v2.28.1

Fixed

• agent-definition-auditor.sh now actually gates in CI — the script's default ROOT was a hardcoded personal absolute path that didn't exist on the CI runner, so the glob silently matched 0 files and the whole audit (F1-F9, FD, B1-B3, G1-G3, P1) passed vacuously. ROOT now resolves relative to the script's own location, and a 0-match glob fails explicitly (exit 2) instead of reporting an empty pass. All 97 genius agents also gained an explicit tools: frontmatter field (previously scoped only via the plugin manifest's blanket "All tools" grant), bringing F8/F9 from 22/119 to 119/119 passing.
• 26 genius agent descriptions repaired — 4 agents (deming, fisher, leguin, ranganathan) shipped with frontmatter description truncated mid-sentence by a JSON double-encoding bug (e.g. "W."); a further 10 had the same corruption undetected until a new quality gate (check FD: description ≥ 40 chars, no broken escape artifacts) was added to the auditor, and 16 more had valid-but-terse descriptions. Since description is the spawn/routing criterion, corrupted values made these agents unroutable; rules/agent-routing-table.md was regenerated so the fix reaches routing consumers.
• Orchestrator now validates a subagent's result against its artifact contract before forwarding it downstream, and isolates a failed subtask so it does not invalidate already-validated independent results — encoded as new Move 6 steps and refusal conditions.
• Encoded the anti-passive-waiting lesson (long-running work is foreground-blocking or terminate-and-handoff, never a sleep/poll loop on a background monitor) as binding §8c in the shared memory contract.

v2.28.0

No CHANGELOG entry found for 2.28.0; using tag message

v2.27.0

Added

• git-historian team agent — regression provenance and abandoned-approach recovery: traces when a behavior changed, which commit introduced it, and surfaces approaches tried-and-reverted so a session doesn't re-walk a dead end.

Changed

• Registered the cortex-viz memory scope in the central registry so cross-agent recall resolves it.

Fixed

• Corrected get_impact drift in rules/agent-reference/codebase-intelligence.md — the doc over-promised the tool's blast-radius (it returns one reverse-edge hop, not transitive/test-aware); wording now matches the implementation.

v2.26.0

Changed

• Spine beat 2 (evidence/sources) now enforces the direction of evidence,
  across all 117 agents (team + genius) via the single generator
  scripts/generate-spine.py and its elaboration
  rules/agent-reference/zetetic-spine.md. The prior wording — "every claim traces
  to a source" — was direction-agnostic: it was satisfied equally by rigor (read a
  source, derive the code from it) and by fabrication (write the code, then find a
  resembling paper and attach the citation as post-hoc justification). Scientific
  rigor is source-first: start from a factual, verifiable, demonstrable source and
  produce the implementation from it. A citation attached after the fact to a
  lookalike paper is fabricated proof, not evidence — even when the citation is real,
  because the code was never derived from it.
• Beat 2 now names three refused failure modes, backed by observed provenance audits
  (invented arXiv ids, fabricated benchmark deltas, constants mislabeled with
  unrelated papers): retrofitted citation (added after the code, chosen for
  resemblance), resemblance ≠ prescription (a paper about your topic is not a
  source for your specific value/equation unless it states it and its experimental
  conditions match yours), and borrowed authority (a hardcoded constant wearing a
  paper's name with no line the paper actually prescribes). The honesty discriminator
  is the provenance comment written before the value plus an explicit
  "engineering-default / hand-tuned" disclaimer when a value is not paper-prescribed —
  never the citation alone.
• No source → "I don't know" and stop is now explicit that this is a gate failure,
  not a formatting gap to fill later: do not ship, then justify.

