v0.35.2

v0.35.2 - 2026-07-15

/quality skill

• Automatic evaluation selection now discovers ready SDK evaluators before
  considering current-session harness judgment. A usable SDK matching the
  invoking Codex or Claude agent is preferred; when no provider match is usable,
  the CLI's deterministic automatic winner is used, and harness judgment is the
  no-SDK fallback.
• Provider-named requests now map directly to that provider's independent SDK
  evaluator without a harness-versus-SDK question. Before writing evaluation
  artifacts, the workflow reports the selected method and whether explicit
  intent, configuration, provider affinity, CLI discovery, or harness fallback
  determined it.
• Skill metadata advances to 0.35.2 and retains
  qualitymd >=0.35.0 <0.36.0 compatibility because this patch changes the
  agent-mediated selection policy without changing the skill-facing CLI
  surface.

Compatibility / migration

• No CLI command, automatic-discovery ordering, JSON receipt, persisted
  evaluation artifact, or QUALITY.md format behavior changes. Existing
  0.35.x CLI installations remain compatible with the updated skill.

Compatibility:

• CLI: v0.35.2
• QUALITY.md specification: 0.12 (Draft)
• /quality skill: 0.35.2, requires qualitymd >=0.35.0 <0.36.0

v0.35.1

v0.35.1 - 2026-07-15

/quality skill

• Evaluation progress now reports preflight, evidence review, report
  generation, meaningful model coverage, and whether attention is needed.
  Request windows, work units, payload schemas, concurrency caps, worker or
  subagent orchestration, and resume loops stay behind the agent interface
  unless a decision or recovery step requires them.
• Default in-session evaluator selection is now concise information. An
  independent evaluator remains discoverable for future invocations; any
  offered current-run change is a real wait-for-answer choice. Evaluator
  selection now precedes the first feedback-log write, so the pre-mutation
  progress boundary matches the actual workflow.
• Skill metadata advances to 0.35.1 and retains
  qualitymd >=0.35.0 <0.36.0 compatibility because this patch changes the
  agent-mediated interface without changing the skill-facing CLI surface.

Documentation

• The agent-mediated UX guide now defines the implementation boundary and shows
  how to translate protocol state into task phases and meaningful coverage.

Compatibility / migration

• No CLI command, receipt, persisted evaluation artifact, or QUALITY.md format
  behavior changes. Existing 0.35.x CLI installations remain compatible with
  the updated skill.

Compatibility:

• CLI: v0.35.1
• QUALITY.md specification: 0.12 (Draft)
• /quality skill: 0.35.1, requires qualitymd >=0.35.0 <0.36.0

v0.35.0

v0.35.0 - 2026-07-15

CLI

• Automatic evaluator discovery now probes and reports both Codex and Claude
  candidates even when the first is ready. Dry-run and run receipts carry each
  candidate's structured verified, assumed, or unchecked authentication
  basis; when multiple runtimes are usable, the selection reason names the
  deterministic ordering decision and every usable runtime not selected.
• Claude readiness now uses the documented non-interactive authentication
  status probe. A verified logged-out runtime is skipped, while an unavailable
  probe remains an explicit authentication assumption; credential values are
  never reported.

/quality skill

• Provider-named evaluator requests that could mean either the current harness
  or a same-provider SDK subprocess now prompt for that transport choice. When
  default precedence selects in-session harness judgment, the explanation
  names the fresh independent SDK alternative and its one-run and durable
  configuration paths.
• Skill metadata advances to 0.35.0 and requires the matching
  qualitymd >=0.35.0 <0.36.0 CLI line for intent-faithful evaluator selection
  and receipt interpretation.

Compatibility / migration

• Automatic evaluator receipts add authenticationBasis to every candidate
  and now report every built-in candidate, including usable runtimes not
  selected by deterministic ordering. Consumers of JSON receipts should accept
  the expanded candidate array and additive field; no persisted evaluation
  artifact migration is required.

