CMI v0.14.1

Maintenance hotfix for Codex consumer activation after field reproduction on a real project.

Fixed

• Codex registry fallback now bootstraps the exact activating CMI version non-interactively instead of using npx --no, so a consumer without a local or cached package can start the CMI MCP server.
• Generated Codex MCP configuration now binds both cwd and CMI_PROJECT_ROOT to the activated repository root so durable Session and Change writes do not resolve against an unrelated client working directory such as /.
• Activation records the root-bound managed-configuration portability limit and requires re-running activation after moving or cloning the project.

Evidence limits

• Field reproduction established these two activation failures and the successful bounded workaround on one real Codex project; this hotfix does not establish universal Codex compatibility.
• Absolute repository-root binding is managed integration state, not portable durable project truth.
• This hotfix does not establish productivity improvement, complete security certification, or v1 readiness.

CMI v0.14.0

Feature release for proactive repository-local agent workflow, portable consumer activation, bounded integration ownership, published Portable Evidence compatibility, and cross-platform/security hardening after v0.13.0.

Added

• Expanded the pinned real corpus from 3 to 7 repositories with Python, Go, Rust, and PHP engineering-validation probes.
• Added the public Deprecation Policy for CLI, MCP, durable-state, and shipped Skill contracts.
• Added the public Portable Schema Compatibility policy and executable v2/v3 compatibility matrix; unsupported bundle schemas fail closed.
• Added a proactive repository-local agent workflow contract covering constraint-first discovery, a live ephemeral checklist, autonomous progress through authorized phases, failure recovery, and evidence-separated reporting.
• Added CMI Provenance Mark v1 for activated consumer workflows, backed by actual durable Session evidence when available and explicitly degraded when it is not.

Changed

• Portable consumer activation now binds MCP and CLI integration to an exact local CMI package entrypoint when valid local package evidence exists, with bounded project-local fallback behavior instead of relying on a bare global cmi command.
• Activation-managed integration boundaries are explicit: normal consumer tasks preserve the bounded CMI-managed blocks in AGENTS.md, .codex/config.toml, and .gitignore; those sections may change only for an explicitly requested CMI integration or activation change.
• Portable Evidence compatibility preserves released v2 semantics, uses v3 manifest-provenance integrity coverage for current writers, and does not promote integrity checks into authenticity or authorship claims.

Fixed

• Hardened the Windows MCP fail-closed portability path so the child process is fully terminated before temporary-directory cleanup, with bounded tolerance only for legitimate transient Windows lock timing.
• Removed the package-bin validation check-then-use race while preserving package-boundary, existence, regular-file, and expected-node-shebang validation with deterministic errors.

Evidence limits

• The real-repository corpus is engineering validation of CMI on pinned source trees; it does not establish target-application correctness, language-complete parsing, universal agent compatibility, or productivity improvement.
• The Provenance Mark records CMI workflow participation and evidence provenance; it is not authorship, authentication, signing, certification, approval, or verification by CMI.
• Portable Evidence digests establish integrity only; they do not authenticate a creator, prove source authorship, or turn a bundle into an authenticated backup.
• The former high-severity CodeQL file-system-race finding is fixed on the release base, but CodeQL is a supporting scan and not complete security certification.
• Existing product-value results remain descriptive-only, production Evidence Contract v2 remains simulation-only and unexposed, and this release does not establish v1 readiness.

CMI v0.13.0

Feature release for operational sharing trust and stronger real-repository/product-evidence validation after v0.12.1.

Added

• Added Operational Trust as an additive read-only cmi-trust binary with doctor and export gates for Git-sharing policy, bounded credential-like-content scanning, stable diagnostics, cross-platform path canonicalization, and fail-closed share-candidate checks.
• Added a pinned real-repository corpus plus failure-preserving execution/transport infrastructure, dedicated CI, machine-readable evidence, and engineering-validation reports that preserve exact revisions and never execute target-repository tests or application commands.
• Expanded the external empirical-study harness with blinded-review provenance, reviewer assurance/blinding boundaries, externally observed timing support, and stricter productValueEligible requirements.
• Added a lightweight Product Value Regression workflow for routine continuation/context/impact feature checks so normal engineering does not require a full claim-grade study.

Changed

• Product-value and real-corpus evidence now distinguish engineering correctness, protocol eligibility, product-value-review eligibility, paired observations, limitations, and claim boundaries instead of collapsing them into a single success signal.
• Real-repository execution now preserves partial artifacts before failure and supports an exact-revision transport fallback while keeping repository identity and revision pinning auditable.
• Public release/security metadata is advanced to the v0.13.0 supported line.

