v1.62.0

Minor Changes

• 0cd07f4: Add paqad-ai join for zero-diff teammate setup, persist onboarding RAG choices in tracked team configuration, and make repeat onboarding manifests byte-stable.

v1.61.1

Patch Changes

• 496950a: fix(#394): feature-development plan/spec must land in the feature bundle

  A feature-development change could produce correct code yet write its plan and spec
  to an invented .paqad/features/<slug>/ directory, never creating the bundle's rigid
  plan.json / specification.json, while the stages still read green. Three fixes
  close that:

  • Rules (RC-1): the feature-development rule pack and requirement-analyst agent now
    instruct paqad-ai plan compile / paqad-ai spec freeze into the bundle and no longer
    name .paqad/plans/* or .paqad/specs/* as durable write targets.
  • Enforcement (RC-2): a planning / specification stage-end whose artifact is not
    the active bundle's plan.json / specification.json is recorded inconclusive (the
    artifact is dropped so no digest is hashed), with a message naming the compile/freeze
    verb. review and the mutation stages are unchanged.
  • Verdict (RC-3): the end-of-change completeness verdict now asserts the bundle
    actually contains a non-empty plan.json and specification.json; missing either
    reads incomplete / Needs your attention, never Safe to merge.

  The dogfooded fix produced both artifacts in its own feature bundle, and no
  .paqad/features/ directory was created.

v1.61.0

Minor Changes

• 9467f47: feat(#389): imperative advisory-host stage protocol + honest JetBrains tiering

  The JetBrains Claude Agent incident (framework followed, but stages unenforced and
  the ledger empty of planning/specification artifacts) exposed that on advisory hosts
  paqad had zero enforcement and relied on the model voluntarily self-recording. The
  bootstrap only said "narrate every stage and the verdict yourself" — too weak. The
  advisory-host section now imperatively requires running the host-independent CLI for
  a feature-development change: paqad-ai stage start/end, paqad-ai plan compile, and
  paqad-ai spec freeze, in order, so the stage-evidence ledger carries real planning +
  specification artifacts even where no lifecycle hook fires.

  The verification-enforcement doc now documents the three JetBrains products honestly:
  Claude Agent / AI Assistant reads CLAUDE.md-equivalent guidance but fires no
  Claude Code hooks (advisory — the incident's host), the Claude Code [Beta] plugin
  runs the real claude CLI so paqad's hooks fire (full, hook-enforced — the supported
  enforcement path), and Junie is advisory. The aiassistant adapter (already
  correctly hooks:false) is now listed in the per-adapter coverage matrix, and a
  reproducible manual advisory-host protocol is documented. No host-detection change and
  no overclaim of enforcement paqad does not have on the host.

• 9467f47: fix(#390): gate feature-evidence bundle creation + report render to the feature-development route

  Feature-evidence artifacts — the feature-evidence/change-<ULID>/ bundle,
  stage-evidence.jsonl, report.html, receipt.json, ai-bom.json, rag.jsonl,
  and _session/<id>.json — were created for every workflow, not only
  feature-development. A root-cause-analysis (or any non-feature) session that emitted
  paqad:stage markers and reached Stop still minted a change-<ULID> bundle and a
  report.html, because both the marker-driven auto-open and the report/receipt render
  gated only on "is a feature pointer active?" (currentFeature() != null), never on the
  routed workflow.

  Both seams now consult the persisted route via isFeatureDevelopmentRoute (a new
  routeIsAffirmativelyNonFeature helper), not the working-tree diff:

  • parseAndRecordMarkers records nothing when the route is affirmatively
    non-feature-development, so no bundle or _session control is minted.
  • The per-feature receipt/AI-BOM projection and renderActiveFeatureReport in
    run-repository-verification.ts write nothing when the route is affirmatively
    non-feature-development, even if a pointer leaks through.

  Cross-provider safe by design: suppression fires only on a route we can prove
  is non-feature. The route seam runs only in Claude's UserPromptSubmit hook, so
  Codex/Gemini sessions — which record markers through the same recorder but never write
  route state — have no route state; an absent route is treated as "unknown" and
  preserves today's recording behaviour rather than silently killing all cross-provider
  feature-evidence. Feature-development behaviour is unchanged.

• 9467f47: feat(#388): stop auto-opening report.html in the browser and remove the feature_report_auto_open config option

  The per-feature report.html no longer opens in the OS browser. The auto-open-on-completion
  path (fired by the completion hook) and the manual paqad-ai feature report --open path are
  both removed, and the sandbox-aware opener (report-open.ts) is deleted — no code path opens a
  browser anymore.

  Removed config key: feature_report_auto_open is gone from the config registry, resolver,
  serializer, listing, the project-profile type and JSON schema, the onboarding default seed, and
  the generated config templates. On the next onboard/update the config-split reconcile (#227)
  prunes the key from team/local .config.* files, so existing projects stop opening the browser
  without manual edits — every still-valid value is preserved.

