Skip to content

Latest commit

 

History

History
306 lines (251 loc) · 18.8 KB

File metadata and controls

306 lines (251 loc) · 18.8 KB

Changelog

All notable changes to this package are recorded here. The format is based on Keep a Changelog, and the package follows Semantic Versioning.

The package version tracks the directive set as a whole. Each authority text also carries its own version inside its header. The versioning rule (see MANIFEST.md): a change to an authority text bumps the package minor or major; a companion-only edit bumps the patch.

1.5.0 - 2026-06-10

Cuts Agentic Architecture Audit Specification v3.5 (was v3.4) and Project Profile Discovery Directive v1.6 (was v1.5) — the second dual-authority cut. Sources the proposals queued by the 2026-06-10 cross-application review (ADR 0003), each premise adversarially re-verified against the v3.4/v1.5 texts before drafting; one proposal (the profile JSON Schema) remains queued as a post-cut package patch. All changes additive; minor package bump per the versioning rule. Companions, MANIFEST.md (including the content-hash binding re-attestation), README.md, AGENTS.md, and CITATION.cff are synced; the drift linter verifies the sync. Rationale recorded in adr/0004-evidence-discipline-and-enforcement-honesty.md.

Agentic Architecture Audit Specification — v3.4 → v3.5 (2026-06-10)

  • Per-cycle id durability (Phase 10, §10.2; optional cycle field in §8.11/§8.12): ids are monotonic-never-reused across cycles or cycle-qualified wherever cited in artifacts that outlive their cycle; the anchored in-run id patterns are unchanged and dated cycle artifacts keep bare ids.
  • Typed citation provenance (§8.1): optional observed_at, valid_until, and evidence_lane (repo-local | host-verifiable | unavailable) extending snapshot_ref/current_ref; Phase 10.5 re-verifies through the lane and records not-run for unavailable.
  • Enforcement-honesty flags (Phase 4, §8.5): claimed-automation-absent — governance/instruction text claiming automated enforcement no implemented gate backs — and instruction-reference-broken — instruction contracts citing repository artifacts that do not resolve in the tree.
  • Review-power proportionality (§9.9, Phase 6, §8.7 review-power-inversion): ordered enforcement-power (blocks-merge > blocks-commit > warn-only > prose-only) and mandated-review scales with a decidable pairwise inversion condition.
  • regression-trap eval-suite mode (Phase 9, §8.10): behavioral traps seeded from observed agent failures, with the existing demotion rule applied — an unharnessed trap corpus is manual-review-only, not coverage.
  • Metalanguage containment (Prime Directive 16, Phase 1, Phase 5, §8.2 audit-taxonomy-collision): audit-internal taxonomy is never exported into a project's controlled vocabulary; machine-readable schema fields stay exempt.
  • Operating-agent baseline drift (Phase 6, §8.7 model-alias-resolution-drift): where the profile records CLI versions, model posture, and alias pins, the audit confirms each pin still resolves to the recorded baseline.
  • Agent-memory truth maintenance (Phase 5, §8.6 optional truth_maintenance, prose flag memory-truth-maintenance-absent): whether agent-held claims have a correction path when repository facts change.
  • Negative-self-test rule for implemented fitness functions (Phase 10 step 8, §8.12 optional negative_self_test, §11.11 note) and a standing-decision-ledger evidence target (§11.10 note).

Project Profile Discovery Directive — v1.5 → v1.6 (2026-06-10)

  • Adds conventions.forbidden_terms (a structured projection of vocabulary-banning rules), an optional agent_surface.operating_agents block (observed agent CLI versions, model posture, and model-alias pins with resolutions and observation dates, citing the committed agent configuration), and regression-trap in the evals baseline — with matching Phase D/E/F discovery steps and §7 validation rules. Purely additive: v1.5 snapshots need no migration.

Also included (companion/tooling, shipped unreleased on main before this cut)

Two patch batches (PRs #11 and #12), both sourced from the 2026-06-10 cross-application review (run against an independently governed agentic repo, "HCS"); the report itself stays untracked per MANIFEST.md ("Not bundled"). Rationale and the full disposition record in adr/0003-cross-application-review-dispositions.md.

