Check plugin data writability through native observations

[jonathanhefner]

Add mcp.stdio.data.writable to the core suite. The existing absolute-path check does not establish that the plugin subprocess can write to its data directory.

Each call to the default server’s observe tool exclusively creates a temporary file in PLUGIN_DATA, writes and closes it, and removes the file it created. The observation preserves filesystem errors for deterministic evaluation. Repeating the check on each call supplies current evidence when a client retains the MCP process across runs.

Cleanup failures appear as warnings in JSON and human summaries without changing a successful writability result. The tool description and annotations disclose the write, and the existing guide records the expanded observation through the same workflow.

Validation:

• All 53 local tests pass, including actual writes, failed writes, cleanup, preservation of existing files, repeated observations, and saved warning rendering.
• Syntax and committed bundle reproducibility checks pass.
• The abbreviated README example parses as JSON, its counts match the displayed results, and all four documented jq queries produce the expected outcomes.
• Independent implementation and documentation review completed.
• CI passes on Linux and Windows with Node 22 and 24.

Native validation used unchanged copies of commit 577aa38, the ordinary guide request, and each client’s native plugin loading mechanism:

Client	Saved report	New writability check
Codex CLI 0.154.0-alpha.6.2	15 pass, 0 fail, 0 not_verified	Pass
GitHub Copilot CLI 1.0.83	13 pass, 2 fail, 0 not_verified	Pass

Both runs saved all seven observations unchanged, matched deterministic evaluation, and removed the temporary probe file. Independent review checked the traces, recorded evidence, and package integrity. Programmatic acceptance of mcp.stdio.data.writable passed for both reports.

Copilot retained its previously observed argument and environment expansion failures: $APC_VALUE expanded in an argument, while ${PLUGIN_DATA} remained literal in a configured environment value. It also batched the four observation calls before recording them, contrary to the guide’s instruction to record each observation immediately. Every recording command completed sequentially, and all returned evidence matched the saved report.