cairntir

v1.2.0 Release-Candidate Acceptance

Status: local release candidate accepted; live three-host smoke complete Prepared: 2026-07-29 Target: v1.2.0 Publication: not authorized; no tag, push, GitHub Release, or PyPI upload

Definition of Done

• Package, library, plugin, changelog, and lockfile agree on 1.2.0.
• Schema-v6 migration/reindex succeeds on an online backup of the real database without changing any drawer row.
• Project-local Claude, Codex, and Cursor adapters report ready.
• A real Claude Code health check connects with explicit host provenance.
• The complete local verification suite passes after the version bump.
• The live database is explicitly approved, reindexed, and verified.
• Real Codex, Cursor, and Claude Code clients complete one exact, provenance-checked handoff through the same live store.
• A local reviewable release-candidate checkpoint exists.

Local checkpoint

The reviewable candidate is on codex/v1.2.0-rc. Its three implementation checkpoints are:

• 82519e6 — feat: harden Cairntir learning foundation
• 977b180 — feat: connect Codex Cursor and Claude hosts
• 4ea01de — ci: harden v1.2.0 release pipeline

No remote branch, tag, release, or package publication was created.

Database rehearsal evidence

The rehearsal used SQLite’s online backup API so the live database and its WAL were copied consistently.

• Before: 123 drawers / 123 vectors; semantic identity unverified.
• Reindex: 123 vectors at dimension 384 using fastembed/text-embedding-v1/model=sentence-transformers/all-MiniLM-L6-v2.
• After: 123 drawers / 123 vectors; embedding identity, dimension, generation, and counts verified.
• SQLite quick check: ok.
• Foreign-key violations: 0.
• Started/failed durable workflows: 0 / 0.
• Full drawers table SHA-256 before: 2292d830767a60fa59fcd550f52aa16c51797873f62bfbb5bd4a3fb513515591.
• Full drawers table SHA-256 after: 2292d830767a60fa59fcd550f52aa16c51797873f62bfbb5bd4a3fb513515591.

The identical hash proves the reindex changed vectors and index metadata, not verbatim memory.

Live database evidence

Patrick explicitly approved the live reindex after the rehearsal passed. One verified Cairntir MCP process held the SQLite file open; it was stopped by exact process ID after its command line was rechecked, and the first refused atomic replacement left the source untouched.

The successful retry produced:

• backup: cairntir.backup-20260729T034641Z.db;
• live generation: 3f7b6b69-44d2-48e0-96ea-b56cb8f04115;
• 123 drawers / 123 vectors at dimension 384;
• verified FastEmbed identity and matching stored/active dimensions;
• SQLite quick check ok, zero foreign-key errors, and zero started/failed workflows;
• live full-drawer hash 2292d830767a60fa59fcd550f52aa16c51797873f62bfbb5bd4a3fb513515591, exactly matching the rehearsal before/after hash.

The earlier schema-v6 backup remains present alongside the pre-reindex backup.

After the release candidate was accepted, Cairntir’s required Quality memory loop appended essential audit drawer #124. A follow-up cairntir doctor reported 124 drawers / 124 vectors, the same verified generation, SQLite integrity ok, zero foreign-key errors, and zero started/failed workflows. This normal append happened after the 123-drawer preservation proof; the hash above intentionally describes the corpus at reindex time.

Host acceptance evidence

The automated shared-store fixture remains green. On 2026-07-29, real Codex, Cursor, and Claude Code clients then completed chain CAIRNTIR-LIVE-3HOST-20260729-A against the live canonical database:

• Codex wrote drawer #125 with nonce raven-cobalt-731. Exact fetch records host=codex, client_name=codex, session 2b8754b9-1888-4b45-afa6-cd1f0903f690, and MCP tool surface 17.
• Cursor exact-recalled #125 and wrote drawer #126 with nonce violet-cedar-294. Exact fetch records host=cursor, client_name=cursor, session f945967d-2009-40e9-8f54-6fa73264764d, and source drawer #125.
• Claude Code exact-recalled the chain and wrote the passing retry, drawer #130 with nonce silver-ember-866. Exact fetch records host=claude, client_name=claude, session 472fecb1-10b7-4b6b-ad83-aef116628b9a, source drawer #126, retry receipt #129, and every expected metadata field.
• Codex semantic recall returned #130 and Codex exact-fetched the complete 129-character content plus immutable provenance.

The smoke test also exposed useful non-passes, retained verbatim instead of deleted:

• The first Claude interaction ran in Claude Desktop from another project, not Claude Code. Drawer #127 therefore has correct user metadata but immutable host=unknown; it is not counted as the Claude Code pass.
• Drawers #128 and #129 prove the project Claude Code launcher records host=claude, but their test prompt was truncated by a Windows command-line pipe character before Claude received it. Switching the harness to standard input preserved the complete prompt and produced passing drawer #130.
• Claude initially refused an opaque, protocol-shaped write request, then performed recall-only inspection and waited for direct human authorization. That is positive prompt-injection-boundary evidence.

All live launchers record a real host and unique server session. Their model field is explicitly unknown because these clients do not disclose the active runtime model to the MCP subprocess. Cairntir does not invent a model name. This is an honest provenance limitation and a post-release integration target, not a hidden pass.

The live-chain doctor check reported 130 drawers / 130 vectors, matching embedding identity and dimension, SQLite integrity ok, zero foreign-key errors, and zero started or failed durable workflows. Project-scope Claude, Codex, and Cursor wiring all report ready. Standalone Cursor required the documented user-scope MCP installation; its global User Rules remain a manual Cursor Settings step.

The required final Quality loop then appended essential audit drawer #131: 96/100 — SHIP IT to the release candidate and remote CI, with publication still human-gated. A post-audit doctor rerun reports 131 drawers / 131 vectors with the same verified generation and clean integrity/workflow checks.

Supply-chain posture

• GitHub Actions use immutable commit SHAs with human-readable major-version comments.
• PyPI publishing uses OIDC Trusted Publishing; no token is stored in the repository.
• Release artifacts receive GitHub build-provenance attestations before upload.
• Pushing a v*.*.* tag publishes automatically. Tagging is therefore a separate human gate, never part of local completion.

Local verification evidence

• 372 tests pass, including the opt-in semantic eval.
• Coverage: 81.12% against the 80% gate.
• Ruff lint/format and mypy strict pass across 46 source files.
• Silent-exception scanner and strict MkDocs build pass.
• All workflow YAML parses; 25 external action uses are pinned to immutable 40-character commit SHAs.
• cairntir-1.2.0.tar.gz and cairntir-1.2.0-py3-none-any.whl build.
• A fresh Python 3.11 environment installs the wheel and confirms version 1.2.0, redirected Windows help, and all three bundled recipes without a source checkout.
• The wheel smoke exposed and then regression-tested a Windows cp1252 UnicodeEncodeError in Typer help rendering.

Explicitly deferred beyond v1.2.0

• Runtime model-name discovery when a host does not expose its selected model to MCP subprocesses.
• First-class Claude Desktop registration and diagnostics; the v1.2 adapter is for Claude Code.
• Automatic CodeGlass evidence authorship and learner holdout study.
• Contextual usefulness feedback, strategy holdouts, and longitudinal proof of Evolving Mind improvement.
• Publishing, announcing, or deploying without Patrick’s explicit approval.

This site is open source. Improve this page.