  Report generation is unchanged: the feature_report flag (default on) and
  renderActiveFeatureReport still write report.html into the feature bundle exactly as before.

Patch Changes

• 9467f47: fix(#387): reject sequential decision ids at write, migrate legacy fixtures

  Decision packets could still be minted with sequential D-{number} ids when an
  agent bypassed the sanctioned writer and hand-authored the JSON, seeded by legacy
  D-1/D-2/D-3 example packets. The creation paths now funnel through a single
  canonical D-<ULID> guard — DecisionStore.writePending rejects a caller-supplied
  non-ULID id, and createPendingDecision already mints one — while read, validate,
  and list stay tolerant of legacy ids so pre-existing packets keep working. Onboard
  and update now best-effort migrate any .paqad/decisions/{pending,resolved}/D-{N}.json
  to a minted ULID id (updating the in-file id and index reference, idempotent on
  already-ULID packets), the repo's own legacy fixtures are migrated, and the bootstrap
  now states imperatively that packets are created only via the decision skill,
  never by hand — including on advisory hosts with no Decision-Pause hook.

• 9467f47: Close two CodeQL high-severity findings surfaced on the evidence/decision paths:

  • Resource exhaustion (js/resource-exhaustion-from-deep-object-traversal) — SchemaValidator no longer sets ajv allErrors: true, so validation short-circuits at the first error instead of traversing an entire (potentially deeply nested, caller-influenced) object to collect every error. verbose is retained so error output still carries the offending value; results now report the first failure rather than all of them.
  • File-system race (js/file-system-race) — the legacy decision-id migration's remapIndex dropped its existsSync check-then-writeFileSync sequence, which was a time-of-check/time-of-use race. It now relies on the existing try/catch (a missing index throws ENOENT, treated as "nothing to remap"), with identical behaviour and no race window.

• 9467f47: fix(#386): batch onboarding's git check-ignore to a single call

  Repository discovery spawned two synchronous git check-ignore subprocesses per
  visited directory, so onboarding stalled for seconds before the provider prompt on
  large repositories. It now walks the tree once into memory and resolves all
  git-ignored paths with a single batched git check-ignore call, cutting spawns
  from O(directories) to O(1) with identical detection output.

v1.60.2

Patch Changes

• 3900daf: Prevent onboarding repository discovery from scanning dot directories, ignored paths, generated trees, and nested Git checkouts.

v1.60.1

Patch Changes

• 54d949b: fix(#380): stage-integrity fixes from the 1.59.0 #356 dogfood RCA

  Two deterministic stage-evidence defects surfaced by the #380 RCA are fixed:

  • Orphan feature bundle ("bug #5"). The SessionStart hook now aligns the
    machine-local ledger-session cache to the live host session id. Previously the
    single-slot cache could still hold a prior session's id at the start of a new
    session, so a shell paqad-ai stage start --title invoked before the first edit
    minted its feature bundle under a stale session while the PreToolUse gate keyed on
    the true host session — orphaning the bundle and blocking the first edit. This
    extends the bug #5 mitigation from finalization to bundle minting.
  • documentation_sync left start-only. recordMarkedStage now forward-closes any
    earlier still-open stage before opening a marked (non-mutation) stage, symmetric with
    the live edit writer. Previously the last mutation stage (e.g. documentation_sync)
    had no later mutation edit to close it, so a following marked review stage opened
    without closing it and the change was left with a dangling open stage.

v1.60.0

Minor Changes

• 2b2ee2c: Planning now sees the codebase: on the feature-development route, the background context
  worker composes a token-capped ## Existing surface section into the session-context
  artifact. It lists the exported symbols that already exist for the files/modules the prompt
  and working set implicate — as signature cards (name(signature) — file:line · called from N places · module) ranked by the repo-map's structural importance (PageRank) — so the model
  reuses what exists instead of rewriting it. Signatures and caller counts come from the
  code-knowledge index when present, falling back to name-only cards otherwise. The section is
  budget-capped (config existing_surface_tokens, default 1000) with an honest truncation
  line, appears only for feature-development (every other route stays token-neutral), and rides
  the existing artifact so Codex/Gemini/advisory hosts get it through the same file. This gives
  the built-but-unconsumed repo-map its first live consumer.

v1.59.0

Minor Changes

• deb1cb9: Make the end-of-change verdict visible, honest, and enforced (#368). The completion backstop now surfaces the one end-of-change receipt to the developer whether the verdict passes, fails, or is inconclusive — a failing "Needs your attention" verdict can no longer be hidden on stderr while a PR ships. A hard-failing feature-development change now blocks the turn at the Stop seam via the documented decision:block channel (exit 2 is a no-op there), bounded by the stop-hook loop guard. A feature-development change with no real paqad-ai checks run report reads as Inconclusive and shows its checks stage as unverified rather than done. The checks.block_on_failure escalation stays script-enforced; review_findings / stale_docs are documented as agent-raised and Decision-Pause-enforced rather than falsely claiming script enforcement. The cross-provider verdict tiering (hook-surfaced on Claude, agent-narrated on Codex/Gemini/advisory) is documented so the promise is truthful per host.

