Skip to content

v3.7.1

Choose a tag to compare

@github-actions github-actions released this 15 Nov 18:54
· 140 commits to main since this release

This is release 3.7.1. Although numbered as a patch by owner decision, it
drops state and commands from before 3.7 and changes 3.7.0 behavior (review
routing, CLI surface); read Changed and Removed before upgrading. A v3.7.0
project upgrades in place.

Added

  • Pi host adapter. cloop init --host pi (or install-pi.sh) installs
    the project-local .pi/extensions/changeloop.ts for Pi >=1.5.2. It registers
    the canonical commands, loads the package-owned agent contract, and bridges
    built-in tool secret/phase guards, lint feedback, session identity, and
    !/!! shell execution. /land <change> grants Land for that change only
    until reload or new user input. Guard coverage is partial: custom/MCP tools
    and other extensions are outside the live boundary, session-digest import is
    unavailable, and no OS sandbox attestation is issued.
  • Shared Pi runtime. cloop init <target> --host pi --shared-runtime keeps
    the runtime, hooks, skills, commands, and OpenSpec schemas in a
    content-addressed bundle under Pi's agent directory. The project holds only
    its .changeloop.json pin, foundation.json, OpenSpec settings and
    artifacts, and .foundation state. A missing or mismatched pin fails closed.
    After the global extension is installed, /changeloop init then /reload
    configures another trusted project. Existing local installations are not
    migrated and remain supported; project settings survive re-initialization.
  • Hotfix regression for its incident. A new hotfix draft must include a
    scenario whose regressionOf equals bugfix.incident.link, in a requirement
    with an executable capability; otherwise change start returns an agent
    EDIT. The change records regressionRequired: true and change revise
    enforces it only for marked changes, so hotfixes created before the upgrade
    are never blocked. The compiled spec shows "Regression of: "; evidence
    claims and protocol pins are unchanged.
  • Project incident records and lessons. Install and upgrade seed
    openspec/incidents/README.md, openspec/incidents/_template.md, and
    openspec/lessons.md only when missing, and never overwrite them.
  • Lessons as grounding. When openspec/lessons.md exists, investigation and
    change intake read it as a grounding source.
  • Advisory project review rules. The AI reviewer packet (review packet schema
    7) now carries the project's REVIEW.md as reviewRules, shown to the
    reviewer as advisory guidance that never changes pass or fail. Review receipts
    record review.reviewRulesHash; editing, deleting, or creating REVIEW.md
    after a review makes it stale (review-rules-changed), while receipts recorded
    before this release stay valid. A missing or empty REVIEW.md keeps the
    previous behavior. Install and upgrade seed a starter REVIEW.md only when
    the project has none, and never overwrite one.
  • test-design skill, vendored unchanged from hashtagf/skills 0.16.0. It
    designs and documents what to verify: test cases, test plans, UAT, coverage
    audits, bug reports, and release go/no-go. testing-fundamentals still owns
    writing automated tests; the router splits the two.

