Status: COMPLETE — published, installed and verified on 2026-09-07. Issue #94 tracks the combined version change; issue #96 tracks publication and installation evidence.
This additive release combines acknowledged task checkpoints and cross-host resumption from PR #93, the Claude tool-schema fix, compaction recovery guidance, and typed rejection of invalid checkpoint input. It retains schema 7, the 21 MCP tools, public exports, and existing dependencies.
Recovery instructions retain the saved wing and task ID across host changes, save exact corrections with the same task, and handle terminal or unavailable tasks explicitly. One capture/update/finish flow replaces overlapping rules. Malformed Unicode and evidence IDs outside SQLite’s integer range fail before writing; normal Unicode and idempotent retries retain their behavior.
The candidate receipt records artifact hashes and local gates at the candidate checkpoint. Its pending publication status remains historical; the completed release operations below supersede it without rewriting frozen acceptance evidence.
The private package harness initially compared an explicit 8,192-character MCP budget with the CLI’s 12,000 default. The same assertions passed with matching inputs; no runtime or frozen test changed. Before its freeze, the boundary tester excluded seven surrogate wire cases rejected by the MCP SDK before tool dispatch. Direct/CLI tests cover that input boundary; both baseline runs remain available in the frozen evidence.
The original delivery evidence remains historical evidence for PR #93. Candidate verification records new source hashes without rewriting those results.
PR #95 passed all 14 checks,
including Linux, macOS and Windows on Python 3.11–3.13. Reviewed head
ba034c7558578f9cdc8c8f1189664490f5694f7b and merged commit
d88beabdf5407b03c72cf60d9d59ca0908630999 have identical trees. Following explicit
maintainer approval, annotated tag v1.11.0 was published on that merged commit.
All five jobs in Release workflow 34102294324 passed: verification, build and provenance, trusted PyPI publication, PyPI verification, and GitHub Release. The full tagged test run passed 1,116 cases, including slow acceptance and model evaluations, at 84.75% coverage. Linux skipped the Windows encoding case and deselected the self-referential publication-tag case; the subsequent publication verification job passed that tag gate.
Version 1.11.0 is available on PyPI and GitHub.
Independent verification matched both distributions across the workflow artifact, GitHub Release and PyPI. Cryptographic verification pinned the release workflow, tag, source commit and run attempt 1.
| Distribution | SHA-256 |
|---|---|
cairntir-1.11.0-py3-none-any.whl |
c0a3dd536bae60081ee7eff97590fbfd63bf76c31a9ccefa7a860c5a791ccaf7 |
cairntir-1.11.0.tar.gz |
1f25ab1153598694a8d6aaed497a4146e5f539728dc9844e3d7bec1d82af2408 |
All 78 wheel members have identical bytes to the accepted local candidate. The archive hash differs only because ZIP creator-platform metadata identifies the Windows candidate and Linux release builders.
The publication receipt records signed provenance policy, artifact hashes and sanitized independent proof. The fresh public-PyPI installation passed dependency checks and matched all 73 package files to the public wheel. Actual separate stdio hosts resumed exact Unicode requests; checkpoint updates, typed rejection, CLI/MCP parity while the server held the store, all 21 object-root schemas, help, recipes and doctor passed.
The authorized package-only upgrade changed the actual global Python installation from 1.10.0 to 1.11.0. An independent audit matched all 73 installed files and static package metadata to the public wheel; required dependencies are unchanged. Windows initially locked the running MCP executable. Only Cairntir server processes were stopped, and the same upgrade then succeeded.
Independent online backups before and after installation are byte-identical: all 22 tables, 1,393 drawers and 1,393 vectors, including every row and schema object, remain unchanged. Schema stays at 7; no migration or reindex occurred. A restricted-permission backup attempt failed before copying; the identical read-only operation passed with filesystem access, with the failed artifact retained. Subsequent release-checkpoint memories are intentional new writes.
The real installed CLI and fresh MCP launcher report 1.11.0. All 21 tool schemas have object roots and advertise checkpoint/resume. Read-only CLI handoff while MCP holds the production store agrees at 8,152/8,192 characters; doctor passes. Actual production Claude Code health reports Connected without a tool-fetch error. No model prompts were submitted. Existing applications remain open and their MCP sessions need reconnecting; fresh probes do not establish session reload.
The user-visible outcome is one published and installed version with concise recovery instructions and reliable task checkpoint boundaries. Independent tests freeze before fixes. Budget: 90 minutes implementation plus 30 minutes reserved for verification; at most two repair rounds. Terminal dispositions are COMPLETE, BLOCKED or EXHAUSTED. Passing tests do not authorize publication.
No schema migration, new dependency, speculative history optimization, automatic conversation capture or billed-model experiment is part of this change. Tests establish checkpoint persistence and policy distribution, not guaranteed model compliance after compaction or measured token savings.