Compatibility:

• CLI: v0.35.0
• QUALITY.md specification: 0.12 (Draft)
• /quality skill: 0.35.0, requires qualitymd >=0.35.0 <0.36.0

v0.34.0

v0.34.0 - 2026-07-15

CLI

• Atomic evaluation, report, lint-repair, and cancellation writes no longer
  leave orphaned temporary directories behind. Provider result files are now
  scoped to each completed evaluator call and cleaned up immediately after the
  checkpoint consumes them.
• Evaluation advice now synthesizes one stakeholder-facing executive summary
  after finding and recommendation ranking. The run report ## Summary renders
  its bottom-line headline, concise narrative, and concrete key points instead
  of exposing the scoped root's aggregation trace.
• evaluation.json advances to schema version 10 and
  EvaluationOutputResult references the persisted EvaluationSummaryResult.
  Report build and status require that advice payload before treating a run as
  reportable.

/quality skill

• Skill metadata advances to 0.34.0 and requires the matching
  qualitymd >=0.34.0 <0.35.0 CLI line for the schema-10 evaluation artifact
  and executive-summary work unit.

Compatibility / migration

• In-flight schema-version-9 evaluation runs cannot resume under this release
  and must be started again. Completed version-9 artifacts remain historical
  data; there is no migration or dual reader.

Compatibility:

• CLI: v0.34.0
• QUALITY.md specification: 0.12 (Draft)
• /quality skill: 0.34.0, requires qualitymd >=0.34.0 <0.35.0

v0.33.0

v0.33.0 - 2026-07-15

CLI

• Evaluation concurrency now comes from the selected evaluator transport rather
  than host CPU count: harness and Codex default to four, while Claude remains
  sequential. Workspace evaluation.concurrency stays the only configuration
  override and is clamped by any evaluator maximum; no CLI flag is added.
• Direct evaluator runs now create their immutable manifest with the selected
  evaluator before issuing work and use a completion-driven bounded pool. Each
  accepted result is persisted before its slot is refilled, so fast calls no
  longer wait behind the slowest sibling and accepted work survives
  interruption. Harness runs retain their bounded rolling request window.
• evaluation.json advances to schema version 9 with structured dispatch
  capabilities and concurrency provenance. In-flight version-8 runs must be
  started again; completed historical artifacts remain historical records.

/quality skill

• Harness evaluation may assign one self-contained outstanding request to each
  native worker while the parent keeps graph, artifact, quality-control, and
  retry authority. Progress distinguishes the runner's outstanding cap from
  requests actually dispatched.

Compatibility / migration

• In-flight schema-version-8 evaluation runs cannot resume under v0.33.0 and
  must be started again. Completed version-8 artifacts remain historical data;
  there is no migration or dual reader.
• Automatic evaluation concurrency no longer follows host CPU count. Harness
  and Codex default to four; Claude is sequential even when a higher workspace
  cap is configured. Set positive evaluation.concurrency in workspace
  configuration when a lower explicit cap is required.

Compatibility:

• CLI: v0.33.0
• QUALITY.md specification: 0.12 (Draft)
• /quality skill: 0.33.0, requires qualitymd >=0.33.0 <0.34.0

v0.32.2

v0.32.2 - 2026-07-15

Internal

• Evaluator prompts now place canonical shared policy, task, model, and
  inspection context before each work-unit delta so provider prompt caching can
  reuse the longest safe prefix across independent sessions. Claude uses its
  supported cache-stable system-prompt preset, and run-local evaluator-call
  logs preserve separately reported cache-read and cache-creation input tokens.
  Evaluator sessions remain fresh and are never resumed or forked for reuse.

/quality skill

• Skill metadata advances to 0.32.2; its compatible CLI line remains
  qualitymd >=0.32.0 <0.33.0 because this patch preserves the skill-facing
  command and evaluation artifact contract.

Compatibility:

• CLI: v0.32.2
• QUALITY.md specification: 0.12 (Draft)
• /quality skill: 0.32.2, requires qualitymd >=0.32.0 <0.33.0

v0.32.1

v0.32.1 - 2026-07-14

CLI

• Evaluation history and next-run numbering now share one recognition rule:
  current evaluation.json metadata takes precedence over the historical data
  manifest and folder prefix, unreadable or invalid metadata falls through, and
  manifest-less run folders whose slug contains quality are no longer
  silently excluded.

