CHANGELOG.md

Changelog

All notable changes to Quaere are documented in this file. The format follows Keep a Changelog and the project uses Semantic Versioning.

Unreleased

Added

• quaere-prospect extension — a 0→1 origination skill for deciding what to build next before any problem is chosen: survey the terrain and its stated non-goals, inventory gaps each tied to a named beneficiary and job, run a mandatory reality gate (the capability may already exist or be a deliberate non-goal), and design a validation probe before any build — classifying each opportunity as verified gap / assumed gap / already covered / wishlist so origination does not collapse into a generic feature wishlist. Sits before quaere-invention in the pipeline (prospect → invention → grounding/evidence → execution). Opt-in: quaere install --skill prospect.
• quaere-crucible extension — a Quaere-style "Grill Me": an evidence-gated adversarial interrogation that pressure-tests a plan, claim, design, PR, or the agent's own proposed plan at the commitment boundary. It decomposes the target into load-bearing claims, grills the most decisive one with a mandatory falsifier and alternative-hypothesis question, and refuses to bless it until each claim survived on new evidence or a defended rebuttal — never on confident-but-unverified pushback — or is logged as an explicit unresolved gap. Returns a terminal verdict (blessed / blessed-narrowed / blocked / unresolved gap), never "looks fine"; intensity is calibrated to stakes. Opt-in: quaere install --skill crucible.

0.6.0 — 2026-06-13

Adds the quaere-naming extension and tightens the existing skills: every prescriptive gate now sits within Codex CLI's ~220-line SKILL.md read cap, the certainty labels (confirmed / confident / locally novel) can no longer be claimed without the probe that earns them, and the eval harness gains an outcome-axis judge on every scenario plus a not_regex assertion type. The CLI interface is unchanged; the published headline numbers still come from the v0.3.1 sweep and predate these skill-body changes.

Added

• quaere-naming extension — a metaphor-driven naming skill: naming brief before any name, conceptual territories instead of thesaurus synonyms, anti-pattern filtering, and a mandatory tool-verified availability gate (never from memory). Opt-in: quaere install --skill naming.
• Two semantic eval fixtures — no-corroborator (no tests, comments, history, or spec available; the honest output is plausible / UNKNOWN, not invented intent) and comprehension-quality (grades the depth of per-unit analysis, not just its shape).
• --isolate-skills eval flag — relocates installed skills for the duration of a sweep so a baseline run cannot silently read them; without it, "baseline" numbers are contaminated by whatever is already installed.
• Outcome-axis judge on every scenario — all 22 eval scenarios now carry at least one llm_judge assertion tagged axis: outcome (was 5/22), so the paired baseline-vs-skill comparison has an outcome signal everywhere, not just form checks.
• not_regex assertion type — negative regex matching for the eval harness; four previously vacuous safety assertions (regex strings passed to the literal not_contains) were converted and now actually fire.

Changed

• quaere-semantic distilled to its measured active core — 256 → 148 lines vs v0.3.1, keeping only the content that ablation showed carries the outcome.
• confident / locally novel certainty labels now require a probe. quaere-semantic grants confident only when a named corroborator (test, caller, git blame, spec, ADR) was actually consulted, and quaere-invention grants locally novel only when the candidate carries a falsifiable kill-probe — neither label can be earned by self-assessment.
• All prescriptive content now sits within the Codex read cap. quaere-evidence compressed 322 → 260 lines (workflow steps 5–9, including the patch-authorization gate and Decision labels, previously sat past line 220 and were structurally unread on the Codex path); quaere-audit gained an in-cap 7-condition Confirmation Rule summary (conditions 2–7 previously truncated mid-list); quaere-grounding gained an in-cap no-network-fallback compact form and a deliverable-template pointer. Only the sanctioned full Handoff/Stop sections and reference pointers remain past the cap.
• Certainty-label loopholes closed. quaere-evidence's lightweight pass now uses the canonical Decision labels (confirmed | rejected | inconclusive | deferred, was accept | reject) and forbids confirmed while the disconfirming probe is "not yet run"; quaere-audit Triage condition 3 requires naming what was actually consulted, else the finding caps at potential; quaere-grounding's Decision label set is unified across all sites (adds documented / local-only, retires the phantom single-source label); quaere-semantic's module-hypothesis verdict renamed held (was confirmed — that word is reserved for probe-gated claims) and its handoff Confirmed inputs are scoped to corroborator-backed fields; quaere-naming gained an unverified (checks blocked) branch so blocked availability checks are never presented as free.
• quaere-invention candidate IDs renamed K-N (was C-N, which collided with quaere-evidence's Review Claim C-N across an explicit handoff); locally novel is emitted as locally novel (unprobed) until its kill-probe has run downstream.
• Validator hardened — the .agent-state check is git-aware (no more false positives on local gitignored state), broken relative references in skill bodies are detected, README.ja.md coverage is enforced, and quaere-naming / the audit Confirmation Rule joined the required-blocks list.

