Inbox Info

Inbox Information Research

Overview

The user reads email with NeoMutt against a Maildir indexed by mu (maildir-utils). This skill wraps mu find / mu view to answer questions about the inbox without scanning Maildir files directly. It also helps convert flagged emails into themed nytid todo items.

Canonical "needs handling" signal: the user flags every email they need to act on (the \Flagged / "starred" IMAP flag). Default to flag:flagged for triage queries — not flag:unread.

This skill is strictly read-only: query mu, suggest nytid todo commands, and let NeoMutt own flag/state changes.

Known Setup

Embedded so the skill activates with zero tool calls. Re-verify with mu info only if a query misbehaves.

• Maildir+ account root (use with mu — it indexes the whole tree): /home/dbosk/mail/kth
• Key leaf folders — encode the lifecycle of a message, not just routing:
  • INBOX/ — waiting / not yet processed. Unread mail lives here by definition; flagged mail here is active work. For triage, this is the folder to target.
  • Processed/ — done. Messages the user has handled and moved out of INBOX. The user intentionally keeps flags here — a flag in Processed means "I had to act on this", retained as a labeled signal for tuning ~/.muttrc.scores and spotting patterns. Do not treat them as actionable triage items, but do not ignore them as noise either — they're training data.
  • Processed-archived/ — archived, not interesting. Older Processed mail. Even less relevant than Processed/ for triage.
• Xapian DB: /home/dbosk/.cache/mu/xapian
• Verified: 2026-05-27

The account root contains subfolders (INBOX/, Sent/, Archive/, Processed/, Processed-archived/, etc.) but no cur/new/tmp of its own. Pointing neomutt -f at the root opens the folder browser, not messages — limits then filter an empty set. Always target a specific leaf for mutt.

Implication for triage queries: unread mail = "waiting to be processed" (the user moves mail to Processed/ when done; unread mail still in INBOX is therefore by-definition actionable). Old unread mail in INBOX is stale — never got around to it. Flagged mail in Processed/ is not a triage target — it's done — but it is a historical record of "things that needed action", deliberately retained for score-tuning. Use the user's flag-discipline to your advantage:

• Flag-cleared anywhere = "I didn't have to act on this" (FYI/noise).
• Flagged in INBOX = "this needs action, still pending".
• Flagged in Processed/ = "this needed action and is now done" (labeled training point).

Lazy fallback — call mu info only if:

• a query returns unexpectedly zero results,
• the user reports the index seems stale,
• or mu find errors out.

Research Workflow

Step 1: Pick the right query

Bound output: always use --maxnum N and --fields to keep results compact.

Step 2: Classify against the user's taxonomy

The authoritative classifier lives in ~/.muttrc.scores. For the full extracted taxonomy — every category regex, every high-signal sender, every noise filter — see references/muttrc-scores-taxonomy.md. Read it when classification is non-trivial or when the user updates their scoring config.

The common categories (priority order) are:

Course codes to recognize:

• DD1301 (datorintro/dataintro)
• DD131[057] (programmering/prg)
• DD1320 (tilda)
• DD1337 (inda)
• DD2520 (tilkry/tillkry)
• DA2215, DD2302 (vetcyb, science/security)
• DA150X (dkex/dkand/kex, bachelor thesis)
• DA231X (exjobb, master thesis)
• LH219V (pedagogy)

Programme tags to recognize: CINEK|CMAST|CITEH|CDATE|CFATE|CTMAT|CTFYS|CLGYM.

High-signal senders (from ~/.muttrc.scores): Celina Soori, Yara Faris, Yasmina Altayy, Liisa Wettemark, Edita Korlat, Francisco Gomes, Bedour Alshaigy, plus institutional addresses (schema@admin.kth.se, service@eecs.kth.se, education-support@eecs.kth.se, etc.). When a flagged email is from one of these, give it its own cluster.

Known noise (excluded from triage even if flagged): Google Scholar, KTH Magazine, IEEE, ACM, SIAM, Nature Briefing, The Download, KTH Alumni, etc. — see negative scores in ~/.muttrc.scores.

