Skip to content

CHANGELOG.md

Latest commit

 

History

History
396 lines (323 loc) · 20.6 KB

CHANGELOG.md

File metadata and controls

396 lines (323 loc) · 20.6 KB

Changelog

All notable changes to this project will be documented here. Format follows Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[2.24.0] — simplifier agent (de-over-engineering) + adversarial-verify simplicity lens

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.

[2.23.0] — CAP-2 reviewer-prefs, history seeding, and doc/count refresh

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.

[2.22.0] — mutation-test suites for the gate cores + gate-integrity fix

Added

  • In-process pytest suites for manifest_gate.py (90.6% mutation score) and semantic_layer.py (97.4%), wiring both critical zones into the mutation policy. acceptance_gate.py mutation coverage at 77.6% (residual = documented equivalents).

Fixed

  • .craftsmanship.conf size-gate scoping: a test suite had set FILE_MAX=2000 globally, weakening the production 500-line gate. Now test-scoped (TEST_FILE_MAX/TEST_FILE_RE); production stays FILE_MAX=500.
  • Routing table regenerated to include the memory-writer agent.

[2.21.0] — mechanical craftsmanship enforcement + hook-layer correctness

Added

  • Craftsmanship checker (tools/craftsmanship-checker.sh + tools/lib/craftsmanship-detectors.sh). Mechanically enforces coding-standards.md §4 size limits and select structural rules. FILE_TOO_LONG (>500 lines) blocks; function/class/parameter/nesting block for recognized languages; grab-bag module names and layer-direction advise. Judgment rules (SRP/OCP/LSP/ISP, rule-of-three, dead-code) are deliberately NOT mechanized — a hook that fakes a verdict it cannot reach just trains you to ignore it. Mirrors zetetic-checker.sh (--staged/--files/--full, exit 0/1/2); wired into the commit/push hooks and a new CI job (hard on newly-added files, informational full-tree sweep).
  • Per-repo .craftsmanship.conf (.craftsmanship.conf.example). Every threshold and per-rule severity (block/advise/off) is team-tunable; defaults are the sourced §4 numbers. A CI drift-guard test pins the defaults to the §4 table in rules/agent-reference/craftsmanship-moves.md.
  • Regression suites tools/tests/hook-layer/ (87 cases) and tools/tests/threshold-drift/ (10 cases), run as a hard CI gate.

Fixed

  • Registry drift, at the root. scripts/setup.sh's hook merge was skip-if-present and silently let hooks/hooks.json and .claude-plugin/plugin.json diverge; it is now a content-equality re-sync. Restored the dropped stop-acceptance-gate.py Stop entry (plugin.json 17 → 18 hooks).
  • macOS hook no-op. 11 stdin-reading hooks used timeout 3 cat; stock macOS ships no timeout/gtimeout, so the payload was zeroed and the hooks silently did nothing. Replaced with a portable bounded read.
  • Fail-open holes. pre-tool-secret-shield.py, stop-acceptance-gate.py, and stop-context-guard.py raised on valid non-object JSON instead of failing open; guarded with an isinstance shape check.
  • secret-shield false blocks + a regression. No longer hard-blocks .env.example/.sample/.template templates or keyring/keychain source files; restored the printenv $SECRET / env NAME secret-read block. All true-positive secret blocks preserved.
  • git-verb guard bypasses. The commit/push gates were silently skipped by git push;, git commit&, (git commit), sudo git commit, env X=1 git commit, and GIT_EDITOR=… git commit. One unified anchor across all 5 hooks closes both classes with no new false positives.
  • claim-gate latency. A per-line subprocess loop (~38–70s on routine edits) is now a 3-pass scan (<1s).
  • Smaller hook fixes: notification-handler.sh pipefail crash, session-end.sh abort in a non-git directory, post-commit-difficulty.sh promiscuous grep, post-tool-error-routing.sh keyword anchoring, pre-edit-layer-check.sh advisory-honesty wording.