Internal

• Evaluation execution now derives frames, work scheduling, protocol requests,
  artifacts, and summaries through pure domain functions and Effect-native
  ordered traversal. Source collections across the runtime are expression-built
  and exported collection types are readonly.

/quality skill

• Skill metadata advances to 0.32.1; its compatible CLI line remains
  qualitymd >=0.32.0 <0.33.0 because this patch preserves the skill-facing
  command and artifact contract.

Compatibility:

• CLI: v0.32.1
• QUALITY.md specification: 0.12 (Draft)
• /quality skill: 0.32.1, requires qualitymd >=0.32.0 <0.33.0

v0.32.0

v0.32.0 - 2026-07-14

Specification

• QUALITY.md specification 0.12 (Draft) clarifies that an area's source
  identifies the subject or starting boundary being judged, not a precomputed
  prompt payload or permission boundary for every supporting file. Evaluation
  methods may inspect separately classified supporting context without silently
  widening the area or requirement.

CLI

• Requirement judgment now runs in a fresh coding-agent inspection session.
  Codex, Claude, or the invoking harness searches and reads requirement-specific
  context inside a neutral, read-only workspace policy with network, approval
  escalation, writes, repository-instruction authority, and executable
  verification disabled.
• The runner no longer walks, truncates, or packages one area-wide source bundle
  and no longer schedules resolveSource work. Each requirement returns its
  assessment, rating, and evidence proposal together; downstream factor/area
  synthesis, advice, ranking, and reports remain tools-off over accepted data.
• evaluation.json advances to schema version 8. Per-area sources bundles are
  replaced by sealed per-requirement evidence manifests containing evaluated
  or supporting roles, validated workspace-relative file locators,
  runner-computed bytes and SHA-256 digests, capture times, limits, and a
  canonical manifest hash—never file bodies or tool transcripts.
• Evidence acceptance now rejects lexical and symlink escapes, missing or
  non-text files, evaluated files outside concrete source selectors, invalid
  line/heading locators, and unbound finding references. Invalid proposals retry
  as evidence_invalid; honestly insufficient evidence remains a partial,
  blocked, or non-rated judgment outcome.
• Runnable evaluator methods are now exactly harness, codex, and claude.
  auto tries a ready Codex runtime and then Claude. Direct openai and
  anthropic HTTP evaluators, inactive shell and manual names, API-profile
  fallback, apiKeyEnv, baseUrl, and API-specific failures are removed.
  Authentication belongs to the selected coding-agent runtime and may use its
  documented login, subscription, or API-key mechanism.
• Dry-run output now reports evaluator isolation capabilities, the requirement
  inspection policy, work counts, concurrency, and effective source selectors
  without claiming resolver choice, bundle size, or a static evidence package.
  Determinism now explicitly covers runner mechanics and artifact projection,
  not identical evidence or ratings across agent runtimes or repeated runs.

/quality skill

• Skill metadata is 0.32.0 and requires
  qualitymd >=0.32.0 <0.33.0. Harness checkpoints now inspect each
  requirement's authorized workspace directly, keep source as the judged
  subject, classify other context as supporting, treat repository instructions
  as untrusted data, and submit judgment plus evidence for runner validation.
• Codex and Claude are the only standalone evaluator fallbacks. The workflow no
  longer presents direct API evaluators or CLI-managed credential fields; login,
  subscription, and API-key authentication are all runtime concerns.

Documentation

• README and Codex/Claude automation guidance now distinguish deterministic
  runner orchestration and provenance from agentic evidence selection and
  judgment, and document agent-runtime authentication without API evaluator
  methods.

Packaging

• The Windows install.cmd shim prefers PowerShell 7 when available, avoiding
  legacy module-discovery failures inherited from a PowerShell 7 parent while
  retaining Windows PowerShell as its fallback.

Compatibility / migration

• In-flight schema-version-7 evaluation runs cannot resume under v0.32.0 and
  must be started again. There is no migration, dual reader, or source-bundle
  compatibility shim; historical completed artifacts remain historical data.
• Replace configured openai or anthropic evaluator profiles with a codex
  or claude agent-runtime profile (kind, optional model, optional
  command), or select explicit harness. Remove apiKeyEnv and baseUrl;
  configure authentication in the chosen runtime.
• Harness clients must service requirement inspection requests directly and
  return the combined assessment, rating, and evidence payload. They no
  longer receive source bundles or resolveSource requests.