Fixed

• check-availability.sh default platform expansion — with no platform arguments the script checked nothing and exited 0; it now expands to domain github npm and exits non-zero on unknown platforms.

0.5.0 — 2026-05-30

Quaere splits into a core set plus opt-in extensions, adds a new quaere-invention extension, and reorganizes every skill body so the prescriptive content survives Codex CLI's ~220-line SKILL.md read cap. The install default changed and the skill bodies changed.

Added

• quaere-invention extension — a divergence skill for non-obvious approaches: name the default basin, break the assumptions that make it feel inevitable, label each candidate's novelty honestly, and design a kill-probe before promoting any idea. Opt-in: quaere install --skill invention.

Changed

• install installs the core set only by default — quaere-evidence, quaere-execution, quaere-grounding, quaere-semantic. Extensions are opt-in via quaere install --extensions (all) or quaere install --skill <name> (repeatable, e.g. --skill audit). quaere-audit moved from the default set to an extension. update stays conservative: it refreshes whatever is already installed and does not remove a previously-installed extension or silently pull in new ones.
• list groups core vs. extensions and shows how to install each uninstalled extension.
• Skill bodies reorganized for the Codex read cap. Codex CLI reads only ~the first 220 lines of a SKILL.md; content past that is structurally unread. Handoff sections were front-loaded, worked examples and long detail moved into each skill's references/, quaere-evidence was trimmed 376 → 316 lines, and a compact "Stop now" guardrail reminder was added within the first 200 lines of quaere-evidence, quaere-grounding, quaere-execution, and quaere-audit (the full Stop condition stays at the end for full-read agents like Claude Code and OpenCode). The decision discipline is unchanged — only where it sits in the file.

Internal

• Layout: skills now live under skills/core/ and skills/extensions/; installed skill names are unchanged.
• Eval harness: --runs N (default 5; forced to 1 under --dry-run) with per-cell pass-rates, an outcome-vs-form assertion axis split, and an exact-binomial McNemar paired baseline-vs-skill comparison. Fixed ordered_sections matching for bold markdown headers and added a Japanese requires_pair alternate.
• validate_skills.py: enforces reachability anchors (Iron Law / Handoff triggers / a stop reminder) within a 200-line cap, description ≤ 1024 chars with no angle brackets, a compact stop reminder strictly shorter than its full section, and a table-of-contents on any references/*.md over 100 lines.
• CI gates pull requests that touch cli/ on pnpm check (lint + typecheck + vitest).

0.4.1 — 2026-05-24

Maintenance release. No user-facing CLI or skill changes — the bundled skills are byte-identical to v0.4.0. The release exists to retire the temporary NPM_TOKEN secret used for the first npm publish and to refresh several CI pins that were two majors behind at v0.4.0 ship.

Internal

• Trusted Publishers (OIDC) replaces NPM_TOKEN on the npm-publish step. The Granular Access Token used for the first publish (configured against the broad "All packages" scope because quaere-cli did not yet exist on npm) is revoked, and the NPM_TOKEN repo secret is deleted. Authentication now flows through GitHub Actions OIDC against the trusted-publisher binding on https://www.npmjs.com/package/quaere-cli/access.
• CI toolchain catch-up: pnpm 9 → 11, pnpm/action-setup v4 → v6, actions/setup-node v4 → v6, Node runtime in CI 22 → 24 (active LTS). cli/package.json engines.node stays at >=22 so users on the maintenance LTS line can still install. sigstore/cosign-installer stays at @v3 because the upstream repo does not ship a v4 major-version alias tag yet.
• cli/pnpm-workspace.yaml added. pnpm 11 retired the pnpm field in package.json and reads per-project settings from this file; it carries the allowBuilds policy required to opt in to esbuild's native-binary postinstall (a transitive dep of tsdown / rolldown / vitest).
• docs/cli-contracts.md and README.ja.md aligned with the TypeScript port: the Rust CLI had --target / --skill / --repo flags that the TS port does not. Contract entries pinned to flags that no longer exist were removed.
• Landing page install commands switched to npx quaere-cli install / bunx / npm install -g to match the README.

