v8.4.0

• Add optional acceptance-criterion IDs. A criterion may be written - AC-01 (REQ-01): ... with
  one parent requirement; tasks, prompts, and EV-* records cite it in the existing Requirement anchors field, and an AC-* anchor implies its parent requirement. Optional Examples: use
  Given/When/Then only as notation. Features without AC-* behave exactly as in 8.3.0.
• Show per-criterion coverage in doctor --evidence-graph (text, --json acceptance_criteria,
  and --html) when a feature uses AC-*. New broken edges, each exiting 1: duplicate-ac-id,
  ac-unknown-requirement, unknown-ac-anchor, and no-ac-evidence. A line that resembles a
  criterion but does not match is reported as malformed AC line and never counts as coverage.
• Clarify contracts (no new Status value, EV-* field, contract kind, or dependency): sensor
  change classes infrastructure repair, automation repair, contract correction, and
  contract weakening in evidence-standard.md; a mock of the seam under test is a shallow
  sensor; check and review flag orphan behavior; a recorded RED states its cause
  (missing-behavior, test-defect, environment, unrelated-failure); acceptance discovery
  separates blocker from assumption, confirms prose before examples, and asks for a
  forbidden-outcome example on security, permission, or data criteria. Skills only cite these
  rules. The tdd baseline version is unchanged.
• Fix a gap in the 1.15.0 note: no script protected the stable false-positive class names. The
  contract check now fails when any of the ten false-positive classes or four sensor change
  classes leaves its reference section.
• Add nine behavior cases to the corpus and a seventh Level A fixture family, acceptance contract
  (F7), whose oracle compares a test's assertion lines so setup repair passes and an edited
  expectation is critical. No real-host campaign ran, so no behavior improvement or cost claim is
  made.
• Add docs/acceptance-driven-development.md and its pt-BR counterpart, and update the idea-to-spec
  golden to AC-* IDs while task-management stays REQ-only.

v8.3.0

• Add a maintainer skill-behavior instrument: 12 fixtures in six families (evidence before
  completion, sensor adequacy, resume freshness, authority, unknown effects, independent
  verification), each with a development and a reserved case grounded in existing corpus IDs and
  system invariants. Oracles read final disk state, sensor stamps, Git-ignored sentinels, and a
  hidden truth table; a test proves every oracle accepts compliant behavior and flags the scripted
  violation as critical.
• Rebuild npm run saf:sandbox-campaign on those oracles: an exit code is an observation, never a
  verdict; a forbidden effect fails even after a timeout; timeouts, budget stops, and missing
  capabilities never count as PASS. Cases come from the manifest, the report version follows
  package.json, a Claude Code adapter joins Codex and Cursor with isolated profiles, and an
  offline stub host exercises the pipeline without model calls.
• Add npm run saf:skill-behavior:compare, an offline baseline/candidate comparison that refuses
  unpaired, duplicated, or mixed data and never aggregates across hosts or models.
• Document the protocol in docs/skill-behavior-evaluation.md, correct docs/inspirations.md (SAF
  contract fields live in saf-contract.yml), update the ROADMAP direction, and clarify that
  rewind and rollback are host capabilities.
• Fix doctor ignoring its homeDir option when resolving user skill directories, installation
  intent, contracts, and autonomy checks.
• Run the first real skill-behavior campaign (12 cases × Claude Code 2.1.287 and codex-cli 0.162.1
  × 1 trial on the 8.3.0 bundle): 24/24 PASS. Pilots first exposed and fixed instrument defects:
  a Codex --sandbox/--approve-for-me conflict, closeout status parsing, Claude Code bundle
  placement (.claude/skills), and headless approval denials, now recorded as a missing
  capability. Claude Code runs Bash inside its own sandbox; positive claims on a real host need a
  sensor command in the harness-captured transcript, so a forged sensor stamp fails.
• Mark skill loading as manually verified for those host versions. One trial and no no-SAF control
  measure neither variance nor SAF's effect, and no skill instruction changed, so no behavior
  improvement or cost claim is made. No new command, flag, runtime dependency, telemetry, or
  public schema.

