cairntir

v1.8.0 Release Acceptance

Status: published and independently verified 2026-08-30 Prepared: 2026-08-25 Tag: v1.8.0 at 8dcb3fb676242af49236fac749ac3a57456cac0b Settlement pull request: #78 Release-record pull request: #79 Publication-repair pull request: #80 Release workflow: run 33346147256

Why a MINOR

This release adds user-visible recovery and finalization capabilities, new CLI options, an opt-in MCP handoff path, and verified shared-store wiring for four additional agent hosts. Users need the changelog to know these capabilities exist and how transcript trust differs from drawer trust.

What is accepted

• Qwen Code, Claude Code, and Codex read bounded JSONL tails from their host-owned project transcript stores;
• host-specific completion markers exclude finished turns, and an exact later drawer suppresses a request already captured on arrival;
• Cursor returns an honest unsupported receipt rather than reading private, undocumented SQLite tables;
• recovered messages are whole or named, never truncated, under a separate character budget;
• recovery is disabled by default on handoff and requires cairntir recover, --recover-from, or recover_transcripts=true;
• reading and handoff never store transcript content; --write N is explicit and writes with trust=untrusted, capture_path=transcript_recovered;
• the kill-after-request fixture opens a fresh Codex handoff, excludes the live session, and returns the prior request verbatim before any drawer exists.
• Finalization Mode freezes independent acceptance tests, protects verification capacity, bounds repair rounds, and requires an explicit terminal result;
• verified MCP wiring covers Claude Code, Cline, Codex, Copilot CLI, Cursor, Gemini CLI, OpenCode, and Qwen Code while transcript recovery remains gated to its separately verified adapter set;
• all seven previously open dependency and documentation pull requests were reconciled into the protected settlement merge without dropping either 2026-08-30 proposal.

Adapter evidence

The parsers were checked against local host-owned transcripts without copying their content into fixtures:

• Qwen Code 0.21.5: type=user, message.parts[].text, and tool-running assistant parts with functionCall; the real killed 2026-08-05 request at 23:25:50Z is recovered from its bounded tail;
• Claude Code 2.1.222: type=user, string or text-block content, tool-result user records, and assistant stop_reason=end_turn completion;
• Codex 0.149.0-alpha.4.1: session_meta.payload.cwd, user response_item/input_text, turn ids, and event_msg/task_complete;
• Cursor: official documentation confirms local SQLite history but does not document a stable transcript schema. Status remains unsupported.

Automated acceptance

• Ruff format and lint
• mypy strict: 53 source files
• complete pytest suite: 744 passed in the protected CI matrix
• silent-exception, release-tag, landed-commitment, seam, and store-health checks
• strict MkDocs build
• sdist and wheel build
• isolated Windows wheel install: 1.8.0, redirected help, recovery help, and four wheel-bundled recipes
• PR #74 CI: lint/type, CodeQL, package build, LongMemEval R@5, and Linux, macOS, and Windows tests on Python 3.11, 3.12, and 3.13
• settlement PR #78 CI run 33343943524 and CodeQL run 33343943482: lint/type, package build, LongMemEval R@5, CodeQL, and all nine OS/Python test jobs passed on integration head f53bdea
• release-record PR #79 passed all 14 required checks before merge as 8dcb3fb
• trusted PyPI publication and independent post-publish tag verification
• GitHub Release with wheel and source distribution
• workflow, GitHub Release, and PyPI artifact SHA-256 digests match
• both artifacts pass GitHub SLSA provenance verification
• fresh public-PyPI Windows/Python 3.11 install: version, redirected help, four wheel-bundled recipes, and dependency check

Publication evidence

The immutable annotated tag v1.8.0 points to the protected release-record merge 8dcb3fb676242af49236fac749ac3a57456cac0b.

The first tag-triggered run, 33345026660, exposed a circular gate: pre-publication tests required 1.8.0 to exist on PyPI before the workflow was allowed to publish it. The failure occurred before the build, PyPI, or GitHub Release jobs; the tag was not moved, deleted, or reused. PR #80 added an exact-tag recovery path, kept all product tests, moved the self-referential tag/PyPI check after trusted publishing, and required GitHub Release creation to wait for that independent check.

Recovery run 33346147256 checked out v1.8.0, verified that its version and commit exactly matched the tag, passed the release gates, built and attested the artifacts, published through PyPI Trusted Publishing, verified the public PyPI files, and created the GitHub Release. PyPI accepted the files before its JSON index exposed them; the post-publish gate passed once that propagation completed.

Artifact digests are identical across the workflow artifact, GitHub Release, and PyPI:

Artifact	SHA-256
cairntir-1.8.0-py3-none-any.whl	74a42bfd5f0db8cb68bf5c3d5a1387c4d73eba19cbc7bd80f09e2da65d44b59e
cairntir-1.8.0.tar.gz	8e3be37c9cb05186314d81db014532dc1a3704747c19d3b8f5018015af5146ee

gh attestation verify --repo pnmcguire480/cairntir verified both artifacts. Because this was an immutable-tag recovery dispatched from main, the SLSA certificate identifies workflow commit 3a7af17d6ffd63cb2930dc56af514a4678d9170d and refs/heads/main; the workflow’s separate exact-tag check is the evidence that the built checkout was v1.8.0 at 8dcb3fb. This provenance distinction is recorded rather than overstated.

A fresh Windows Python 3.11 environment installed cairntir==1.8.0 using only public PyPI. cairntir version returned 1.8.0, --help rendered cleanly, all four bundled recipes loaded from site-packages, and the 64-package environment passed uv pip check.

Migration

None. The database schema and embedding space are unchanged. Transcript reads are opt-in and local. No reindex or transcript import is performed.

Publication authorization

Patrick explicitly requested that 1.8.0 be released on 2026-08-30. The tag points to the exact protected-branch release-record merge. Trusted publishing, provenance verification, artifact digest comparison, and a fresh public-PyPI Windows install all passed.

This site is open source. Improve this page.