0.4.0 — 2026-05-24

Distribution switches from the Rust binary chain (Homebrew + cargo + curl|sh) to a single npm package, quaere-cli. The skill bodies are unchanged — the in-tree eval numbers carry over.

Breaking

• brew install, cargo install quaere-cli, and curl ... install.sh | sh no longer ship new releases. v0.3.2 binaries remain downloadable from GitHub Releases for users on the old install paths, but no further versions follow them. The replacement is npx quaere-cli install.
• scripts/install.sh is deleted. The QUAERE_REPO / QUAERE_VERSION / QUAERE_INSTALL_DIR / QUAERE_SKILLS environment overrides go with it.

Added

• npm package quaere-cli (TypeScript, citty-based CLI, ~12 KB bundled). Same surface as before: install [claude|codex|all], list, doctor, update. The package also exposes quaere as a binary alias for users with a global install.
• npm provenance attestation (Sigstore OIDC) on every release. npm audit signatures verifies that the published tarball was built by the release workflow at the exact tag, replacing the cosign signature that previously covered the binary.

Removed

• cli/ Rust source (cli/src/, Cargo.toml, build.rs). The directory now holds the npm package.
• scripts/install.sh (curl|sh installer) and tests/test_validator_parity.py (Python ↔ Rust validator parity test).
• release.yml: cross-compile matrix (build job), bump-homebrew, cargo-publish.
• ci.yml: cli (rust toolchain) and parity jobs.
• pages.yml: install.sh sync into the static site.

Migration

npx quaere-cli install      # replaces every old install method

Users on the old Homebrew tap can brew uninstall haru0416-dev/quaere/quaere to clean up.

0.3.2 — 2026-05-22

Security release. Addresses the four findings from the v0.3.1 audit (F-001 — F-004) and adds recurrence prevention for an agent-write incident discovered during the audit cycle.

Security

• Release SHA256SUMS is now signed with cosign keyless OIDC. install.sh requires cosign and verifies the signature against the release workflow's certificate identity (refs/tags/vX.Y.Z) before computing the archive checksum. A github.com write compromise can no longer impersonate a signed release without also running the workflow. (F-004, CWE-345.) v0.3.2 is the first signed release; older tags must be installed via cargo install quaere-cli --version <X.Y.Z>.
• Terminal-Bench adapter gates credential forwarding on a dataset allowlist. Host ~/.codex/auth.json, OPENAI_API_KEY, and ANTHROPIC_API_KEY are forwarded into per-task containers only when tb run --dataset names terminal-bench-core (currently the sole curated entry) or the operator sets QUAERE_TB_ALLOW_CRED_FWD=1. Anything else fails closed and emits a single stderr line so the operator can see why the run aborted. (F-001, CWE-200.)
• quaere eval no longer walks up from CWD. --repo or $QUAERE_REPO is now required; running the binary from inside an attacker-controlled directory can no longer execute a planted evals/run_skill_evals.py. (F-002, CWE-426/CWE-427.)

Fixed

• Homebrew tap bump regex pairs url lines with their sha256 lines by target instead of substituting positionally; the four release SHAs no longer risk being bound to the wrong artifact URL if the formula is ever reordered. (F-003, CWE-665.)
• cosign install link in install.sh, README.md, and README.ja.md corrected (/cosign/system_config/installation/; the prior path 404'd).
• evals/terminal_bench/install_scripts/install-with-skill.sh now bootstraps cosign before invoking the curl|sh installer so the new signature verification works inside the Debian Bookworm task container (apt has no cosign). Pinned to v3.0.6 with SHA256 verification of the downloaded binary on both amd64 and arm64.

Changed

• Roadmap moved from inline README to docs/roadmap.md; CLI behavior contracts moved to docs/cli-contracts.md. README gains a short ## Docs section linking the three reference documents.

Internal

• evals/run_skill_evals.py sets GIT_CEILING_DIRECTORIES to the scenario workspace's parent before spawning the agent so a git commit inside the workspace cannot walk up and land changes in the parent Quaere repo. (One incident during the v0.3.1 in-tree sweep had to be removed via git filter-repo before this push.)
• eval-manual.yml artifact upload now drops eval-results/**/workspace/** and **/.git/**. eval-terminal-bench.yml drops .cast recordings and sessions/** so a task that printed env or cat $HOME/.codex/auth.json cannot embed credentials in a public artifact.