v8.2.0

• Restrict the privileged release workflow to a successful CI run caused by a push to this
  repository, so a fork pull request on a branch named main cannot reach it. The release job
  now installs with --ignore-scripts, pins its npm CLI, serializes runs, passes step outputs
  through env:, and every GitHub Action is pinned to a commit SHA (test/workflow-policy.test.ts).
• Make state durable: install intent, provenance, workspace config, project context, loop state,
  Git exclude blocks, and evidence graph HTML are written atomically; clean-upgrade staging stays
  beside each original with a cross-device fallback, so recovery works when the temporary
  directory is another filesystem; git and npm subprocesses have timeouts.
• Exit with 1, not 2, for expected state refusals such as missing Git metadata for project
  scope, an interrupted apply, unsupported intent, or unsafe provenance, as documented in
  docs/cli-interaction.md.
• Fix the rich uninstall --purge --plan hint to the canonical uninstall --yes --purge, accept
  the documented --preset aliases (man, assist, assisted, auto), and detect the branch of
  a freshly initialized repository without commits.
• Remove the environment override that could redirect the update check to another registry.
• Refactor the CLI entrypoint into a thin router with dedicated upgrade, welcome, output, and
  config-installation homes; remove unused exports and dead code.
• Add focused tests and src/ coverage floors; npm run check no longer builds twice.
• Extend provider-neutrality and shared-reference guards, require every source module in the
  architecture layout, block maintainer-local paths in publishable files, and wrap skill prose at
  100 characters without changing any obligation.
• Package contents: CHANGELOG.md, ROADMAP.md, and CONTRIBUTING.md are no longer shipped in the
  npm tarball; they remain in the repository.
• Sync the Portuguese README, skills usage guide, and language profile guide with English.

v8.1.2

• Fix duplicate Checks labels in human-readable doctor output.
• Expand black-box coverage for read-only, offline, non-interactive, and failure paths in the CLI
  update flow.
• Preserve the existing selective skill reconciliation and zero-runtime-dependency boundaries.

v8.1.1

• Make CLI upgrades pin the verified package version and reconcile only affected managed files.
• Preserve locally modified skills when package sources are unchanged, while updating skills when
  the published bundle changes.
• Clarify upgrade status, provenance, and read-only planning behavior across supported contexts.

v8.1.0

• Add bounded continuity for resumed tasks through compact v8 evidence-report references,
  canonical authority re-grounding, active-successor selection, and fail-closed freshness guidance.
• Protect the continuity behavior corpus with required IDs and a deterministic missing-case
  validator regression.
• Add the published-source TypeScript boundary to the normal typecheck and preserve local-first
  behavior, zero runtime dependencies, and human release authority.

v8.0.0

• Replace Evidence Graph v4 coverage with the Markdown-native v1 evidence contract, declared-input
  freshness, deterministic report supersession, and unified text, JSON, and HTML projections.
• Require explicit maintainer audit baselines and preserve legacy reports as visible context rather
  than current proof.
• Preserve local-first behavior, zero runtime dependencies, and human release authority.

v7.17.0

• Add deterministic internal evidence records with tree manifests, bounded sanitization, and
  fail-closed freshness classification for maintainer release evidence.
• Keep mechanical consumer-closure findings separate from semantic review findings and preserve
  explicit omission and invalidation states.
• Add direct, inherited, and not-declared provenance to the contract audit while validating
  unknown extends references and cycles without introducing a public schema or runtime registry.
• Preserve the 12-Skill roster, zero runtime dependencies, local-first behavior, and human release
  authority. Host execution and semantic authority review remain separate evidence limits.

v7.16.0

• Clarify Skill selection and conditional context, including authorized local PR preparation
  and repair, without changing remote publication authority or the 12-Skill roster.
• Strengthen observable obligations and enumerated requirement-to-task-to-evidence coverage
  within existing specification, prompt, check, and validation contracts.
• Add stable finding IDs, typed evidence, and resolution tables for local review and re-review;
  repair evidence remains subject to reviewer confirmation.
• Expand positive and negative routing examples for every Skill and add structural regression
  checks for narrative cases and review carry-over. These checks do not certify host behavior.

v7.15.1

• Fix interactive Team/project setup so user-only coding-agent targets are not passed to project
  installation, show the configured specs root, and preserve shared specs visibility through
  workspace initialization.