Pre-cut gate hardening (2026-06-10, follow-through on ADR 0003):

  • Drift linter, three new commit-blocking rules, each shipped with a negative self-test wired into --self-test: bare per-cycle FF-/F- ids flagged on every tracked surface except the authority texts, CHANGELOG.md, adr/, and examples/ (mechanizes the identifier convention); MANIFEST.md content-hash binding of the three derived files to the current authority texts (the review's P-013; --print-bindings regenerates the lines, cut-time re-attestation documented in CONTRIBUTING.md); and AGENTS.md must quote the directive's current directive_version/audit_spec_target schema-identifier literals, forcing the instruction file to update with any cut that changes either literal. MANIFEST.md's automation claims updated to state exactly what is and is not gated.
  • DECISIONS.md standing decision ledger (the review's P-012, part (a)): durable D-NNN ids seeding the thirteen dispositioned proposals with their current statuses (restated in package terms), the two rejected review defects with reasons, and the open decisions; the restatement-policing check is explicitly recorded as not built.
  • Inline (per AGENTS.md) cites on the hook-bypass-ban restatements in companions/kickoff-prompt.md, .github/pull_request_template.md, and CONTRIBUTING.md (closes the residual the review's R-004 left behind).
  • ADR 0003 marked Accepted (operator-approved; its patch scope landed via PR #11).

Review dispositions (2026-06-10, landed via PR #11):

  • Identifier convention adopted (weak form of the queued P-001): self-audit FF/F ids are per-cycle and are now cycle-qualified on live, undated surfaces — README.md, .pre-commit-config.yaml hook display names (hook ids unchanged), CONTRIBUTING.md, and scripts/check_drift.py display strings (e.g. 2026-06-07/FF-004, 2026-06-08/FF-001). Convention recorded in MANIFEST.md. Resolves the review's confirmed defects R-001/R-002.
  • Two review defects refuted on adversarial re-verification and recorded in ADR 0003 so they are not re-raised: R-003 (.codex/ is gitignored personal scaffolding, never shipped) and R-004 (the --no-verify ban has eight statements, not five, with AGENTS.md as the named canonical home).
  • Review-input handling: .gitignore now excludes /improvement-report-*.md; the convention (untracked input evidence, transcribed via ADR) is documented in AGENTS.md.
  • Instruction-file currency (AGENTS.md, CLAUDE.md): the stale directive_version example literal updated to the v1.5 value and the pinned audit_spec_target v3.1 consumption baseline explained (do not "fix" it as drift); drift-linter behavior, scope, and commit-blocking rules documented; the .agents/skills/run-agentic-audit/SKILL.md derived surface, ADR convention, release/provenance steps, and published repo slug recorded.
  • Queued for a future deliberate cut (Audit Spec v3.5 / Profile Directive v1.6): thirteen additive proposals, dispositioned in ADR 0003's table — shipped by this release's authority cut, except the profile JSON Schema (still queued as a post-cut package patch) and the content-hash binding (already shipped in the gate hardening).

1.4.0 - 2026-06-09

Cuts Agentic Architecture Audit Specification v3.4 (was v3.3) and Project Profile Discovery Directive v1.5 (was v1.4) — the first dual-authority cut. Driven by a verified standards-currency review (2026-06-09) of the mid-2026 agentic landscape: MCP's task/extension/ deprecation lifecycle, the cross-tool Agent Skills standard, agent identity and delegated commerce, and the customization supply chain. All changes are additive; minor package bump per the versioning rule. Companions, MANIFEST.md, README.md, and CITATION.cff are synced; the FF-004 drift linter verifies the sync. Rationale recorded in adr/0002-skills-as-contracts-and-protocol-object-lifecycle.md.