Notes

• Content-only change to the injected <zetetic-spine> blocks; no tool, hook, or
  memory-contract surface changed. scripts/generate-spine.py --check is clean
  (idempotent) and all seven release-gate memory suites pass locally.

v2.25.0

No CHANGELOG entry found for 2.25.0; using tag message

v2.24.0

Added

• New simplifier team agent (agents/simplifier.md). De-over-engineers code
  that already works and breaks no hard rule but carries more complexity than its
  problem requires — premature abstraction, needless indirection, speculative
  generality, premature optimization, and drifted duplication. Distinct trigger from
  refactorer: refactorer fixes hard-rule violations; simplifier removes
  superfluous complexity that violates no hard rule (already-functional,
  already-compliant code never invokes refactorer, so over-engineering needs its own
  trigger — folding both into one agent would be an SRP violation in agent design).
  Behavior-preserving (tests pass before and after, one simplification per commit),
  stakes-calibrated, and language-/project-agnostic (idiom-mapping across Python,
  TypeScript, Go, Rust, Java, Swift). Its over-engineering heuristics are an open,
  non-exhaustive catalog (YAGNI, rule-of-three, needless indirection, premature
  optimization, speculative generality, drifted duplication, plus KISS, Gall's Law,
  Ousterhout shallow modules, dead code, boolean blindness, …) — each removal must
  name and source its principle per the zetetic §8 standard. New simplifier memory
  scope (memory/scope-registry.json).
• A2 / CR-4 — adversarial-verify pre-verdict workflow now carries a fifth,
  perspective-diverse lens: simplicity (agentType simplifier), prompted to
  REFUTE by hunting superfluous complexity that no hard rule forbids. Joins the four
  existing refute lenses (residual-fp, missed-cases, robustness, test-adequacy);
  synthesis stays deterministic and fail-closed. (The A2/adversarial-verify work
  shipped in code prior to this release without a CHANGELOG entry; recorded here.)

Changed

• code-reviewer Move 5 now hands off superfluous-complexity smells (over-
  engineering with no hard-rule violation) to simplifier as an advisory note,
  distinct from the §4 size-breach blocking gate. Output format and blind-spot
  hand-off list updated accordingly.
• README badge, body, CONTRIBUTING.md, and the marketplace plugin/metadata
  descriptions refreshed: 97 reasoning patterns + 21 team agents (118 total).
  rules/agent-routing-table.md regenerated (118 agents).
• memory/scope-coverage.md: simplifier added to the team table; counts reconciled
  to 20 scope-owning team agents, 117 tabulated agents, 27 distinct registry
  scopes (7 systemic + 18 team + 1 research + 1 genius). This doc tabulates only
  scope-owning agents, so its total is intentionally one less than the 118 agent
  definition files the README/marketplace count — memory-writer is a scribe that
  owns no scope. A new footnote in the doc records the distinction.

Tests

• tools/tests/adversarial-verify/core.test.mjs updated for the five-lens core
  (lens count, lens-keys ordering, distinct-agentType count, five-element fixtures).
  13/13 deterministic synthesis-core tests pass.

v2.23.0

Added

• CAP-2 — adapt to a demanding team lead. New reviewer-prefs memory scope
  (memory/scope-registry.json) holding a lead's standing review preferences.
  Owners are the lead (_user) and the orchestrator/curator; the reviewed agents
  (engineer, refactorer, code-reviewer) are readers only, so an agent cannot
  invent its own prefs. The three agents read the scope at the top of their workflow
  with a fixed precedence — a coding-standards.md blocking rule always outranks a
  preference — and a graceful fallback when the scope is absent.
• tools/seed-reviewer-prefs.sh — bootstraps a lead's prefs from evidence of how
  they already work (their gh review comments + merged PRs), emitting a
  provenance-tagged status: inferred draft. It presents raw evidence only;
  abstracting it into confirmed preferences is the lead/orchestrator's judgment step
  (§9 — no faked judgment). Injection-safe, owner-only writes, deterministic exit
  codes. 19-case test suite under tools/tests/seed-reviewer-prefs/.

Changed

• README badge, assets/banner.svg, CONTRIBUTING.md, and the marketplace plugin
  description refreshed to the current reproducible counts: 64 skills, 18 hooks,
  35 tools, 25 commands, 97 reasoning patterns + 20 team agents (117 total),
  288 tests.

