Research · skill report
astro-brain, described the same way as everything else here.
Codger got a benchmark because there were comparable projects to measure it against. This one gets a description instead: what it does, how it reaches your vault, and where it is judgement rather than measurement. Same skeleton, one honest substitution.
← Back to the skill pageGoal
The context dies with the session. That's the problem.
An agent finishes a task holding everything that mattered: which approach was chosen and why the other one was rejected, which bug had which root cause, what is still open. Then the session ends and all of it is gone. Next week the same ground gets re-explored from zero, and the notes that would have prevented it were never written, because writing them is a chore nobody does at the end of a long day. The question is whether the agent that already has the context can put it down itself, in a format a human can read and edit, and keep the result coherent without supervision.
Method · pass one
Write the session, then the project.
A dated session note gains one section per task: what changed, what was decided and why, the root cause of anything that broke, and the open threads. It appends, never overwrites, so a day with six tasks is six sections and not the last one. The project note it touched is updated in place instead, so it stays a current picture rather than a log. Both are linked from the vault's index in the same pass that creates them, which is what makes “no orphans” a property of the write rather than something the cleanup has to fix afterwards.
Notes carry no frontmatter. Astro adopts an external Markdown file, mints it a stable id and takes the title from the filename, so a hand-written header risks colliding with an id that already exists. The vault watcher picks the file up live: no restart, no import step, no interaction with the app at all.
Method · pass two
Then clean up after everyone, not just after itself.
The coherence pass runs over the whole vault every time, regardless of what this session touched. Broken wiki-links are fixed by correcting the spelling or writing the missing note as a real stub, never by deleting the link. Orphans are linked from the index rather than deleted. Near-duplicate titles are merged into whichever has more inbound links, with every link that named the deleted title repointed. It reports counts, and “vault coherent” is a normal and common answer.
Constraints
What it refuses to do is most of the design.
It never guesses a vault path and never creates a vault: if Astro's config is missing or names no open vault, it does nothing and says so. If the open vault is plainly about a different subject than the work, it writes nothing and names the vault instead of cross-posting. It stays silent on turns that changed no files, on typos and version bumps, and when a run today already recorded the same task. And no key, token, .env content or gitignored file ever enters a note: it references the file that holds the secret.
Limits
Where this is judgement, not measurement.
- There is no benchmark. A session note is either useful in three months or it isn't, and that can't be scored in an afternoon. This page describes behaviour; it does not claim a measured result.
- Duplicate detection compares titles, not meaning. Session 2026-07-28 against session 2026-07-28 is caught; two notes about the same decision under genuinely different names are not.
- The vault-subject check is a judgement call. It compares the existing notes against the work and stops when they plainly disagree, but a vault covering several subjects at once gives it nothing to disagree with.
- Deliberate ghost links are protected by reading the sentence around them, which is prose comprehension rather than a rule. The seed vault ships one such link precisely to keep that behaviour honest.
- Writing is append-first, but a same-day re-run has to recognise its own earlier entry to update it instead of duplicating it. That match is by task description.
Code
The skill is the specification.
There is no binary and no runtime to inspect. The skill is one Markdown file, and everything on this page is in it, in the order it executes.
Next
What would turn this into a result.
The honest test isn't whether the notes look good the day they are written, it's whether they answer a question months later that would otherwise have cost an afternoon. That is the memory-and-retrieval track: retrieval quality measured against a vault that has actually grown, over months rather than minutes. When there are numbers they go on the research page, losses included.
The research tracks →