Fixed

• Fixed the CLI large-stdout completion path so successful commands can flush bounded large output before process exit instead of truncating externally consumed results.
• Hardened real-corpus orchestration and transport error handling so failed repositories remain visible as failed evidence rather than disappearing from the aggregate.

Evidence limits

• The real-repository corpus is engineering validation of CMI behavior on pinned public repositories; it is not a productivity benchmark and does not prove application correctness in those repositories.
• The first controlled product-value pilot remains descriptive-only: answer-quality results were mixed/neutral, and a favorable continuation/handoff wall-clock observation does not establish a causal productivity effect.
• No pair is productValueEligible without blinded externally-verified human review; current agent-blinded reviews remain QA evidence only.
• cmi-trust is a conservative pre-share guard, not DLP, malware scanning, authentication, or proof that content is safe to disclose.
• This release does not establish proven productivity, universal-agent effectiveness, production Evidence Contract negotiation, or v1 readiness.

CMI v0.12.1

Maintenance hardening release for CLI/MCP request boundaries and Portable Evidence trust integrity after v0.12.0.

Fixed

• Hardened top-level CLI parsing so unknown short options, duplicate single-value flags, and extra positional arguments on fixed-arity commands fail closed instead of being absorbed or ignored.
• Added dependency-free runtime validation for the JSON-Schema subset advertised by CMI MCP tools, including required fields, enums, numeric bounds, nested conditions, strict object boundaries, and additionalProperties: false; malformed known-tool calls now fail with JSON-RPC -32602 before business logic or durable writes.
• Hardened Portable Evidence restore/rebind artifact addressing against cross-platform case/NFC aliases, Windows reserved names and unsafe path segments, and case-insensitive portable-provenance.json collisions.
• Revalidated complete generated rebind provenance before reuse and re-ran project compatibility immediately before staged restore commit so source/repository/policy TOCTOU changes fail with CMI_PORTABLE_READ_RACE and do not install partial .codex-memory state.
• Moved new Portable Evidence writers to manifest schema v3 and added a separate integrity.digest (manifest-provenance-v1) that binds manifest provenance metadata which was outside the released v2 core identity.digest.
• Prevented unbound v2 project.location metadata from promoting restore compatibility to exact; v3 origin location may contribute to exact only because it is integrity-bound.
• Preserved released v2 read/restore behavior and the bounded relocated-v2 rebind-provenance compatibility path without allowing old unbound-origin exact provenance to regain trust.

Added

• Added negative Evidence Contract compatibility mutation tests that reject protected consumer-semantic breaks at exact evidence addresses while retaining additive-field compatibility.
• Added simulation-only dual-version upgrade, unsupported-version refusal, capability-discovery, and discovery-to-negotiation TOCTOU gates for a possible future Evidence Contract evolution.
• Added an explicit PRODUCTION_CONTRACT_SURFACE_NO_GO gate: runtime Evidence Contract discovery/negotiation remains absent while only v1 is runtime-supported and no named production consumer requires negotiation.
• Added regression coverage for the new CLI/MCP malformed-request boundaries, cross-platform Portable Evidence restore/rebind hardening, and Portable Evidence v3 manifest-provenance integrity.

Compatibility

• Evidence Contract v2 remains simulation-only and is not advertised as runtime support; no production discovery endpoint, negotiation parameter, handshake token, CLI flag, MCP method, or automatic downgrade/upgrade behavior is added by this release.
• Portable Evidence schema v2 remains inspectable/restorable under its released core identity algorithm. New writers emit schema v3 with separate manifest-provenance integrity coverage.
• The deterministic core identity.digest is not silently redefined for released v2 bundles; v3 adds integrity coverage rather than reinterpreting old v2 digests.

Evidence limits

• Portable Evidence digests provide integrity checking, not authentication, signatures, source authorship proof, or backup authenticity; authenticated remains false.
• The Evidence Contract versioning/discovery/handshake work in this release is executable simulation, not a production interoperability claim.
• This release does not establish productivity improvement, time savings, universal-agent effectiveness, or v1 readiness.

CMI v0.12.0

Feature release adding Evidence-Anchored Rule Intelligence so CMI can connect reviewed rules and findings to concrete source locations without promoting source matches into product truth.