Changed

  • Internal: the advance, change-lifecycle, proof-readiness, and agent-planner
    runtime factories now require the collaborators the composition root always
    supplies and fail at construction naming a missing one; the null guards for
    those impossible states are gone. Never-passed parameters became constants,
    and target-edit and atomic-start rollback logic moved out of foundation.mjs
    into their workflow modules. No output, CLI, or protocol change.
  • Breaking (CLI surface): primitives that advance already performs are
    now internal: proof readiness, proof advance, proof run,
    proof collect, evidence detect, evidence init, evidence doctor, and
    budget checkpoint. They are hidden from help, the default describe
    listing, and the public CLI tables, but still dispatch (and answer
    --help/describe <name>), so existing scripts keep working. Every runtime
    next/resume/recovery route that named one (receipt validity recovery,
    readiness verify/retry choices, wiring actions, waiver output, Build
    completion, verification plans, update advisories, budget and blocker
    recovery, the moved-target Land option) now names
    advance <change> --through proven|archived or advance <change> --inspect.
    The public command golden fixture drops from 76 to 68 commands.
  • The deterministic suite now has two tiers (repository-only). Plain
    .claude/tests/run-all.sh is the fast local gate before commit (about three
    minutes): every suite except the SLOW tier of real-Git end-to-end,
    mutation, installer, and topology suites and the exclusive benchmark.
    --quick is the same gate. New run-all.sh --full runs the complete set
    (the former default); CI runs it as --release on every push and pull
    request, and it must be green before merge or release. --affected still
    selects slow suites whose owned files change.
  • Each harness command asks Git less: repeated git status, git ls-files -s, and git rev-parse HEAD listings of an unchanged workspace within one
    process are answered from a per-process memo instead of a new spawn. The
    memo is forgotten on any file write the process makes into that workspace
    (or into a .git directory), on any child process other than a read-only
    Git query, on a lock wait, and on every clearSnapshotCache. In one
    build→prove→land→sync flow, advance --through proven went from 22 to 10
    Git spawns, Land from 48 to 29, and the whole flow from 134 to 89.
  • Installation copies each managed directory's files in one batch instead of
    forking dirname, mkdir, and cp per file (about 1,450 processes down to
    about 320). A fresh install is roughly twice as fast; the installed tree is
    byte-identical.
  • The deterministic test gate is faster and steadier: one pool worker per core,
    the slowest suites start first, the workload suite runs once (inside the
    sentinel), end-to-end fixtures copy one installed consumer instead of
    reinstalling, and run-all.sh --quick skips only the slowest suites for the
    inner loop. Six test files that no runner named now run in the gate, and a
    registration check keeps new ones from going unrun.
  • run-all.sh clears every host-exported FOUNDATION_* variable and
    CLAUDE_FOUNDATION_PROJECT before running suites, so a run started from an
    agent session (Claude or Pi) tests its fixtures instead of the caller's
    project and guardrail mode. The affected-suite selector also works when the
    checkout sits under a symlinked directory such as macOS /tmp.
  • The default run-all.sh gate no longer runs the release-only openspec native sentinel, which reran the whole workload suite alone after the pool
    (about six minutes). run-all.sh --release runs it with the full gate; the
    release workflow, CI, and --affected for benchmark edits still run it, and
    npm run release:preflight checks that the release workflow uses
    --release. The multi-repository topology contracts and the concurrent
    repositories seam are split into parallel rows and files with every
    assertion kept, so neither holds the tail of the gate.
  • The AI reviewer now checks scope discipline in the changed hunks:
    unrequested abstraction, dead or unused code, speculative options, guards or
    retries that hide a symptom, and untraceable or unrelated edits. These are
    minor findings, major only when they leave a scoped claim false or unproven;
    root-cause fixes and behavior-preserving preparatory refactors are not
    flagged. Packet schema, receipts, and protocol pins are unchanged.
  • ux-ui-principles is now the unchanged upstream hashtagf/skills 0.16.0
    copy instead of the adapted bundle, including its runtime assets/ (about
    13 MB) and scripts/; upstream evals/ is not shipped. Upgrade removes the
    adapted bundle's files, including references/upstream.md and
    references/change-loop-handoff.md. Each vendored skill records its source
    in VENDORED.md.
  • Every non-lifecycle skill now works on its own: guidance that needs
    OpenSpec or Change Loop is a conditional "With OpenSpec/Change Loop …
    otherwise …" section, links to other skills are optional reading, and dead
    references are removed; references and scripts follow the rewritten
    SKILL.md files. Skill descriptions name sharper English and Thai triggers
    and the neighboring skill to use instead. coding-discipline is rewritten:
    fix the root cause where the first wrong state appears, treat a
    behavior-preserving preparatory refactor as its own in-scope slice, record
    unrelated issues as follow-ups instead of fixing them, validate inputs before
    irreversible effects, and verify each outcome on the real artifact. The
    router loads it for a bug fix or Build edit whose scope, assumptions, or diff
    size the always-on digest does not settle.