Fixed

• memory/scope-coverage.md miscount: the checkpoints systemic scope was never
  tabulated. Reconciled to 26 scopes (7 systemic + 17 team + 1 research + 1 genius);
  the doc's embedded verification command now matches.

v2.22.0

No CHANGELOG entry found for 2.22.0; using tag message

v2.20.0

Added

• Closed-loop autonomous build (.claude/workflows/autonomous-build-loop.js).
  Drives a build task to a candidate on an isolated iteration branch:
  refine → plan → verify-plan → orchestrator build → best-effort in-loop
  acceptance checks → iterate until green or the budget is spent. Repo-generic —
  repoPath, gateRunner, and an optional base gateConfig are inputs, so the
  loop runs from any working directory against a repo that need not contain this
  tooling. It drafts and converges a candidate; it does not self-certify.
• Deterministic acceptance gate (tools/acceptance_gate.py,
  tools/acceptance-gate.sh). Runs configured command gates and aggregates
  their exit codes — a gate passes iff its command exits 0, never a model grading
  its own output (Huang et al., arXiv:2310.01798). Gates any repo via --root,
  evaluates the committed tip via --diff-base/--diff-head in a throwaway
  worktree, rejects an empty diff, and fails closed. Unit-tested under
  tools/tests/acceptance-gate/ (incl. external-repo-from-a-foreign-cwd and
  fail-closed cases).
• Real-exec Stop-hook gate (hooks/stop-acceptance-gate.py) — the build
  loop's gate component (invoked by the workflow, not registered as a global
  lifecycle hook).
• Self-hosted web ingestion engine (tools/web_ingest.py,
  tools/web_extract.py, tools/web-ingest.sh). A dependency-free replica of
  Firecrawl's self-hostable core (scrape / map / crawl): stdlib fetch,
  robots.txt respected, main-content markdown extraction, conditional-GET
  caching. No web search and no LLM extraction by design; TLS verification is
  never disabled.
• Query-indexed semantic layer over Cortex (tools/semantic_layer.py,
  tools/semantic-layer.sh, memory/semantic-layer.yaml,
  memory/semantic-layer.schema.yaml). A YAML index keyed by query + intent
  (ingest / verify / compare / monitor) with fresh / stale /
  superseded states. Never writes Cortex itself — the agent owns the write and
  passes back the cortex_id as a pointer.
• Manifest-membership gate (tools/manifest_gate.py, tools/manifest-gate.sh,
  tests under tools/tests/manifest-gate/). Fail-closed grounding check: every
  fact's source must be a URL the web-ingest engine actually fetched this
  session. Complements the semantic layer's presence check with a membership
  check; a pure stdin → stdout filter.

Changed

• README: broader refresh. Corrected counts to ground truth — agents
  116 → 117 (a 20th team-role agent, memory-writer), skills 63 → 64; documented
  the autonomous build loop and the new knowledge-ingestion / semantic-layer
  subsystems; refreshed the genius-trigger source comment (recounted 2026-06-23).
• Marketplace manifest counts refreshed (20 team agents, 64 skills,
  25 commands, 24 logical tools, 17 registered hooks).

Fixed

• web-ingest follows 308 redirects and records the final resolved URL.
• Manifest-gate comments reworded to satisfy the absolute-claim checker (§8).

v2.19.1

Fixed

• Release workflow no longer references a non-existent tests/run-all.sh.
  The tag-time release.yml invoked bash tests/run-all.sh, which has never
  existed, so every tagged release failed at the test step with exit 127. It
  now runs the same suite as ci.yml (the scripts/test-memory-*.sh and
  scripts/test-agent-id-propagation.sh suites) plus the structural auditor.
• Corrected the zetetic-checker invocation from the unsupported --tree
  flag to --full; valid modes are --staged, --files, and --full.
• Added jq to the release job's apt dependencies (required by the memory
  suites, matching CI).