Changed

  • Mutation testing stays operationalized as test-engineer Move 8 (in coding-standards.md §3.2 and craftsmanship-moves.md); a real per-stack runner integration in the acceptance gate is a tracked follow-up — an inert draft was not shipped (no current caller, per §9).
  • The craftsmanship CI gate is a ratchet: newly-added files must fully comply; the legacy tree's pre-existing §4 debt (e.g. scripts/setup.sh) is surfaced informationally, not blocked, until refactored.

[2.20.0] — autonomous build loop + self-hosted knowledge ingestion

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).

[2.19.1] — fix Release workflow test paths

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).

[2.19.0] — genius corpus inherits the full tool set

Changed

  • All 97 genius agents drop the explicit tools: front-matter line. Each genius now inherits the full session tool set instead of pinning a hardcoded allow-list that had drifted from the live tool registry; pinning silently starved an agent of any tool the list omitted. Single uniform change — 97 files, 97 deletions, reasoning sections untouched.
  • Validated by a full isolation sweep before release: with the plugin disabled and an un-namespaced clone live, all 97 genius agents spawned and responded, each confirming file/search tools visible — 97/97, zero failures.

[2.18.0] — letta-code follow-up: lean genius corpus, compact routing, reflective checkpoints, memory contract hardening

Changed

  • R1 completed for the genius corpus: all 97 genius agents split into lean core + on-demand reference stubs (same two-tier move 2.17.0 applied to the 19 team agents). Doc-covered protocol sections deleted; memory/token-budget/ worktree replaced by parameterized stubs keeping every safety-critical invariant inline; uniform reference-docs index appended. 5,169,780 → 3,248,016 chars (37.2%, ~4.8K tokens saved per spawn per agent). Reasoning sections byte-identical before/after, asserted by the rollout script.
  • R2: routing never reads full files. New generated rules/agent-routing-table.md (~25KB, name + shape keywords + description for all 116 agents, from frontmatter via scripts/generate-routing-table.py) replaces full Reads of the 132KB INDEX.md in genius:route, genius:index and the orchestrator (~30K tokens saved per routing decision). pre-commit warns when the table is stale.
  • R3: checkpoint stubs follow the letta summary schema — goals / file references (paths + line ranges) / errors and fixes / current state / next steps, ≤500 words, tool outputs clipped to 2K chars, frontmatter description retrieval cue. Resume contract: checkpoint + ONE targeted recall, never re-reading what the checkpoint summarizes. All 116 agent token-budget stubs and the shared token-budget.md doc teach the schema.
  • R4: reflection at WARN, not at HARD. stop-context-guard.py's WARN firing is now a one-time blocking reflection (like letta's compaction event): the model spawns the new budgeted memory-writer agent (haiku, ≤16K context) to persist the semantic checkpoint + cortex:remember entries while headroom remains, then resumes the task — the HARD block becomes a formality.

Added

  • R5: mandatory description: frontmatter on memory .md files, enforced at the memory-tool.sh chokepoint on create/rethink (instructive error, MEMORY_NO_DESC_CHECK=1 test escape hatch). Contract §4.8.
  • R6: conflict-aware memory verbs — rethink <path> <text> [expected_sha] (atomic whole-file rewrite, letta memory_rethink) and sha <path> (CAS token); str_replace gains optional compare-and-swap. Contract §3.6b/§3.6c/§4.7; exposed via the memory_extensions MCP tool.
  • agents/memory-writer.md — single-purpose budgeted reflection scribe.

[2.17.0] — Lean team agents: core + on-demand reference docs

Changed

  • All 19 team agent definitions split into a lean core plus on-demand reference docs. Shared protocol detail (token budget, memory protocol and architecture, worktree protocol, effort calibration, codebase intelligence, dynamic workflows, mid-task system messages) moved to 8 docs under rules/agent-reference/, referenced from a uniform index table in every agent. Definitions shrink 987,888 → 625,267 chars (~36.7%, ~5.0K tokens saved per spawn per agent). Inline stubs remain self-sufficient for safety-critical invariants (checkpoint thresholds, memory scoping, worktree commit rules); reference docs are elaboration and recovery material, validated via headless fresh-session runs.
  • Agent frontmatter parameterized (agent_topic, memory_scope, model); Haiku agents carry 170K/~120K budgets plus an escalate-to-orchestrator line, Opus agents 200K/~180K. Fixed latex-engineer/professor MEMORY_AGENT_ID=haiku bug (now agent name) and the orchestrator's dangling <dynamic-workflows> prose reference.