  Also (AC-D2, from the second reproduction on the #355 session): the evidence receipt's verification_result is now three-way — FAILED only on a genuine failure, INCONCLUSIVE when a measure could not run (a blocked ratchet from unwired/absent tooling) or could not be judged, and PASSED when every row passed. A couldn't-verify is no longer reported as a failure, so an unwired quality-ratchet measure can no longer manufacture a false FAILED on the attestation.

v1.58.0

Minor Changes

• ee46d89: New routed workflow #10 codebase-health (backed by health-retest): a project health check-up on demand, at the same tier as pentest and design-test. Say "check my project's health" and paqad scans for six kinds of junk — dead code, unused packages, outdated or risky packages, leaked secrets, stale docs, and copy-paste AI slop — and lands every finding in one report with proof (real tool output, never an AI opinion), a plain-words reason it matters, and what to do about it.

  Detection is a single deterministic verb, paqad-ai health run, that costs zero model tokens: it reuses the code-knowledge index (dead code, unused deps), the 9-ecosystem dependency inventory, OSV/native-audit vulnerability data, registry deprecation metadata, and the chunk-similarity index, shelling out to osv-scanner, gitleaks, jscpd, and knip when they are on PATH and degrading gracefully (labelled fallbacks or blocked_checks) when they are not. Findings carry stable content-addressed HL- ids, split honestly into a "Proven" section (deterministic) and a "Needs judgment" section (ai-judged, confidence-scored), and never expose secret bytes — only a file:line + rule + fingerprint.

  Finding fingerprints use SHA-256 (a stable content-addressed dedup key, not a security primitive; secret bytes never reach it). All scanners are invoked from Node, so the workflow is cross-platform (Windows included). A baseline ratchet marks findings new-since-baseline vs pre-existing so a legacy project isn't drowned, --offline skips the network categories, and paqad-ai health retest re-runs the evidence and reclassifies each finding fixed | still-open | needs-manual-verification by its stable id. The workflow is classifier-routed (health phrasings route to codebase-health; pentest phrasings still route to pentest), surfaces in the dashboard, and its run summaries flow to the SIEM export.

v1.57.0

Minor Changes

• f76f2dd: Per-feature evidence report (issue #371): a feature's whole on-disk record — plan, frozen spec, every stage with real durations, the rules that ran, what retrieval found, the verification receipt, the AI-BOM, the commit trail, and the review — is now rendered into one self-contained, human-readable report.html by a pure script (no LLM, no network), saved next to the JSON it came from and byte-for-byte identical on every provider. A normal person can read it, it prints cleanly to PDF for an auditor, it stays local and git-ignored, and it makes zero external requests.

  Every section communicates gracefully when its data is absent: a missing plan, an unrun retrieval, or (on a non-enterprise install) an absent verification receipt and AI-BOM each render an explicit plain-English note instead of a blank gap. The verification receipt honestly says whether its hash-chain integrity recomputes, and a failed gate is shown as plainly and prominently as a passing one.

  New: paqad-ai feature report [ref] [--out <file>] [--open] [--quiet]. The report is regenerated automatically at end-of-change (covering Claude Code, Codex, and Gemini through the one verification backstop) and after every commit/merge (covering advisory hosts through the git hooks), always best-effort so it can never disrupt the feature-development stages or the verification verdict. Two config flags: feature_report (default on) and feature_report_auto_open (default off; sandbox-aware, skips CI/SSH/remote/headless).

  The page is laid out like a dashboard: a ▸ paqad header, an at-a-glance overview of stat tiles (verdict, stages, duration, rules, retrieval, receipt, AI-BOM, delivery) with colored indicators, and a CSS-only submenu that drills into each section's detail — so the first glance is the summary and every detail is one click deeper, all still zero-JS and self-contained.

v1.56.0

Minor Changes

• 997bbc0: RAG retrieval was dark: local-model cosine scores on hybrid-fused results rarely reached the 0.75 precision floor, so every query fell back below-floor and the model never received a slice. Retrieval now uses floor-with-relief — the 0.75 floor still marks high-confidence hits, but when nothing clears it the top slices at or above a new rag_relief_floor (default 0.35, chosen from live probe data) are delivered tagged low-confidence instead of nothing. When retrieval still finds nothing, the session-context artifact carries one honest line naming the best score instead of silently omitting the section.

  Also wires the previously-defined-but-never-recorded used RAG-evidence event: the background context worker now records what it actually delivered into the artifact (injected sections, slice/pointer counts, top score, bytes) into the per-feature .paqad/ledger/feature-evidence/<feature>/rag.jsonl, so you can see whether RAG was used — not just that retrieval ran. Adds a paqad-ai rag probe "<query>" verb that prints pre-floor top-10 fused scores for diagnosis.