Measured

Terminal-Bench (terminal-bench-core==0.1.1, 80 tasks, --global-agent-timeout-sec 1800):

mode	resolved	accuracy
Baseline	41 / 80	51.25%
With skill	42 / 80	52.50%
Δ	—	+1.25 pp

Narrow positive, well inside run-to-run variance. The honest reading is "Quaere does not regress Terminal-Bench performance"; do not market this as a measurable improvement. The v0.3.1 Δ of −17.5 pp on the same dataset is superseded — that number was dominated by install-pipeline / agent_timeout failures (76 timeouts in the worst rerun), not by skill quality. With the v0.3.2 install pipeline (cosign-verified, container bootstrap) and the raised agent timeout, agent_timeout count drops from 76 to 7. See docs/evaluation.md for the per-task breakdown.

0.3.1 — 2026-05-21

Patch release: rebuild the Linux release binaries against musl libc so they no longer require a specific GLIBC version. v0.3.0's x86_64-unknown-linux-gnu artifact was built against GLIBC 2.39 (Ubuntu 24.04 builder) and refused to run on Debian Bookworm (2.36), Ubuntu 22.04 (2.35), and older. The musl static binaries are GLIBC-independent and run unchanged on Alpine through RHEL 7 class systems.

Changed

• Linux release artifacts are now x86_64-unknown-linux-musl and aarch64-unknown-linux-musl. The host-facing tarball name pattern changes accordingly (quaere-vX.Y.Z-<target>.tar.gz). The curl one-liner picks the right artifact via uname -m and continues to work transparently.
• release.yml switches the linux x86_64 builder to native ubuntu-latest with musl-tools and the linux aarch64 builder to cross (via taiki-e/install-action) so that ring's C bits cross-compile with a musl-aware toolchain.

Fixed

• evals/terminal_bench/README.md's tb run smoke examples used flags that no longer exist in the current terminal-bench CLI (--task, --output-dir, TB_AGENT_REGISTRY). Replaced with the canonical form (--agent-import-path, --task-id, --output-path, --dataset 'terminal-bench-core==0.1.1') that the v0.3.0-era smoke run actually used.

Measured

In-tree sweep (18 scenarios / 106 assertions, OAuth Codex single-run):

mode	assertion pass rate	scenarios
Baseline	53% (56 / 106)	0 / 18
With skill	91% (96 / 106)	10 / 18
Δ	+37.7 pp	+10

This is the first sweep where the v0.3 line was measured directly; v0.2.1 reported +27 pp on 97 assertions.

Terminal-Bench (terminal-bench-core==0.1.1, 80 tasks, both modes via evals/terminal_bench/):

mode	resolved	accuracy
Baseline	45 / 80	56.25%
With skill	31 / 80	38.75%
Δ	—	−17.5 pp

Of 20 regressions, 17 hit agent_timeout; the agent_timeout count jumped from 9 (baseline) to 33 (with-skill). Quaere's deliberation pass widens wall time on time-boxed tasks past the grader's budget. The in-tree number is the load-bearing measurement; the Terminal-Bench number is reported for honesty, not as a quality verdict. See docs/evaluation.md.

0.3.0 — 2026-05-20

Minor release: introduces an agent-aware install pipeline (quaere install claude / codex / all) so the same CLI deploys to both Claude Code (~/.claude/skills/) and Codex CLI (~/.agents/skills/). The curl one-liner now ends with an explicit quaere install all step and a Commands: block, and quaere install --force swaps skills atomically so a mid-extract failure no longer leaves dest half-written.

Added

Agent-aware install pipeline

• quaere install accepts an agent positional (claude / codex / all). claude targets ~/.claude/skills/, codex targets ~/.agents/skills/, and all writes both, deduplicating by canonical real path so symlinked targets do not produce double-writes. --target is preserved as a hidden escape hatch for advanced relocation.
• The curl installer (scripts/install.sh) runs quaere install all as an explicit post-download step instead of leaving deployment implicit. The QUAERE_SKILLS=0 env var skips this step for binary-only installs.
• Install output ends with a Commands: block listing every deployed skill, so users see the slash commands immediately after curl ... | sh.