Agentic Architecture Audit Specification — v3.3 → v3.4 (2026-06-09)

  • Agent-skill packages become first-class contracts and authority surfaces (§5.1, §6.3, Phases 4 and 6, §8.5, §8.7, §9.1, §11.3, §14): frontmatter contracts, capability pre-approvals (e.g., allowed-tools), bundled executables, and source/marketplace provenance are inventoried and flagged — prose instructions remain governance, checked in Phase 8.
  • Protocol-object lifecycle status is recorded against the pinned protocol revision (Phase 0, Phase 4, §8.5 protocol_object_status, §14), so the inventory tracks deprecations and extensions instead of hard-coding an object list; protocol tasks, extension declarations, server-delivered UI surfaces, and registry/discovery metadata become inventoryable objects.
  • Agent identity and delegated commerce join Phase 6 and the §9.7 authority matrix (commerce checks apply only where an agent can initiate or approve payments).
  • Customization supply-chain checks (skills/hooks/plugins/protocol-server configuration consumed by repo-operating agents) join the agent-operability lens; the §11.6 isolation baseline is calibrated (OS-level or micro-VM sandbox, write-allowlists, default-deny egress).
  • Hosted-eval platform-lifecycle flag (Phase 9); refreshed §14 reference anchors.
  • Additive only: v3.3 audits remain valid; v1.4 profiles are consumed unchanged.

Project Profile Discovery Directive — v1.4 → v1.5 (2026-06-09)

  • Adds agent_surface.skills, MCP task/extension/server-delivered-UI/deprecated-objects fields, a commerce protocol-surface block, and committed repo-operating-agent configuration + agent-identity-attestation fields in the authority baselines, with matching Phase D/E/F discovery steps and validation rules. Purely additive: v1.4 snapshots need no migration.

Also included (companion/tooling, from the 2026-06-08 self-audit follow-ups)

Resolves the three remaining findings from the 2026-06-08 self-audit (shipped unreleased on main before this cut):

  • Drift linter (scripts/check_drift.py): the section-reference check now resolves each reference against the specific authority text its line names, instead of accepting any number present in either text. This closes self-audit F-001 (FF-001), where a reference attributed to the profile directive but naming an audit-spec-only section passed. Correctly-attributed and unattributed references are unaffected.
  • scripts/check_drift.py --self-test, wired as its own pre-commit hook, locks in the FF-001 negative test so the F-001 regression cannot return silently.
  • CONTRIBUTING.md and SECURITY.md (self-audit F-003 / FF-003): contribution and release workflow, and security scope with private vulnerability reporting.
  • A release-provenance verification discipline (self-audit F-002 / FF-002), documented in CONTRIBUTING.md and pointed to from MANIFEST.md: release claims must separate repo-local, GitHub-hosted, and unavailable evidence lanes.

1.3.0 - 2026-06-08

Cuts Agentic Architecture Audit Specification v3.3 (was v3.2). Additive: adds the agent-operability strategic-theme lens (§2.3, §11.12) — it reweights existing dimensions (§11.3 / §11.6 / §11.10 / §11.11) at 1.5× and adds no scored dimension, preserving the min-aggregation rule — plus a repo-operating-agent authority clause (§11.6), a developer-agent-workflow scope boundary (§0), and Appendix D. The rationale (a theme, deliberately not a scored dimension) is recorded in the package's first ADR, adr/0001-agent-operability-as-theme.md.

No phase or scoring change for existing surfaces, so v3.2 audits remain valid. The Profile Directive stays v1.4, consumed unchanged. Companions, MANIFEST.md, and CITATION.cff are synced; the FF-004 drift linter verifies the sync. Minor package bump per the versioning rule.

Agentic Architecture Audit Specification — v3.2 → v3.3 (2026-06-08)

Resolves self-audit F-008 by adopting agent-operability as a theme lens rather than a scored dimension — avoiding the orthogonality / min-aggregation cost a dimension would impose. Surfaced a separate confirmed finding F-012 (MANIFEST over-claims check_drift.py's section-number coverage), tracked as its own workstream.

1.2.0 - 2026-06-07

Cuts Agentic Architecture Audit Specification v3.2 (was v3.1). Additive change: recognizes AGENTS.md as the canonical cross-agent instruction standard, with tool-specific files (CLAUDE.md, GEMINI.md, .cursor/rules) treated as bridges that should import or defer to it. Phase 4 now inventories this cross-agent instruction contract and flags cross-agent-instruction-drift; the Contract schema (§8.5) and Contract-discipline rubric (§11.3) gain the corresponding surface, flag, and scoring language.

No phase was removed and no behavior changed for existing surfaces, so v3.1 audits remain valid. The Profile Directive stays at v1.4 and its profiles are consumed unchanged — its spec references are left at their v3.1 baseline, which is additively compatible. Companions, MANIFEST.md, and CITATION.cff are synced; the FF-004 drift linter verifies the sync. Minor package bump per the versioning rule.