Step 3: Handle ambiguous subjects (students often write bad ones)

Many emails arrive with vague subjects: "Hi", "Hej", "Question", "Help", "Course", a bare first name, or just a course code. Don't trust subject regex alone. Two-tier approach:

1. Subject regex first (cheap) — sort flagged emails into the obvious buckets.
2. Body query for ambiguous matches: mu find 'flag:flagged AND body:/ladok|betyg/' --fields 'd f s' catches grade-adjacent emails whose subjects don't mention grades.
3. Read the full message with mu view <path> only for the borderline cases — never as a default step.
4. If still unclassified, put it in a residual "uncategorized flagged" group rather than forcing a wrong bucket.

Step 4: Group flagged emails into themed todos

When the user asks for todos from flagged email, propose themed batches — never one todo per message. Cluster by category, course code, sender, or topic, then suggest one nytid todo add per group.

For the full rules with rationale and worked examples, see references/todo-conversion-rules.md. Headlines:

• NEVER use --who dan-claude. Email-derived todos are for the user themselves. dan-claude doesn't answer the user's email. Omit --who entirely so the default (current user) applies.
• Missing grades is always its own group, separate from other course traffic, even when only 1–2 messages match.
• Every todo description must embed a review command so the user can re-open the relevant emails with one paste.

Step 5: Choose the right review-command form

Two forms. Default to Form A unless you specifically need mu's richer query language for read-only browsing.

Form A — mutt limit on the real Maildir (DEFAULT). State changes (mark read, reply, unflag) persist to the real inbox:

neomutt -f /home/dbosk/mail/kth/INBOX \
  -e 'push "<limit>~F ~s \"betyg\\|ladok\\|grade\"<enter>"'

Note INBOX, not the account root — see "Known Setup" above.

Mutt patterns: ~F flagged, ~s "regex" subject, ~b "regex" body, ~f "regex" from, !~s "regex" negation. Combine with space (AND) or | (OR within a single regex). Date filters: ~d "<3m" newer than 3 months ago (within last 3 months), ~d ">3m" older than 3 months ago. Offsets: d day, w week, m month, y year. Absolute range: ~d "01/27/26-02/27/26" (US MM/DD/YY). When embedding in a shell string, escape inner double quotes: ~d \"<3m\".

Form B — mu virtual maildir, READ-ONLY browsing only. Use when a query can't be expressed as mutt patterns (cross-folder, mu's richer operators):

mu find 'flag:flagged AND (subject:/grade|betyg|ladok/ OR body:/ladok|betyg/)' \
  --format links --linksdir ~/.cache/mu-results/missing-grades --clearlinks \
  && neomutt -f ~/.cache/mu-results/missing-grades

Form B caveat — symlinks don't propagate state. Marking a message as read in the linksdir renames the symlink, not the underlying file. State changes do NOT reach the real inbox. Pick Form A whenever the user will mutate state.

Use a per-group slug under ~/.cache/mu-results/<slug>/ for Form B so concurrent uses don't collide.

Step 6: Suggest the todo (don't run it)

Print the command for the user to run. Prefer --command (-c) over --description for the review command — that way nytid todo start <id> runs it directly, no copy-paste. Put context/rationale in --description for view, put the executable command in -c.

nytid todo add MAIL \
  -t "Resolve missing grades — 5 student emails" \
  --description "5 flagged emails about missing grades (Ladok/betyg)." \
  -c "neomutt -f /home/dbosk/mail/kth/INBOX -e 'push \"<limit>~F ~s \\\"betyg|ladok|grade\\\"<enter>\"'"

Escape ladder (bash → nytid storage → shell on start → mutt -e parser → mutt push string). Inside a bash double-quoted argument, use \\\" for every literal " that mutt's push-string parser must see — \\ becomes \ and \" becomes ", yielding the stored \" that mutt unescapes back to ". When you splice variables into the template, make sure they're *bare* (no quotes baked in) and let the template add the \\\" around them.