Changed

• quaere install --force now swaps a skill atomically. The new content is staged at <target>/.<name>.staging, the previous dest is renamed to <target>/.<name>.backup, staging is renamed into place, and only then is the backup removed. A mid-extract I/O failure leaves dest at the previous complete content; previously a remove_dir_all -> extract sequence could half-write the dest while the manifest still claimed it as installed. Crash residue (.staging, .backup) is silently skipped by quaere doctor's orphan policy (dot-prefixed) and reclaimed on the next install.
• Agent::targets and paths::resolve_target share a single source for the default install paths via paths::claude_default / paths::codex_default, eliminating the previous duplicate hardcodes.

Fixed

• quaere doctor flags frontmatter values that contain : when not quoted, matching scripts/validate_skills.py. tests/test_validator_parity.py pins the cross-validator agreement so future drift fails CI.
• validate_skills.py MAX_SKILL_LINES aligned with the Rust doctor at 500 (was 600).
• evals/run_skill_evals.py builds the runner prompt with printf instead of heredoc, preventing shell-injection when scenario prompts contain EOF-like delimiters.
• scripts/install.sh passes --strip-components=1 to tar so the extracted binary lands at the expected path on all release tarball layouts.
• release.yml workflow_dispatch checks out the requested tag ref instead of HEAD, so manual re-runs build the tag the user asked for, not the current main.

Internal

• .gitignore excludes evals/results/ so per-run eval output does not accidentally land in commits.

0.2.1 — 2026-05-20

Infrastructure release: lands the eval-harness graders and Terminal-Bench adapter that v0.3 will use to publish external-benchmark numbers, plus a Skill content pass addressing post-v0.2.0 review feedback. Treated as a PATCH bump because the Skill changes are additive structural improvements — no Iron Laws changed, no existing gates weakened.

Added

Eval harness graders (ADR-0008)

• behavior assertion type — per-run resource thresholds against runtime metadata (max_tool_calls / min_tool_calls / max_duration_seconds / max_tokens_output / max_tokens_input). Records status: "manual" when the runner has not emitted the relevant metric, so adoption can land before runner-side metadata extraction does.
• llm_judge assertion type — LLM grades output against a free-form rubric. Default-off via --enable-llm-judge (per-PR CI cost stays at zero). Three transport backends share the same grading contract:
  • anthropic (default): Anthropic SDK + ANTHROPIC_API_KEY.
  • openai-compat: openai SDK + --llm-judge-base-url (or $QUAERE_LLM_JUDGE_BASE_URL) for Ollama, vLLM, LM Studio, LocalAI, LiteLLM, etc. Enables fully local eval flows.
  • codex: shells out to codex exec --output-last-message, reusing Codex CLI's existing auth and endpoint. No extra Python SDK required.
• Responses cached at evals/.llm_judge_cache/<sha>.json keyed by sha256(backend:model + rubric + output). The backend prefix in the key prevents cross-backend cache hits.
• Sample assertions added to evals/scenarios.json for both new types (two behavior, one llm_judge) so the schema is exercised end-to-end.

Terminal-Bench adapter (ADR-0007)

• evals/terminal_bench/ package with two AbstractInstalledAgent subclasses:

  • quaere-tb-codex-baseline — Codex CLI only.
  • quaere-tb-codex-with-skill — Codex CLI + quaere install.

  Both share env, model selection, and command shape; the only difference is whether ~/.claude/skills/quaere-* is present inside the per-task container. Lazy-imports the optional terminal-bench dependency so the package can be inspected without it installed.

• evals/terminal_bench/install_scripts/install-baseline.sh and install-with-skill.sh — pin Codex CLI to 0.128.0 by default; the baseline script refuses to run when ~/.claude/skills/quaere-* exists (treats a leaked treatment as a void run); the with-skill script verifies the five expected skills landed before exiting.

• .github/workflows/eval-terminal-bench.yml — manual-trigger (workflow_dispatch) workflow that runs tb run against the registered agents. Inputs choose agent (baseline / with-skill / both), scope (smoke / full / task_id), and codex_version. 4h timeout cap. Never PR-triggered.