Added

• Added structured evidence anchors that can carry project-relative file paths, line ranges, symbols, features, and commit references.
• Added portable evidence syntax for source:, symbol:, feature:, and commit: references, plus bounded extraction from existing finding evidence and related files.
• Added explicit finding verification semantics: suspected → observed → established, while durable resolved remains the lifecycle resolution state.
• Ambient Intelligence now tells agents to inspect affected source and capture line/symbol/feature/commit evidence when reviewed project rules are relevant to a task.
• Closing Intelligence now carries and renders evidence anchors and verification state, and violationEstablished is tied to established verification rather than reviewed-rule relevance alone.
• Extended the findings registry schema additively with optional verificationState and evidenceAnchors fields while retaining schema version 1 compatibility.
• Added focused evidence-anchor tests and docs/EVIDENCE_ANCHORED_RULE_INTELLIGENCE.md covering the trust boundary and intended lifecycle.

Evidence limits

• Source or static evidence can establish that code exists at a cited location; it does not by itself prove a runtime-visible, user-visible, design, architecture, or policy violation.
• established requires verification appropriate to the rule, such as a test, runtime/browser observation, or explicit human review; reviewed-rule relevance remains advisory until then.
• Line ranges can drift as source changes, so CMI pairs them with symbol/feature/commit context where available; this is not compiler-grade symbol tracking.
• This release does not add an autonomous generic rule scanner and does not hard-code any Project 001-specific rule or detector.
• No productivity, time-savings, universal-agent, or v1-readiness claim is established by this release.

CMI v0.11.2

Maintenance patch for graph-drift signal quality discovered through repeated real-project Project 001 sessions.

Fixed

• Kept stale graph evidence fail-closed for graph/impact claims while distinguishing the narrow case where every stale source node is fully explained by the just-completed session's attributed source mutations.
• Downgraded that expected current-session source-only drift from a material Closing warning to a non-blocking refresh reminder; pre-existing, unexplained, missing-node, source-set, resolver/workspace, scan-policy, discovery, and generated-format drift remain material.
• Added bounded stale/missing graph path evidence so session attribution can explain expected cache invalidation without declaring the graph current.
• Updated managed activation guidance to require refresh before relying on stale graph/impact evidence while explicitly forbidding cosmetic scans merely to produce Closing CLEAN.
• Added regression coverage for both uncommitted and committed-clean-worktree source changes, including the exact field pattern observed in Project 001.

Evidence limits

• This patch does not make stale graph evidence healthy and does not auto-run cmi scan; graph/impact evidence remains blocked until generated intelligence is refreshed.
• The Project 001 observation establishes a concrete signal-quality defect and regression scenario, not a universal productivity or agent-effectiveness claim.
• Historical prediction-gap and other finding semantics are unchanged by this patch.

CMI v0.11.1

Maintenance release for the public licensing and project-identity cutover. No CMI product-behavior change is claimed by this release.

Added

• Added LICENSING.md documenting the source-available licensing model, the legacy MIT boundary, and the separate commercial-license path.
• Added NOTICE and BRAND_POLICY.md so official-project provenance, attribution, and fork/brand boundaries travel with the repository and npm package.
• Added a structured real-repository field-feedback issue template focused on useful signals, noisy/misleading behavior, missing behavior, reproducible evidence, and public-data safety.

Changed

• Repository source after the 2026-08-11 licensing cutover is offered under PolyForm Perimeter License 1.0.1 rather than MIT.
• v0.11.0 and all earlier public releases remain under the MIT terms shipped with those versions; this release does not retroactively revoke previously granted MIT rights.
• npm package metadata now uses SEE LICENSE IN LICENSE and includes LICENSING.md, NOTICE, and BRAND_POLICY.md in the published package.
• README terminology now describes post-cutover CMI source as source-available, documents the licensing boundary, and invites evidence-based field feedback from real repositories.
• Material code contributions require pre-coordination until a formal contributor licensing agreement is available; bug reports, field evidence, documentation suggestions, and design feedback remain welcome.

Evidence limits

• The licensing/identity changes reduce ambiguity around future competing products and official-project provenance; they do not erase rights attached to prior MIT releases.
• PolyForm Perimeter is a source-available license and this release should not be described as OSI-approved open source.
• No new productivity, effectiveness, compatibility, or universal-agent claim is established by the licensing maintenance release.
• Product field evidence continues to be bounded by the documented v0.11.0 Grok/Codex results unless separately re-evaluated.

CMI v0.11.0

Feature-complete planned Skill inventory and post-v0.10.0 agent-integration work. Publication remains separately authorized after this release-preparation commit.

