Catch up - sweep the inbox into the brain

Runs on: local-writes - files things into brain/rants/; on a read-only surface it drafts what it would file and says so.

The operator captured thoughts somewhere else - a voice-note app, an email to self, saved messages, a text file dictated into a synced folder. This skill gets them into the brain with provenance, without asking the operator to structure anything.

Sources, in sweep order

1. capture/inbox/ - any .txt or .md file in this folder is treated as a capture. This is the drop zone: synced-folder dictation, exported voice notes, anything that can land as a text file.
2. Pasted text - if the operator pasted content with this request, treat the paste as one capture.
3. Connected meeting-notes tool - if a meeting_notes MCP is bound in stack.json and reachable, offer to pull transcripts since the last sweep. Never claim to have checked a tool that is not connected.

Procedure

1. Sweep. For each capture, create brain/rants/<YYYY-MM-DD>-<slug>.md with frontmatter:

   ---
   captured: <date from the file's own timestamp or content, else today>
   source: <inbox filename | paste | meeting-notes tool>
   processed: false
   ---

   Body = the raw text, corrected ONLY per the name pass below. Never rewrite, summarize, or clean up the operator's words at this stage - raw in, raw kept.

2. Name pass (transcription errors are the rule, not the exception). Dictation mangles proper nouns. Before filing:
   • Read context/names.md (the names glossary). For any name in the capture that closely matches a glossary entry or a listed mishearing, substitute the canonical spelling. Substitute ONLY on a glossary match.
   • A proper noun that matches nothing stays as heard, marked (sp?) - for example Jansi (sp?). Collect all (sp?) names and ask the operator about them in ONE batch question at the end of the sweep, not one by one.
   • The batch question names the file the corrections go to: "I will add your corrections to your names list (context/names.md) so they stick." When the operator corrects a name, apply it AND append the wrong-to-right pair to the glossary's mishearings list, so the same mistake never survives twice. If they say not to keep a correction, apply it to this capture only.
   • Never "fix" a name, number, price, or date you are unsure of. An unsure span stays as heard with the marker. A confident wrong substitution is how fabricated facts enter a brain; the marker is how they stay out.
3. Move, don't copy. A swept inbox file moves to capture/inbox/.processed/ (create if missing) so the inbox reads empty when it IS empty. Nothing is deleted.
4. Report and offer. One line per capture filed (filed brain/rants/2026-07-08-site-visit.md - 2 names confirmed, 1 marked (sp?)), then offer: "Say 'dream' (or run /dream) and I will distil these into patterns, flags, and decisions." Do not auto-run /dream.

If two sources cover the same event

When the same conversation arrives twice (say, a meeting-notes tool and the operator's own voice memo), reconcile by this hierarchy - higher wins:

1. A document or screenshot (authoritative for names, numbers, terms)
2. What both sources independently agree on
3. A single source - file it, but mark facts only it carries as (single-source)
4. The operator's own correction - overrides everything above

A proper noun the two sources disagree on is surfaced as provisional, never silently picked.

When one source is all you have

Most of the time it is. You recorded a call on one device and there is no second version of it. That is the normal case, and it has a failure mode worth naming because it does not look like a failure.

A single recording with no second source leaves nearly every proper noun unconfirmed. Not wrong, unconfirmed. A person's surname, a company name, a product name you heard once and never saw written down. The capture reads fluently and confidently, so the marked names quietly become facts the next time anything reads that file.

The fix is never a better transcript. Re-recording the same audio reproduces the same gaps, because the information was never in the audio. What resolves an unconfirmed name is a second source, a document, or you confirming it. Three things follow:

• When a capture is single-source, say so in the sweep report. "One source, so the names below are unconfirmed" costs one line and changes how the next reader treats the file.
• An unconfirmed name that reaches a client-facing document or another person's inbox is the expensive version of this. Confirm before it crosses that line, never after.
• If the same class of capture problem shows up twice - always single-source, always the same tool dropping speakers - that is a problem with how you capture, not with that day's notes. Fix the capture path.

Media you consume, not just meetings you attend

A podcast, an audiobook, a course, a long video you were half-listening to. Those land the same way as a meeting and they are worth capturing, because most of what changes how you work arrives that way rather than in a call.

They route differently, though. A meeting updates a person or a client record. Media updates your method: it belongs as a proposed insertion into whatever holds how you do the work, through knowledge-capture. Media in, proposed edits to your own playbook out, and you approve them. Never a silent rewrite of your method because something in your headphones sounded good.

The one thing to filter: capture the training, not the entertainment. A recording that ran all afternoon may hold one useful arc and three hours of unrelated life. File the arc, skip the rest, and say how many you skipped.

Rules

• Raw in, raw kept. No summarizing, no cleanup beyond the glossary name pass.
• Every capture gets processed: false so /dream can find it. Never mark processed here.
• Ask about unknown names once, in one batch. Never write an invented identity, venture, or fact to make a capture look complete.
• If capture/inbox/ does not exist, create it with its README from templates/capture/inbox/README.md and tell the operator what it is for.
• No em dashes or en dashes. Hyphens only.

Record the run (the closing act, when this runs as a seat)

If roles/employees.yaml carries the capture-filer row, close with one line so the run leaves a trace whether or not anyone was watching:

python scripts/agent_runs.py record --seat capture-filer --trigger "inbox sweep"         --read "capture/inbox/,context/names.md" --produced "brain/rants/" --outcome ok

Use --outcome refused (with --could-not "<why>") when there was nothing in the inbox to file, and --outcome failed --could-not "<why>" when it broke - the script requires the reason for both, so a failure with no reason is never a silent no-record. A refusal is not a failure and the log distinguishes them. Skip this silently if the script or the registry is absent, and never mention it in your reply - it is bookkeeping, not output.