Skill content improvements (review feedback)

• quaere-evidence: Added ## Output contract section (冒頭契約) with the 8 required output sections in canonical order, plus a ### Lightweight evidence pass template — same gate run with fewer words for small-scope investigations.
• quaere-audit: Added Tier companion decisions and Tier promotion probe as required output blocks in every Standard/Deep report. These make explicit which companion skills were invoked and whether bulk/cross-tenant exposure was checked before tier promotion.
• quaere-semantic: Added ## Meaningful unit selection section — 6 criteria for which units to analyze in depth; pure local helpers not meeting any criterion can be collapsed into their caller.
• quaere-grounding: Added ## Claim result matrix table and locally observed as a new decision label between "confirmed" and "inconclusive". Clarified that sources derived from the same dependency graph (package.json + lockfile + installed types) are not independent corroborators.
• quaere-execution: Added Verification contract block with labeled Before check / After check forms. Clarified that generated files, lockfiles, snapshots, and build output require justification or must be reverted.
• All 5 skills: Standardized the ## Handoff to other skills section with a common 6-field block template (From skill / Blocking question / Confirmed inputs / Inconclusive inputs / Required next skill / Stop condition).
• evals: Added 4 adversarial scenarios to scenarios.json — urgency pressure (pressure-to-skip-evidence-gate), baseless assertion (baseless-security-assertion), skip-test pressure (skip-tests-pressure), and docs blind-trust (blind-trust-external-docs) — each asserting that the relevant skill's Iron Law gate holds under pressure.
• validator: scripts/validate_skills.py checks for required structural blocks per skill (Output contract, Lightweight evidence pass, Tier companion decisions, Tier promotion probe, Meaningful unit selection, Claim result matrix, Verification contract). Line budget raised from 500 to 600 to accommodate the additions.

Validation + docs

• scripts/validate_skills.py accepts the two new assertion types in its allowlist.
• README.md / README.ja.md gain the new grader rows in the assertion-type table, three subsections covering all llm_judge backends, and a Roadmap upgrade for Terminal-Bench from "v0.3 target" to "v0.3 in progress" with a pointer to the new adapter.
• tests/test_run_skill_evals.py covers the new graders end-to-end (6 behavior tests, 11 llm_judge tests including codex backend dispatch and a backend-specific cache-key regression).
• tests/test_terminal_bench_adapter.py covers the package's lazy-import discipline and missing-dependency error path.

Measured effect

Single-run via Codex CLI against 18 scenarios / 97 deterministic assertions (14 original + 4 new adversarial):

mode	assertion pass rate	scenario-level
Baseline (no skill)	60% (58 / 97)	0 / 18 pass
With skill	87% (84 / 97)	8 pass · 4 inconclusive · 6 fail
Δ	+27 pp

4 inconclusive runs have zero failed assertions; they contain not_in_baseline assertions that require a paired baseline run to resolve. 4 with-skill timeouts account for most failures (complex multi-step skill prompts). The 4 adversarial scenarios score 92% with-skill vs ~46% baseline.

Known limitations (deferred to v0.3.x)

• behavior assertions in the in-tree eval all resolve to status: "manual" because the Codex CLI runner does not yet parse out tool_calls / tokens_* metadata. The grader is ready; the runner-side metric extraction lands separately.
• Terminal-Bench smoke / full sweep / leaderboard runs have not executed; the adapter is wired but the numbers are pending.

0.2.0 — 2026-05-20

Eval-driven quality pass on the v0.1.0 skill set and harness. The 9 with-skill residual failures from the v0.1.0 evaluation are addressed across two work streams without adding new skills.

Measured effect

Two independent baseline + with-skill sweeps via Codex CLI 0.128.0 against the 14 scenarios with 90 deterministic assertions, --scenarios-extra evals/scenarios.ja.json enabled:

mode	range (2 runs)
Baseline (no skill)	63.9 – 65.1%
With skill	91.1 – 94.4%
Δ	+27 to +29 pp

Compared to v0.1.0 (with-skill 89.8% / Δ +28 pp), v0.2.0 lifts the upper bound of with-skill by ~+4.6 pp and leaves Δ in the same band. Run-to-run variance is real for stochastic LLM output; numbers are a range, not a point estimate.

All four Stream A targeted assertion flips confirmed across the runs (regex disconfirm/falsify in ci-failure-evidence-before-patch; requires_pair Backing→source-type in weak-review-claim-rejection; ordered_sections in authorized-implementation-loop; ordered_sections in no-web-local-grounding). Both post-measurement residuals also closed at the assertion level: markdown-tolerant tier regex now matches **Tier:** Standard / **Tier:**\nStandard, and 攻撃者 / 主体 / 攻撃元 alternates landed in the spec-grounded-security-audit ordered_sections position 4.