Removed

  • Breaking: state from releases before 3.7 is no longer read. Finish or
    abandon a change with the release that started it before upgrading.
    • migrate and the .workflow/ reader are gone.
    • evidence upgrade and evidence schema 1 are gone (protocol pin
      evidenceSchema is now ["2"]); an evidence v1 packet is refused with
      that route.
    • The retired commit-based multi-repository Land saga is gone, with its
      land record and land pointers commands. Land always delivers uncommitted
      workspaces; land resume only reports the current plan. A change 3.7.0 or
      earlier left inside that saga is refused instead of replayed: advance
      asks the legacy-land-saga decision (finish it with 3.7.0, then upgrade,
      or change abandon it), naming .foundation/runtime/<change>.json and the
      saga fields; changes, doctor, and the upgrade report name it too.
    • A single-number execution.packetBytes is refused at runtime; the
      installer converts it to the same task/review/repository/global limits.
  • The historical phase-orchestrator benchmark runner, its baselines, the
    .workflow/ scenario fixtures, and the unregistered phase matrix were
    removed from the repository (not shipped).
  • Breaking: only semantic drafts v4 (JSON) and v5 (folder) are accepted.
    Drafts v1, v2, v3 and unversioned JSON drafts return an EDIT that names the
    v5 template; change new --draft now always routes to change start.
    change new remains as an internal fixture primitive and leaves the public
    command list.
    A change 3.7.0 already started from a v3 draft keeps working and can still
    be amended with change amend <change> <amendment.json>.
  • Breaking: the deprecated CLI aliases are gone. Use the canonical
    commands instead: validate → change validate, proof plan →
    proof readiness, proof finish → proof run, agents task →
    packet <change> --task <task>, and the runtime <command> pass-through →
    the matching public command (runtime hash → hash, runtime new →
    change new). Each alias had printed a deprecation warning since it was
    replaced; installed commands, skills, and hooks never call them.
  • Breaking: the legacy review routing is gone; review always routes by risk
    tier.
    workflow.reviewPolicy accepts only risk-tiered and
    workflow.reviewCircuit only full-delta, which are now also the runtime
    defaults. A legacy value is refused with the route to rerun
    claude-foundation init, and install/upgrade rewrites both keys to the
    shipped values. Every review receipt now binds to an authority dispatch
    attempt, and RESOLVED no longer prints review: required; it names the
    tier and model class. In a 3.7.0 project that opted into legacy, upgrade says
    so, an unrecorded legacy-circuit review request is replaced by a current one
    automatically, and a change proven under it gets one current review on its
    next advance.
  • Twelve unused internal runtime functions (old proposal renderers and
    helpers replaced in 3.7). No command, output, or file format changes.
  • Breaking (advance protocol 7): advance output no longer carries
    resumeCommand, the duplicate of resume kept for v2 host adapters. No
    shipped adapter, hook, command, or skill read it; read resume. Deliver and
    budget checkpoint outputs keep their own resumeCommand.
  • Breaking (grounding schema ["2", "3"]): a grounding.yaml version 1
    is refused with the accepted versions. No release wrote one: grounding
    shipped at version 2 in 3.2.22, and 3.7.0 writes version 2 or 3. Version 2
    stays; current templates and change creation still write it.
  • The pass-through modules runtime/evidence/codex-reviewer.mjs (and its
    createCodexReviewerRuntime alias) and hooks/phase-guard-policy.mjs.
    Import runtime/evidence/configured-reviewer.mjs and
    runtime/core/shell-mutation-policy.mjs; upgrade removes the old files.