Added

• Added one-time cmi activate integration for Codex plus a bounded generic mode so supported agents can use CMI after project activation without requiring CMI-specific user prompts.

• Added Ambient Agent Intelligence: read-only ambient task routing through cmi ambient and MCP get_ambient_task_brief, with conservative short-prompt intent classification and evidence-linked context/workflow guidance.

• Added CMI Closing Intelligence: a bounded end-of-work read model and branded ### CMI Intelligence footer that surfaces up to three evidence-based cross-session, verification, finding, and reviewed-consistency alerts plus a clean fallback.

• Added eight Agent Skills open-format adapters as portable repository/package artifacts:

  • cmi-ambient-brief
  • cmi-continue
  • cmi-evidence-health
  • cmi-closing
  • cmi-memory-review
  • cmi-work-session
  • cmi-change-loop
  • cmi-activate

• Added npm distribution of the eight Skill artifacts under package files → skills/ with package-smoke coverage for packed identity and no auto-activation on install.

• Added Skill contract tests for all eight adapters plus distribution invariants.

Fixed

• Preserved intentionally unfinished / partial / paused / review-pending Change Intelligence records as active when a work session ends (session completion != Change completion). Partial progress remains visible under active Changes and Closing Intelligence surfaces it as a non-blocking reminder; explicit abandonment remains terminal.
• Required durable closed-session evidence for Closing Intelligence rather than synthesizing Closing from health-only state.
• Distinguished missing graph evidence from graph drift and kept read-only session scope separate from mutation evidence.
• Resolved project-local CMI CLI fallback resolution for packaged/local entrypoint usage.

Changed

• Managed activation instructions now require supported agents to retrieve Closing Intelligence before ending substantial work and append a concise user-visible CMI footer when authoritative Closing exists.
• Session continuation/handoff remains historical evidence that must be re-checked against current baseline, per-ID Change lifecycle, and open findings.
• Skill distribution contract: the npm package ships Skill artifacts; npm installation does not activate or install them into agent runtimes; cmi activate does not install Skills; CMI has no native Skill loader.
• Unchanged scans keep tracked architecture.md byte-stable by excluding volatile scan timestamps/parser reuse/duration metrics from the human-reviewable architecture summary.
• Active work-session state is stored under ignored transient CMI storage and materializes into durable reviewable sessions/ evidence only when finalized.
• Repository baselines preserve product-scope compatibility while separately reporting raw Git cleanliness and omitted untracked CMI-internal paths.
• Existing relative CSS/static imports are classified as non-code local dependencies instead of unresolved source imports.
• Generic memory-gap and regression-test suggestions are suppressed when no task-specific files, boundaries, or topic evidence support them.

Evidence limits

• Final Codex S0–S7 field acceptance on subject c05098fa82ddf85a4443e3769801baf78e12c200 was not completed. Mission 1.8B was blocked before S0 because the available ChatGPT-auth Codex runtime reported model capacity exhaustion, and an API-key-auth fallback was not available in the operator environment. This is recorded as a runtime limitation, not evidence that CMI passed or failed those final scenarios.
• Package shipment of Skills does not prove runtime Skill discovery or automatic selection by any agent.
• No universal Codex/Grok Skill installation path is claimed; observed paths remain runtime/version-specific evidence.
• Ambient intent routing is deterministic advisory classification, not autonomous authorization or proof of user intent.
• Agent activation cannot force clients that ignore repository instructions or MCP to follow CMI.
• Reviewed design/architecture/policy relevance in Closing Intelligence is a consistency-check cue, not proof of a violation.
• Static parsing and impact/boundary output remain heuristic/advisory rather than compiler-grade.
• This release does not establish productivity improvement, time savings, general product value, or v1 readiness.

CMI v0.10.0

This release candidate contains the reviewed Phase 1–3 work after v0.9.2. It is prepared for release review; publication remains separately authorized.

