# Markdown product validation

This report covers the new explicit save/resume skill. The historical strict-engine matrix remains in [compatibility](compatibility.md). The beta is still a candidate.

## Final supported candidate

Reference runtime and package source: `b29423bdc07464e8b3ea0fc68d1575755f3b7e86`. [Final CI](https://github.com/caotaiyi/remvu/actions/runs/39138270476) passed all **eight supported combinations**: Linux/macOS Python 3.9, 3.10 and 3.12; Windows Python 3.10 and 3.12. Each discovered 170 tests; Linux/macOS passed 169 with one platform skip, Windows passed 166 with four platform skips. Both supported Windows jobs additionally passed ten repetitions of six distinct hook writers with zero warnings (twenty trials total). The local macOS Python 3.9.6 run passed 169 with one skip, and the fresh archive lifecycle smoke passed.

All four ZIPs and SHA256SUMS were compared byte for byte with the final CI artifacts and matched. The Markdown-only archive contains ten files, zero Python files, and is 18,086 bytes. These checks do not certify native skill discovery or GUI behavior. Windows Python 3.9 is explicitly unsupported, with its failed stress runs retained below and a strict manual reproduction workflow. It is **not** counted as a passing combination.

The final README also states that save must happen while the original agent can respond. A missing note cannot reconstruct an unsaved conversation after quota exhaustion. The default skill is a manual handoff workflow, not automatic recovery of all project activity.

Documentation-only updates after the reference commit do not change the tested runtime or candidate archives. The beta remains an unpublished draft while the native client matrix is incomplete.

## Local mechanics (initial run)

The local macOS Python 3.9 test run passed: 167 discovered, 166 passed and one Windows-only skip. Two new tests install the skill into a fresh directory with no Python binary in the installer's PATH and reject existing/symlinked skill targets. This is an installer-mechanics check, not native slash-command or GUI certification. Python is used to run developer tests only.

The main skill uses host file tools and installs no hooks. The optional engine Stop callback now returns without a blocking decision, while pending evidence and observation-gap protections remain. Existing preservation, recovery, privacy and concurrency tests passed. The first new [CI run](https://github.com/caotaiyi/remvu/actions/runs/39135848723) passed the six Linux/macOS jobs but failed Windows tests that decoded UTF-8 skill text with the system default cp1252 codec. The affected tests now explicitly read and write UTF-8; this preserves their original assertions. A subsequent [run](https://github.com/caotaiyi/remvu/actions/runs/39136493048) completed its Windows unit-test stage (167 discovered, with four platform skips), but the Python 3.9 repeated six-writer pressure check failed on trial 2: the five-second ledger budget expired. Its Python 3.10/3.12 pressure checks passed. The normal recorder wait is now eight seconds within its ten-second host configuration; shutdown remains one second. A held-writer regression now waits 5.5 seconds, beyond the old budget, without weakening the six-event/no-warning assertions. This is an optional recorder check, not a Markdown dependency or native-client certification. The next [run](https://github.com/caotaiyi/remvu/actions/runs/39136985254) exposed a separate intermittent Windows `os.replace` WinError 5 while replacing state.json. The exact external holder was not identified. Atomic writers now retry only Windows errors 5/32/33 for at most 0.5 seconds, keeping the same temporary file and never unlinking the old destination. Persistent denial still fails. Added tests verify transient recovery for all three writers, bounded failure, old-file preservation and temporary-file cleanup. The [60fc871 run](https://github.com/caotaiyi/remvu/actions/runs/39137595616) passed eight combinations but Windows Python 3.9 still exceeded the eight-second budget in repeated pressure. We therefore narrowed the supported managed-engine range to macOS/Linux Python 3.9+ and Windows Python 3.10+. Windows 3.9 is explicitly rejected by managed installation before state/config writes, with a no-Python Markdown alternative. Its failed combinations are not counted as passing. The strict six-event/no-warning test remains unchanged and can be reproduced with the [manual experimental workflow](../.github/workflows/experimental-windows39.yml); no test is marked xfail or ignored in supported configurations. Final CI is recorded below when complete.

## Real native pilot

[Two-task pilot measurements](../evals/results/markdown-pilot-2026-10-18.json) preserve one setup failure: Claude Code refused a .git/info/exclude edit, and the producer did not save private content without protection. In the other task, both ordinary HANDOFF.md and Remvu receivers completed all five behavior cases on their first attempt. The Remvu file was 4,212 bytes versus 814 bytes for the ordinary handoff; receiver times were 9.574 and 8.974 seconds respectively. This small trial shows no advantage in speed or size.

The workflow now uses a regular project .gitignore rule before writing private notes, and the note template is shorter. Those changes require a new run. The pilot has not been discarded or relabeled as passed.

## Ten-task comparison

[Protocol and stopping rules](../evals/markdown-protocol.md). The native ten-task run completed in both directions (five tasks per direction). Receivers had only the synthetic code and one handoff file; no Remvu runtime was installed. [Measurements and actual usage](../evals/results/markdown-ten-2026-10-18.json) · [All twenty final code files](../evals/results/markdown-ten-final-code-2026-10-18.json).

| Measure | Ordinary HANDOFF.md | Remvu Markdown |
| --- | ---: | ---: |
| First-attempt tasks passing all fixed cases | 9/10 | 9/10 |
| Tasks passing after at most one correction | 10/10 | 10/10 |
| Evaluator-requested correction rounds | 1 | 1 |
| Sum of receiver elapsed seconds, including corrections | 203.805 | 184.128 |
| Sum of handoff bytes | 10,039 | 34,614 |
| Rejected shortcuts seen in final code review | 0 | 0 |

Both arms failed the same internal-BOM case initially and passed after actual failing-case feedback. That fixture has an underspecified boundary: its accepted starting trim() removes internal U+FEFF, but the fixed rubric preserves it. This correction is not evidence of lost project memory. The frozen fixture and raw result remain visible rather than being silently changed after the run. Final-source review does not measure discarded or attempted approaches. Native usage counters are kept per call; overlapping provider counter views are not summed. Producer usage covers both handoffs jointly and cannot establish the cost of one save.

This run found equal completion and correction outcomes, with larger Remvu files. The measured elapsed totals are one synthetic run with varying native model defaults, caching and two workers; they are not evidence of a speed or token-saving advantage. Task times overlap and these sums are not experiment wall time. Do not promote the timing difference as a savings claim. For a simple task where HANDOFF.md already preserves the required context, this experiment shows no completion benefit from adopting Remvu.

The main product improvement is reduced installation friction and an explicit handoff convention. Its value for long-running projects, source fidelity and repeated sessions still needs suitable tests. Full provider records and private notes are not published.

## Not yet established

Native Windows PowerShell installation, terminal skill discovery, desktop and VS Code entry points, long-term multi-session use, simultaneous Markdown writers and enforced content fingerprints are not certified. Markdown editing cannot promise the strict engine's transactional guarantees. Hidden reasoning is excluded. No cost or token reduction is established.