Fixed

  • Path containment is decided by one shared rule (core/paths.mjs) in the
    runtime and hooks. Several checks treated a child directory whose name
    starts with .. (for example ..cache) as outside its root; it is now
    inside, while .., ../x, and absolute paths elsewhere stay outside.
    Duplicated validation-field, isObject/unique, realpath, and SHA-256
    helpers now share one implementation each; validation messages are
    byte-identical.
  • Test and benchmark tooling (repository only). run-all.sh puts stand-in
    claude/codex first on PATH for every suite and fails a suite that calls a
    model (readiness probes get a deterministic "not logged in"). The benchmark
    lab no longer records an empty baseline product.patch when runner coverage
    under .foundation/ overflowed the diff buffer, isolates the operator's
    user-level Claude Code settings and MCP servers for every arm (recorded in
    provenance.hostSettings), fills requestedModel/actualModel from the host
    stream, and aggregate.mjs --compare shows every arm and keeps Change Loop
    sources apart (by commit or --label).
  • advance no longer stops with "ran RUN_INVALIDATED_EVIDENCE twice without
    changing delivery state" on a proven change whose receipts were re-earned on
    an unchanged workspace (for example a 3.7.0 change re-reviewed after
    upgrade). Proof advance reused the old proof, which no longer bound the new
    receipts; it now finalizes a new one.
  • change start --inspect on a v1–v3 draft returned DONE although
    change start then refused it; it now returns the draft repair with the
    supported formats.
  • A review policy with no named reviewer and fallbackReviewers: ["main-session"] (the fix the missing-CLI error recommends) now reviews in
    the main session. Before, validation accepted it but advance repeated a
    "without a recorded configured reviewer failure" repair forever. A named,
    installed reviewer still runs first; the main session is reached only after
    its recorded failure.
  • Blocker or major findings from a human review now must be closed. The next
    review, human or AI, full or delta, carries them as closureFindings and
    must list them in verifiedFindingIds (or fail with a blocker or major
    finding for what remains). Before, closure was owed only to AI findings, so
    a later review could pass without checking the human blocker, and a
    deterministic repair closure could pass over it. AI-only review chains are
    unchanged.
  • A review receipt whose finding counts were edited after it was recorded
    (for example review.findings.verified raised to 99) no longer reads valid:
    the counts must match the finding rows and verified IDs the dispatched
    attempt binds, as they always do when the harness records the receipt.
  • After the harness rebuilt a corrupt review attempt chain, every later
    review dispatch, a human review included, failed with "a completed AI
    response has no matching recorded receipt". The rebuilt placeholders carry
    no response, so they no longer count as one that owes a receipt.
  • Hooks now work for a project reached through a symlinked path (macOS
    /tmp, a symlinked home directory). Node entry points compared the
    logical script path against their resolved module URL, so the secrets guard
    refused secret reads instead of showing a redacted copy, and the
    detached-authority, shell-route, and dev-terminal hooks silently did nothing.
  • An AI review whose cheaper first round (fast or standard model) was
    interrupted while its receipt was recorded now recovers instead of failing
    "configured review result integrity or reviewer provenance does not match";
    recovery compares against the reviewer the request was dispatched to.
    authority abort now refuses a request whose AI result completed without a
    receipt, which previously blocked every later review of the change.
  • A harness command run inside one project no longer changes another project
    pinned by the host session's CLAUDE_FOUNDATION_PROJECT (the Pi adapter
    exports it to every shell). The runtime and CLI refuse with a clear message
    unless the cwd or the runtime belongs to the pinned project or its sandbox;
    --project still selects a project. Shell suites run alone also drop the
    host session's project, identity, and guardrail variables.
  • skill-creator helper scripts (run_eval.py, run_loop.py,
    improve_description.py, package_skill.py) run by path as well as with
    python -m, package_skill.py prints its real location in usage text, and
    quick_validate.py names the missing PyYAML dependency instead of a
    traceback. ui-ux-pro-max shows search.py relative to the skill's own
    directory instead of a skills/... path that did not exist in installed
    projects.