Agentic Architecture Audit Specification — v3.1 → v3.2 (2026-06-07)

Additive: first-class cross-agent instruction contract handling. Surfaced by F-007 from running the audit on this package itself — the spec had treated AGENTS.md as interchangeable governance rather than the AAIF/Linux-Foundation canonical standard with a canonical↔bridge structure.

1.1.0 - 2026-06-07

Cuts Project Profile Discovery Directive v1.4 (was v1.3). Additive vocabulary change: adds a documentation value to the project.kind enum so documentation, specification, and methodology packages classify cleanly instead of being coerced to library. No schema fields added or removed and no validation behavior changed, so a v1.3 profile snapshot needs no migration.

The Agentic Architecture Audit Specification is unaffected and stays at v3.1. Companions, MANIFEST.md, and CITATION.cff are synced to the new directive version; the FF-004 drift linter (scripts/check_drift.py) verifies the sync. This is a minor package bump (an authority-text change that is backward-compatible) per the versioning rule.

Project Profile Discovery Directive — v1.3 → v1.4 (2026-06-07)

Additive: project.kind gains documentation. The gap was surfaced by running the audit on this package itself — a documentation/methodology package had no fitting project.kind value and was coerced to library.

1.0.0 - 2026-06-06

First public release. No code or directive behavior changed from the internal package; this release renames the files to stable names, reconciles the internal cross-references to those names, and adds the public-repo scaffolding (license, citation, manifest, examples, hygiene baseline).

Contents at this release:

  • Agentic Architecture Audit Specification v3.1 (2026-05-08) — audit-spec.md
  • Project Profile Discovery Directive v1.3 (2026-05-23) — profile-directive.md
  • Companions — companions/explainer.md, companions/kickoff-prompt.md
  • Drift-control manifest — MANIFEST.md

File renames from the internal package (content preserved):

  • agentic-audit-spec-v3.mdaudit-spec.md
  • project-profile-directive.mdprofile-directive.md
  • audit-spec-friendly-explainer.mdcompanions/explainer.md
  • audit-kickoff-prompt.mdcompanions/kickoff-prompt.md
  • audit-directive-set-manifest.mdMANIFEST.md

Authority-text lineage

The package's first public version is 1.0.0, but the authority texts have their own history. The relevant lineage carried into this release:

Project Profile Discovery Directive — v1.2 → v1.3 (2026-05-23)

Validation clarifications only; no schema fields added or removed, so a v1.2 snapshot needs no field migration. The audit specification was unaffected and stayed at v3.1.

  • Citation fidelity is enforced: a cited evidence excerpt that names a symbol or quotes text must actually appear at the cited lines, and a single-line citation must point at the named construct, not an adjacent comment, import, or enclosing block.
  • Phase H performs a content-level evidence check, not just path-exists/range-in-bounds, and records the verification method and any corrected mismatches.
  • Dependency versions are reconciled from lockfiles when a lockfile is in scope, rather than cited as manifest ranges.
  • model_providers[].access_via is checked for consistency with the hosted/local authority baseline, with added guidance for mapping local model execution onto the enum.

Prompted by a real first-profile run that passed path/range validation yet still shipped two evidence excerpts naming classes absent at the cited lines and one off-by-many line number — the exact failure mode the audit's "do not trust profile claims" rule exists to catch downstream, now also caught upstream at profile time.

Agentic Architecture Audit Specification — v3.0 → v3.1 (2026-05-08)

Preserves the v3 audit philosophy and adds targeted coverage for:

  • protocol-specific contract surfaces, especially MCP and A2A;
  • workflow descriptions, overlays, schema dialects, and protocol versions;
  • background, paused, resumable, durable, callback-mediated, and event-triggered execution;
  • authority matrices covering approval mode, precedence, bypass modes, protected paths, secondary credentials, callbacks, hosted/local boundaries, and token delegation;
  • separate state and memory classes;
  • separate runtime/action, content, and build/source provenance classes;
  • semantic-convention version and stability pinning;
  • server-exposed prompts and privileged-context injection boundaries;
  • eval coverage for protocol surfaces, approval paths, async lifecycle, and memory lifecycle.