[2.16.0] — Per-model context thresholds via shared config

Changed

  • stop-context-guard.py re-vendored from session-optimizer v1.1.0. Thresholds are now per-model and loaded from ~/.claude/ctxguard-thresholds.json (embedded fallback when the config is absent or malformed; first substring match on the lowercased model id wins): Fable 5 / Mythos warn 120K hard 160K (2x Opus rates — carrying rent and the 5-min cache-expiry resume penalty bite twice as hard), Haiku 4.5 warn 120K hard 170K (200K IS the window; leave headroom for the checkpoint turn), Opus/Sonnet warn 180K hard 200K (cost discipline; window is 1M). The hard-cap block message now reports the per-model budget instead of a fixed 200K.
  • <token-budget> model-limits table updated in all 117 agent docs. Adds the Claude Fable 5 row (160K hard cap, ~120K checkpoint), corrects Haiku's cap to 170K, and points to ctxguard-thresholds.json as the authoritative source shared with the statusline and the Stop guard.

Added

  • hooks/ctxguard-thresholds.json — vendored copy of the shared threshold config. session-start.sh seeds it to ~/.claude/ctxguard-thresholds.json when absent (idempotent — never overwrites user edits, so tuned thresholds survive plugin updates).

[2.15.0] — Complete the plugin hook manifest

Fixed

  • Missing hooks in the plugin manifest. .claude-plugin/plugin.json had drifted from hooks/hooks.json: the inline hooks block omitted three hooks that the canonical wiring defines, so they never registered when the plugin loaded — pre-tool-secret-shield.py (PreToolUse), stop-context-guard.py (Stop, the context-budget guard from session-optimizer), and session-end-memory-drain.sh (Stop). The manifest now mirrors hooks/hooks.json exactly (17 hook commands). The Stop block fires all three lifecycle hooks; PreToolUse re-includes the secret shield.

Changed

  • Marketplace entry hook count corrected (16 → 17).

[2.14.0] — Public-readiness baseline

Added

  • Public-readiness baseline: CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md.
  • GitHub issue templates (bug / feature / audit-finding) and PR template with audit-cycle checklist.
  • prd-spec-generator row in the companion-projects table.

Changed

  • LICENSE copyright corrected to Clément Deust (sole independent author); ecosystem-context preamble + explicit non-affiliation statement added.
  • LinkedIn post first-comment options refined for algorithm-aware reach.

[2.13.1] — Tier-1 visibility + memory MCP + PII scanner

Added

  • Memory MCP. Local replica of Anthropic's managed-agent memory_20250818 tool with scope-based ACL, queue isolation, and full MCP wire compatibility. 241 tests passing across functional, ACL, concurrency, stale-lock, MCP, and PII suites.
  • PII / secret scrubbing on memory write path (contract §7.2).
  • pre-tool-secret-shield hook — blocks any agent from reading .env, .aws/credentials, *.pem, *.key, or shell-history files.
  • PII scanner daemon. Persistent process eliminates Python cold-start; median scan time reduced 34→8 ms.
  • Memory contract on every agent. memory_scope frontmatter + memory body block added to all 19 team agents and all 97 genius agents (so each agent declares what it persists and where).
  • README rewrite (Tier 1 visibility), 6 supporting docs, full CI matrix, Codespaces config (subsequently removed per cross-check feedback).

Changed

  • CI concurrency suite made Linux-portable (was macOS-specific).

Documentation

  • LinkedIn post series introducing zetetic (rewritten in plain prose; no em-dashes).

For older releases (v2.13.0 and earlier), see git history. The project predates this CHANGELOG; pre-2.13.1 versioning was driven by tag-only release notes on GitHub.