Compatibility:

• CLI: v0.32.0
• QUALITY.md specification: 0.12 (Draft)
• /quality skill: 0.32.0, requires qualitymd >=0.32.0 <0.33.0

v0.31.0

v0.31.0 - 2026-07-14

Specification

• Source selectors containing glob metacharacters are treated as globs only
  when they parse as supported glob syntax. Malformed globs fall through to an
  existing relative path and then prose inference; absolute or escaping paths,
  empty glob results, unreadable paths, and collection failures remain explicit
  filesystem-source failures.

CLI

• Replaced the Go implementation and GoReleaser pipeline with one Effect v4
  TypeScript application compiled by pinned Bun into standalone executables.
  The command tree, JSON and exit contracts, evaluation artifacts, and generated
  reports remain compatible; the repository no longer ships Go source,
  go install, a Go fallback, or a Node sidecar.
• Codex and Claude evaluations now use their supported TypeScript agent SDKs;
  direct OpenAI and Anthropic API profiles and harness checkpoints remain
  available. Evaluator capabilities, readiness evidence, bounded source access,
  cancellation, retry, and call metadata are explicit rather than normalized
  across providers.
• Each area gets one immutable, hashed context assembled from its source bundle,
  frame, rating criteria, and body guidance. Every requirement judgment runs in
  a fresh provider session or thread so sibling transcripts cannot leak into one
  another, and resume reconstructs context without provider-retained sessions.
• Release archives and npm platform packages now cover Darwin, Windows, Linux
  glibc, and Linux musl on arm64 and x64. Checksums are mandatory, installers
  detect libc, and release repair preserves already-published Bun artifacts
  because compiled output is not byte-reproducible.
• Harness-backed evaluation runs are no longer pinned to concurrency 1: the
  runner now keeps a rolling window of dependency-ready work requests
  outstanding — up to the run's resolved evaluation.concurrency — topping
  the window up as results are accepted, so the invoking harness can judge
  independent requests in parallel or fan them out to subagents.
  Omitted-concurrency harness runs now use the shared automatic default
  (max(2, NumCPU*2)) instead of forced 1; receipts, dry-run previews, and
  progress output state the resolved concurrency and window width.
• awaiting_evaluator receipts carry evaluatorRequests — the outstanding
  bounded work-request set — replacing the singular evaluatorRequest, and
  --evaluator-result accepts one result envelope or a JSON array covering
  any subset of the outstanding requests. Not-yet-submitted requests stay
  outstanding at no retry cost; a schema-invalid or failed member re-emits
  for its retry attempt (with its lastFailure named) without touching other
  members' accepted results.
• evaluation.json bumps to schema version 7: the harness checkpoint state
  is now the plural pendingEvaluatorCalls. Clean break: an in-flight
  awaiting run created before the upgrade cannot resume across it — re-run
  it; completed runs are unaffected. qualitymd evaluation status lists
  every outstanding request for an awaiting run.

/quality skill

• Skill metadata is now 0.31.0 and requires qualitymd >=0.31.0 <0.32.0.
  Evaluation guidance distinguishes SDK agent runtimes from direct API
  evaluators, reports capability remedies, and treats area context and
  requirement session isolation as runner-owned boundaries.
• The evaluate workflow's checkpoint loop services the outstanding request
  set: it may delegate independent requests to subagents, submits results as
  they become ready (one envelope or several per call), and names the window
  width on the first windowed receipt.

Compatibility / migration

• The Go implementation, go install path, and GoReleaser assets are removed.
  Install the standalone archive, npm launcher, or Homebrew cask instead.
• In-flight runner artifacts awaiting evaluator input use a clean schema break
  from version 6 to version 7 and should be started fresh; completed evaluation
  artifacts remain readable.
• Harness clients should read evaluatorRequests and
  pendingEvaluatorCalls, and may submit one result envelope or an array. The
  former singular fields are no longer emitted.
• The QUALITY.md specification stays at 0.11 (Draft); malformed-glob handling
  clarifies source-selection behavior without changing the document schema.