For adding email context to an existing todo, suggest: nytid todo note <ID> -m "Saw a follow-up email from <sender> on <date>".

See the nytid-todo skill for broader todo-workflow details, including auto-parenting behavior (a new nytid todo add without --top-level parents under the currently active todo — handy when the user has an "Clear inbox" todo in-progress and wants each bucket as a child).

Quick Reference

--fields letters: d date, f from, t to, s subject, l path, i message-id, g flags.

Output Handling

• Default fields: d f s. Add l only when feeding mu view.
• For large result sets, summarize aggregates ("12 flagged from this month, mostly course ops") instead of dumping the full list.
• Use --threads for topic/thread questions.
• Output is tab-delimited by default; pipe through column -t -s$'\t' for readability when showing to the user.

Scope

Strictly read-only. Never run mu index, mu remove, mu mkdir, or anything that writes to the Maildir or Xapian DB. Never compose, send, or modify mail state — defer to NeoMutt. Never create nytid todo items directly; only suggest commands. Don't print email bodies unless the user asks.

Common Pitfalls

Stale index. Results miss new mail until mu index runs. If the user reports missing emails, run mu info to check the last index time and suggest they reindex.

Shell quoting. Always single-quote mu queries: 'flag:flagged', 'from:foo AND date:1w..'.

mu regex /.../ is fragile. Spaces and parentheses inside the slashes silently match zero results — e.g. subject:/applied crypto/ and subject:/(foo|bar)/ both return nothing instead of erroring. For multi-word terms or grouped alternations, use boolean OR of word terms: (subject:applied OR subject:crypto), (subject:foo OR subject:bar). Single-word alternations like subject:/foo|bar|baz/ do work. When a count looks suspiciously low, sanity-check by running each clause separately with --fields i | wc -l.

Folder lifecycle is asymmetric — flag in Processed/ ≠ actionable. Tempting reflex: query flag:flagged everywhere and assume each hit is a triage target. Wrong here. Flags in Processed/ and Processed-archived/ are intentional — the user keeps them as a labeled "had-to-act" signal for tuning ~/.muttrc.scores. They are not noise to ignore, but they are not work-to-do either. The only triage-relevant folder is INBOX/. Always constrain mu queries with maildir:/INBOX when building todos for "what needs handling":

mu find 'flag:unread AND maildir:/INBOX AND date:3m.. AND <category>'

If a folder-distribution check is ever needed (e.g. to confirm a category really is empty in INBOX vs. just sitting in Processed):

mu find '<query>' --fields l --maxnum 500 \
  | sed 's|^/home/dbosk/mail/kth/||; s|/cur/.*||; s|/new/.*||' \
  | sort | uniq -c | sort -rn

A bucket with zero unread/flagged in INBOX is correctly empty — the user has already processed it (the data is in Processed/, which is done). Don't try to "rescue" Processed mail into a triage view.

Filter choice cheat-sheet (all assume maildir:/INBOX):

Score-tuning queries (use Processed flagged mail)

If the user asks "which senders/subjects keep needing action?" or wants to improve ~/.muttrc.scores, that's exactly what flagged mail in Processed/ is for — it's the user's hand-labeled corpus of "things that turned out to need action". Useful queries:

# Top senders of acted-on mail in the last year
mu find 'flag:flagged AND maildir:/Processed AND date:1y..' --fields f \
  | sort | uniq -c | sort -rn | head -20

# Common subject tokens in acted-on mail
mu find 'flag:flagged AND maildir:/Processed AND date:6m..' --fields s \
  | tr ' ' '\n' | sort | uniq -c | sort -rn | head -30

Compare against ~/.muttrc.scores to find under-scored senders/terms.

Form B is read-only. Symlink renames stay in the linksdir; the real inbox is untouched. Use Form A whenever state must persist.

Trusting subjects. Student emails frequently have unhelpful subjects. Use body queries or mu view for ambiguous cases.

Wrong todo ownership. Never assign email-derived todos to dan-claude. They belong to the user.