Changed

Skill templates (Stream A)

• skills/quaere-evidence/SKILL.md — Hypothesis blocks now require six labeled lines (H-id / Based on / Prediction / Falsifier / Disconfirming probe / Alternative); Review Claim blocks grow from 8 to 10 fields with Falsifier: and Disconfirming probe: appended after Suggested probe:. Backing: value must name a source type from {spec | invariant | test | policy | contract | RFC | ADR}. Worked example updated to the new shape.
• skills/quaere-grounding/SKILL.md — No-network fallback strategy is now a fixed ### 10. heading inside ## Workflow, replacing the loose ## No-network fallback H2. Output format carries a labeled No-network fallback strategy: applied | not needed block with the three numbered steps (cached/dated → stale-data notice → canonical URLs).
• skills/quaere-execution/SKILL.md — Final gate: is now a mandatory labeled line at the end of the Study section, including the explicit skip form Final gate: skipped because <reason>; best next command: <cmd>.

Eval harness (Stream B)

• evals/run_skill_evals.py — requires_pair assertions accept an optional skip_when regex; when the antecedent is present but the consequent is missing, a matching skip_when treats the unconditional pair as vacuously satisfied. Used by the targeted-verification and diff-review assertions in authorized-implementation-loop.
• evals/scenarios.json — Tightened spec-grounded-security-audit:
  • audit tier declared now requires the canonical Tier: Triage|Standard|Deep labeled line, not prose mentions.
  • findings classified now requires Confirmed: / Potential: / Rejected: / Inconclusive: as labeled lines.
  • Added confirmed finding references a derived security property or spec clause so confirmed findings must point to a named Property, spec clause, RFC, or invariant.
• evals/run_skill_evals.py — new --scenarios-extra PATH flag merges a sibling JSON file at load time. Regex/string fields get |-alternation appended; ordered_sections.patterns / not_in_baseline.patterns get per-position alternation; contains_any.texts get array concatenation. Unknown scenario ids or assertion names error out so stale extras cannot silently drop alternates.

Added

• eval-fixtures/no-web-local-grounding/ — new workspace fixture: a pinned slimsync 0.7.3 POSIX shell stand-in with --version / --help, a dated vendored manual mirroring an unreachable canonical URL, plus src/sync.sh and .syncignore. Designed to make baseline and with-skill outputs diverge by colliding with rsync --exclude-from=FILE model memory.
• evals/scenarios.ja.json — Japanese-locale alternates for the eight assertions whose tokens had previously been embedded directly into scenarios.json. The main file is now English-only; locale tokens are merged in only when the user passes --scenarios-extra evals/scenarios.ja.json. This keeps the public surface single-language while preserving the language-bias-removal goal from ADR-0006 Stream B-1.
• tests/test_run_skill_evals.py — new unit tests covering the skip_when clause and the --scenarios-extra merge semantics (regex alternation, per-position alternation with empty-slot pass-through, array concatenation, type mismatch and length mismatch errors).

Distribution

• Homebrew tap haru0416-dev/homebrew-quaere is live with a working formula. Install via brew install haru0416-dev/quaere/quaere. The in-tree Formula/quaere.rb stub is removed; the tap repository is the canonical source going forward.
• First publish to crates.io as quaere-cli. Install via cargo install quaere-cli && quaere install. The v0.1.0 CHANGELOG previously claimed a crates.io publish in error; v0.2.0 is the first version actually on the registry.
• cli/Cargo.toml bumped to 0.2.0 so the release binary self-identifies correctly. The v0.1.0 binary embedded version was correctly 0.1.0, but a missed bump at tag time would have caused v0.2.0 binaries to report quaere 0.1.0 — caught and fixed before the release artifacts were finalized.

Web

• web/index.html — added explicit :focus-visible outline, wrapped main content in <main>, and grouped the bottom resources into <nav aria-label="Project resources"> to address WCAG 2.4.7 (Focus Visible) and 1.3.1 (Info and Relationships) findings.
• web/CNAME — removed. The Pages publishing source is a custom Actions workflow, and the current GitHub docs state that any existing CNAME file in workflow-source Pages is ignored and is not required. The custom domain is held in Pages settings only.
• GitHub Pages https_enforced toggled to true via REST API. http://quaere.dev/ now 301-redirects to https://quaere.dev/.