Added

• Added bounded portable project-evidence bundles with deterministic manifests, SHA-256 artifact verification, path-independent identity, exact/relocated/Git-checkout/content-only compatibility outcomes, explicit restore and rebind operations, destination-conflict protection, and recorded rebind provenance.
• Added executable provenance for the actual runtime/script, package root and version, source-checkout revision and cleanliness where available, install kind, observable candidates, and genuine multi-install ambiguity, with CLI/MCP parity.
• Added actionable uninitialized-project recovery and configuration/evidence health diagnostics across status, doctor, search/context, prepare, and impact; human and JSON trust-critical outcomes now share blocked semantics.
• Added persistence compatibility evidence for the audited v0.5.0 config/memory/index/graph floor, v0.7.0 changes, v0.8.0 sessions, v0.9.0 findings, and v0.9.1 evaluations, including no-rewrite checks and the exact bounded v0.8.0 fallback exception.
• Added fail-closed handling for future or corrupt durable/config/generated formats, preserving bytes and refusing ordinary downgrade or overwrite paths.
• Added regression coverage for portable evidence, executable provenance, operational diagnostics, MCP gating, persistence compatibility, future-format protection, and adversarial filesystem cases.
• Added maintainer/evaluator-side empirical study ledger and harness support for reproducible paired plain-vs-CMI study bookkeeping. The harness is not an agent-facing CMI command and does not establish productivity, time-savings, or general product-value evidence.

Changed

• Repository-baseline summaries now omit only untracked local .codex-memory/ state; tracked, staged, renamed, and ordinary project changes remain visible.
• Portable evidence now binds the bounded scan, ignore, resolver, and workspace inputs needed to reproduce source boundaries after relocation.
• Durable compatibility is read-only/no-rewrite for the audited historical floor; generated state may be rebuilt only when its format is obsolete and supported, while unsupported state remains blocked.
• The release candidate keeps MCP mutation tools hidden or rejected by default; explicit write mode is required for portable-evidence and durable mutations.

Compatibility

• The audited historical floor is bounded and representative, not a promise to support every pre-v1 commit or every future schema. No explicit migration command is required for the audited fixtures.
• Future memory metadata, configuration, graph, and index formats fail closed without ordinary scan/refresh mutation or byte overwrite.

Evidence limits

• Static parsing and impact output remain heuristic/advisory rather than compiler-grade or complete runtime analysis.
• Portable bundles provide integrity checking, not authentication, backup authenticity, or source-authorship proof.
• Executable provenance reports observable runtime/install evidence; it does not prove source authorship and preserves ambiguity when multiple installs are visible.
• Evaluation and empirical records remain observational/caller-attested where documented. This release does not independently prove productivity, time savings, or general product value.
• Study 001 remains incomplete and Study 003 remains unreconciled; no new empirical study was run. This release does not claim v1 readiness.

CMI v0.9.2

Fixed

• Enforced the default MCP read-only boundary by removing scan mutation from safe tool discovery and rejecting direct scan calls before generated CMI caches can be written unless write mode is explicitly enabled.
• Made graph freshness cover the complete source candidate set plus resolver, workspace, scan, and ignore inputs so new files and configuration drift block impact analysis instead of returning stale confidence.
• Separated source/project freshness from semantic memory review: source refresh now yields fresh-source, while reviewed-current requires explicit review provenance that remains current relative to source refresh.
• Made unsafe, unreadable, symlinked, or oversized durable memory fail closed as blocked evidence rather than silently appearing empty; search and refresh cannot use blocked memory as trusted input.
• Made an existing invalid, unsafe, or oversized findings registry fail closed with CMI_FINDINGS_BLOCKED so normal finding/session mutations cannot overwrite corrupted evidence.
• Hardened CLI automation contracts: unknown options fail, JSON-mode errors remain machine-readable, and blocked impact exits non-zero while preserving structured output.
• Aligned memory-review JSON Schema conditions with runtime provenance validation.
• Reduced representative heuristic parser errors by ignoring JavaScript import-like text in comments/ordinary strings and resolving reviewed Python relative-import and Rust crate:: cases.
• Aligned cmi doctor with the Unified Evidence Health model so blocked graph, impact, durable-memory, or overall evidence states fail diagnostics instead of being reported as warning/pass states.

Added

• Regression coverage for safe MCP direct-call bypass attempts, source-set and resolver drift, semantic-review separation, blocked durable memory, corrupt findings no-overwrite behavior, strict CLI contracts, graph/schema freshness, parser edge cases, and blocked doctor diagnostics.
• docs/INDEPENDENT_REVIEW_REMEDIATION.md documenting the finding-by-finding remediation contract and the empirical gaps that remain intentionally open.
• docs/EMPIRICAL_VALIDATION.md defining a paired controlled comparison protocol for plain Codex/Git/source search versus Codex + CMI without converting synthetic or caller-attested data into productivity claims.

Evidence limits

• CMI remains heuristic/advisory rather than compiler-grade semantic analysis.
• Evaluation repository/reviewer provenance remains caller-attested unless a stronger external assurance process is used.
• This release improves trust boundaries and testable correctness; it does not claim that CMI productivity value is independently proven across real repositories.