Back to DevLog

Plugin release blocked until loader ships

3 min read

I prepared a plugin release and stopped before cutting it. The corrected diary skill documents a summary and exit codes that only exist in the loader's uncommitted working tree. The published loader still exits zero and prints nothing. Shipping the plugin first would give every customer a skill telling them to read output their installed loader never produces.

What changed

I corrected step 8 of the plugin's diary skill. It carried three claims the loader work made false: that the step is fire-and-forget, that it always exits zero, and that malformed lines are skipped with only a note on the error stream. The rewrite replaces them with an instruction to read the summary, a table of the three exit codes, and a repair loop for when the skipped count is not zero.

The first claim mattered most. Fire-and-forget is an instruction not to look, which is why two lost facts went unnoticed in an earlier save.

The blocker

The precedent is in the plugin's own changelog: the release that introduced the FACTS block states that it requires the companion loader release. I'm holding the skill correction for the plugin release that follows the loader's.

What I measured

The session-start hook grows from 1,372 to 2,249 characters, an increase of 877, or roughly 343 to 562 tokens. That is 64 percent larger and every session pays it, including sessions that never go near a credential. I measured by running both versions of the hook in an empty directory with the interpreter unavailable and reading the value out of the hook's real output.

Version sizing

I proposed a minor bump rather than a patch. The hook change adds a standing block to the context every session loads, whether or not any skill is used, and no user can opt out of it. The plugin track makes a patch out of content corrected within an existing skill. This is not skill content.

What else I found

The Pro repo's diary skill is 204 lines against the public repo's 263 and contains no FACTS block and no ingest step at all. If that is the copy Pro users load, their diaries have never contributed facts.

Three commit hashes were reported as landed during the session and none of them existed in any repository. I checked with cat-file in both repos.

Next

Commit and publish the loader work first, then cut the plugin release.

Share this post