Known limitations (deferred to v0.3)

The v0.2 Codex sweep left these residual failure modes that this release does not address. All belong to the same axis — skills/quaere-*.md Output format does not push the model strongly enough toward canonical section headers and tier-companion sub-decisions:

• ci-failure-evidence-before-patch ledger sections appear in canonical order — Defense section header position drift.
• weak-review-claim-rejection claim, defense, decision in order — content is right, but the model paraphrases rather than emitting labeled sections.
• triage-tier-confirmation-rule Triage tier vocabulary absent in baseline — with-skill output is missing the canonical phrase promote to Standard / tier promotion.
• spec-grounded-security-audit Standard tier requires grounding and evidence-gated review — newly exposed by the v0.2 markdown-tolerant tier regex; when the audit declares Standard tier, the model does not consistently emit the External grounding + Evidence-gated review sub-decision pair.

v0.3 targets template-strengthening across these axes.

Skipped

• Stream B-5 of the internal v0.2 plan (language-tagged assertion sets) is intentionally not implemented. The alternation approach (English-only scenarios.json + opt-in scenarios.ja.json via --scenarios-extra) already covers the dual-language need without scenario-level duplication; revisit only if a third language or significantly larger token sets are added.

0.1.0 — 2026-05-19

Initial release. Quaere is a process-correction skill set for coding agents, relaunched from the archived agent-skills project.

Added

Skills

• skills/quaere-semantic — semantic review of unfamiliar code with What / Why / Invariants / Failure / Connections per unit, mutation-invariance test.
• skills/quaere-grounding — external-fact grounding for SDKs/APIs/CLIs with dual-axis (source + claim-credibility) gate and executable probes.
• skills/quaere-evidence — evidence-gated review for unclear bugs, CI failures, and risky PR claims; Toulmin claim/data/warrant structure.
• skills/quaere-execution — authorized multi-step implementation with Plan/Do/Study/Act units and explicit commit authorization.
• skills/quaere-audit — security-property-driven audit with attack-surface mapping, exploitability gates, and tier-based confirmation rules.

CLI (quaere)

• quaere install — extract the embedded skill set to ~/.claude/skills/ (or any --target). Additive across runs; merges with existing manifest. Validates --skill names early and rejects unknown ones.
• quaere list — show installed skills and the recorded CLI version.
• quaere doctor — verify SKILL.md frontmatter, names, line budget, and orphans. quaere-* orphans are errors; other directories are informational.
• quaere update — check GitHub Releases for a newer version using semver comparison; handles 404/403/429/5xx with specific guidance.
• quaere eval — wrapper for evals/run_skill_evals.py with auto-detected Quaere checkout.
• quaere version — print the CLI version.

Distribution

• scripts/install.sh — POSIX curl | sh installer with SHA256 verification, four-target detection, and PATH hint.
• .github/workflows/release.yml — tag-triggered cross-platform build (x86_64 + aarch64 for Linux and macOS), skill tarball, SHA256SUMS, and gh release create.
• Formula/quaere.rb — Homebrew formula stub (tap repo planned for v0.2).

Evaluation infrastructure

• evals/scenarios.json — 14 portable scenarios across the five skills with 88 deterministic assertions (contains, not_contains, regex, ordered_sections, requires_pair, min_section_count, not_in_baseline).
• evals/run_skill_evals.py — runner that drives any local agent CLI through baseline/with-skill modes and writes isolated prompts, outputs, and grades.
• eval-fixtures/ — six workspace fixtures pinned for the relevant scenarios.

Tests and CI

• scripts/validate_skills.py plus 34 Python unit tests covering frontmatter, naming, line budget, and scenarios.json structure.
• tests/golden/quaere-*.json plus a 6-test runner enforcing per-skill required H2 sections, forbidden patterns, and line budgets.
• tests/test_validator_parity.py plus a CI parity job that runs both validators against the same eight SKILL.md mutations and asserts they agree.
• Matrix CI: Python 3.11/3.12/3.13 + Rust stable/beta + parity job.
• Manual eval-manual.yml workflow for real-LLM evaluation, gated behind workflow_dispatch.

Migration from agent-skills

Old skill names are renamed wholesale; there are no compatibility shims.

Old	New
semantic-review	quaere-semantic
external-grounding	quaere-grounding
evidence-gated-review	quaere-evidence
execution-loop	quaere-execution
security-audit-loop	quaere-audit

To migrate from a prior agent-skills install:

rm -rf ~/.claude/skills/{semantic-review,external-grounding,evidence-gated-review,execution-loop,security-audit-loop}
curl -fsSL https://raw.githubusercontent.com/haru0416-dev/quaere/main/scripts/install.sh | sh