Compatibility:

• CLI: v0.31.0
• QUALITY.md specification: 0.11 (Draft)
• /quality skill: 0.31.0, requires qualitymd >=0.31.0 <0.32.0

v0.30.0

v0.30.0 - 2026-07-11

Specification

• The format now commits to non-filesystem source selectors: source stays a
  single string, and its kind is detected from the selector itself — glob
  metacharacters make a glob, an existing filesystem entry is a path, and any
  other selector describes the evaluated material in prose (for example
  "open tickets in the support queue"), resolved by evaluating tools. The
  frontmatter shape and quality.schema.json are unchanged. The specification
  version moves to 0.11 (Draft): what a source selector means in a
  conforming document has changed.

CLI

• Evaluation runs resolve every source selector through a per-kind resolver.
  Path and glob selectors keep the deterministic workspace walk; a prose
  selector is dispatched to the invoking harness as a resolveSource work
  request on the existing checkpoint transport, and the returned material is
  validated, capped, hashed, and captured into the run artifact before any
  dependent judgment. Selector kinds are detected and pinned at run creation
  and honored on resume.
• A selector kind the selected evaluator cannot resolve now fails the run at
  plan time with the new selector_unsupported category — naming the
  selector, its detected kind, and the remedy — distinct from
  source_unavailable, which keeps meaning the named material is missing.
• evaluation.json bumps to schema version 6 and gains a per-area sources
  provenance record: each area's selector, detected kind, serving resolver
  (and harness runtime when harness-resolved), bundle hash, capture time, and
  per-file hashes — with captured content kept for harness-gathered evidence
  so resume re-judges against the exact evidence of record.
• qualitymd evaluation run --dry-run --json and run receipts now include
  the per-area source dispatch plan (selector, detected kind, resolver).
• Evaluation runs now resolve area sources as the format specification
  defines: a source-less root area evaluates the QUALITY.md file's directory,
  a source-less child area inherits the nearest ancestor's source, and glob
  selectors (for example docs/**/*.md) are expanded. Previously these were
  silently packaged as empty evidence.
• An area source that resolves to no readable files now fails the run loudly
  with source_unavailable, naming the unresolved selector, instead of
  judging the area against an empty source bundle.
• Source packaging skips symlinked directories and files, sockets, and
  devices instead of crashing on them (is a directory), so repositories with
  committed symlinks evaluate and resume cleanly.
• qualitymd lint no longer reports spec-permitted extension frontmatter as
  invalid-frontmatter: a key that names no model property is now the
  warning-severity unknown-key advisory, so conforming documents carrying
  extension properties lint valid and load. Misplaced root keys on nested
  areas remain errors.
• quality.schema.json now accepts any non-empty scalar (string, number, or
  boolean) for content scalars such as assessment, rating-level
  criterion, and ratings override values, matching lint; its $comment
  no longer claims a rating-level ordering check that no tool performs.

/quality skill

• The harness checkpoint loop now serves resolveSource resolution requests
  alongside judgment requests: the skill gathers exactly the material the
  request's selector describes and returns it verbatim, or reports
  source_unavailable when it does not exist.
• /quality skill metadata now declares version 0.30.0 and requires the
  qualitymd CLI 0.30.x line.

Compatibility / migration

• The QUALITY.md specification moves to 0.11 (Draft). A source selector is
  now defined for non-filesystem material: a selector that is neither a glob
  nor an existing filesystem path is prose describing the evaluated material,
  where before only filesystem selectors were defined. Existing path and glob
  selectors are unchanged.
• Runner evaluation.json artifacts now use schema version 6 and add the
  per-area sources provenance record; in-flight version 5 runner runs should
  be started fresh.
• A selector kind the selected evaluator cannot resolve now fails the run at
  plan time with the new selector_unsupported category. source_unavailable
  keeps its narrower meaning: the named material is missing.

Compatibility:

• CLI: v0.30.0
• QUALITY.md specification: 0.11 (Draft)
• /quality skill: 0.30.0, requires qualitymd >=0.30.0 <0.31.0