose fire on the Windows artifact vs. staged goose 1.41.0 routes to the Windows walk (skew note: guard verified Holmes-side against 1.43.0 — a bump is a separate Martin decision, not taken).
Freshness note: src/lib/provider-policy.test.ts stale model-id fixtures RESOLVED — moot: the file was deleted with the TS policy layer (Track 1).

Staging obligations — human

  1. Fund the API account to finish 0c — RESOLVED 2026-07-19: human funded the account in-band; re-run produced exit 0, verdict ROUND-TRIP COMPLETE, streamed completion "pong" (transcript in docs/audit/evidence/). D-05 residual closed. Key-storage note stands: never store under ANTHROPIC_API_KEY in claude.ai/code env settings (reserved, auto-stripped); MY_ANTHROPIC_API_KEY is the working convention.
  2. F-015: supply AC doc v3 — RESOLVED 2026-07-18: v3 landed at root, normalized into docs/acceptance/ (hash-gated), root removed, v2 to history; v3-conformance pass closed F-021…F-024.
  3. Mark the acdl-gate and supply-chain workflows required status checks on main — DONE 2026-07-19 (Martin). acdl-joint-gate and sbom-and-cve-scan are both required status checks on main; both gates are now binding, not advisory. Completes AC-DL-2 c7's "required gate" clause and makes the 0e CVE gate binding.
  4. Still-unlocated upstream artifacts (F-009/F-011): kickoff v2, Iterative quality validation process.md, claude-code-epistemic-integration-prompt.md.
  5. DeepSeek alias primary re-verification (api-docs.deepseek.com egress) — carry to next open-egress session and Phase RC.

Resume point

Session 10 (2026-07-22) — Alfred embedding evidence landed; RC embedding blocker cleared. Alfred's Stage-1 Holmes integration merged to Alfred main (PRs #21–#23, CI green): compiled holmes-guard adopted, provider-lockdown.ts retired, the artifact-level guard test green in Alfred CI (obligation 1 — the RC unblock — satisfied; run URL in the cross-repo table above). The four perimeter obligations flipped (one — signed-update rollback — stays PENDING-Martin as Alfred's beta.4 publish gate, not a Holmes RC blocker). Phase RC may proceed. Confirm this flip against wherever the RC gate is formally tracked before cutting the RC tag; when it is cut, Alfred re-pins its Holmes dep from 63f877a to the tag.

Phase 2.5 build (Session 8) — the HARD GATE is green. Locks 2.5a/2.5b/2.5c/2.5d verified with executed evidence (see the Phase 2.5 inventory); the mandatory adversarial pass ran (14 claims, 12 held, 3 breaches found+fixed: F-034/F-035/F-036); Beta Scope Decision produced as D-14 PROPOSED (docs/beta-scope-decision-brief.md) — the agent argues both cuts, Martin rules. A-09 (uncertainty_statement schema delta) PROPOSED. 130 tests green; all standing gates clean; Cargo.lock unchanged. Nothing in Phase 3 starts until Martin rules D-14 — and Phase 3 only happens under cut (B) or post-beta under (A). Merge of this PR additionally waits on the upstream source review (connector re-sync). Carried: reader process-separation live leg + calibration real-outcome data → Phase RC; the Phase 2 live legs (Neo4j 2a, Ollama 2c) still run on any open-egress host; 0e provenance/attestation → RC. Prior: Phase 0 CLOSED (PRs #7–#9); Phase 1 (PR #10); Phase 2 (PR #11); housekeeping (PR #14). Every fresh session: git fetch origin first; goose + Neo4j are container-ephemeral.