Forgotten work

Two things make email feel like dropped balls: promises you typed into a draft and forgot about, and inbound threads that quietly aged past the speed at which you usually reply. mxr surfaces both, using only local data — extracted commitments from sent mail, and an “owed reply” lens that ranks threads against the recipient’s typical cadence.

The one-line mental model

Commitments come out of your sent mail; the owed-reply lens looks at threads where you owe in. Both are deterministic ledgers — no LLM is required for the owed lens; the commitments extractor uses an LLM only to confirm a deterministic prefilter match.

Commitments — promises you made

When mxr send DRAFT_ID --check (or any real send) runs the safety pipeline, it scans the draft body for explicit outgoing commitments: “I’ll send the deck Friday”, “I’ll get back to you next week”, “I can review by EOD Monday”. Each match becomes a candidate scoped to the draft. On successful send, candidates promote into the canonical contact_commitments ledger and show up in mxr commitments.

# Before sending: see the candidates the pipeline extracted.
mxr send DRAFT_ID --check --format json \
  | jq '.issues[] | select(.code == "commitment_candidate")'

# After a real send: candidate is now in the ledger.
mxr commitments --status open --format json

What you get: a JSON array of rows like { id, contact_email, direction, what, by_when, evidence_msg_id, status, created_at }. direction is yours (you owe work) or theirs (they owe you). Every row has a message id you can mxr cat to read the original.

List, filter, and resolve

# List everything still open across all contacts.
mxr commitments --status open --format json

# Just one person:
mxr commitments --contact alice@example.com --format json

# Mark done after you've shipped it.
mxr commitments resolve COMMITMENT_ID

# Daily standup: what did I commit to this week that's still open?
mxr commitments --status open --format json \
  | jq -r '.[] | select(.created_at > (now - 7*24*3600))
           | "\(.by_when // "unscheduled")\t\(.contact_email)\t\(.what)"' \
  | sort

What you get: one row per open commitment from the last 7 days, tab-separated, sorted by due date — pipe into column -t or paste straight into a standup doc.

When the LLM is off

The extractor’s prefilter is pure regex on first-person commitment markers (I'll, I will, I can, I'll send, …) and due-phrases (weekday, date, tomorrow, next week, by EOD). When the LLM is disabled or unreachable, deterministic prefilter matches still surface as low-confidence candidates in the safety report — they just don’t get promoted into the ledger automatically. Run with --check to see them and decide.

mxr send DRAFT_ID --check --no-llm --format json \
  | jq '.issues[] | select(.code == "commitment_candidate")'

Owed-reply lens — threads where you’re the bottleneck

mxr owed ranks threads where the latest message is inbound and you haven’t replied. It scores each thread by waiting_days / expected_days, where expected_days is the recipient’s typical cadence from reply_pairs (falling back to a global p50, then to 7 days). The same set powers the is:owed-reply search operator.

# Top 20 threads you owe, ranked overdue-first.
mxr owed --format json | jq -r '.[0:20]
  | sort_by(-.overdue_score)
  | .[]
  | "\(.overdue_score | tostring | .[0:4])\t\(.counterparty_email)\t\(.subject)"'

What you get: tab-separated rows score \t sender \t subject. Score 1.0 = exactly at typical cadence; 3.0 = three times longer than usual.

# Persistent sidebar lens (TUI and web).
mxr saved add owed 'is:owed-reply'

# Same set, scriptable.
mxr search 'is:owed-reply' --format ids

Two equivalent forms

mxr owed --format json returns the structured row (with overdue_score, waiting_days, cadence_days_p50). mxr search 'is:owed-reply' returns the underlying message envelopes through the search stack. Use owed when you want the ranking metadata; use search when you want to compose with other operators (is:owed-reply from:acme.com).

Exclude noise

The lens already excludes list senders, screener-denied senders, trash, and spam — so newsletters never pollute it. To narrow further:

# Only threads waiting >= 14 days.
mxr owed --since 14 --format json

# Only threads whose latest inbound landed in the last 60 days
# (skip ancient unanswered relics).
mxr owed --within 60 --format json

# Combine: aged in, but not too old.
mxr owed --since 7 --within 60 --format json

In real life

• Friday standup prep: mxr commitments --status open --format json | jq '.[] | select(.direction == "yours") | {who: .contact_email, what, due: .by_when}' — list every open promise across every contact, grouped by who’s waiting.
• Inbox-zero focus session: mxr owed --since 3 --format ids | head -10 | xargs -I{} mxr thread {} --format json — pull the top 10 overdue threads, read them in batch, reply down the list.
• Quarter-end review: mxr commitments --status open --format json | jq '[.[] | select(.direction == "theirs")] | length' — count how many things people owe you that have been sitting open.
• Pre-send sanity: before queuing 8 replies in a row, check mxr commitments --contact alice@example.com so you don’t promise the same deck twice.

Agent prompts that work

"For every open commitment in `mxr commitments --status open --format
json` where `direction == "yours"` and `by_when` is in the past, draft a
short status-update reply with `mxr draft-assist <evidence_msg_id>`.
Don't send — show me the drafts."

"List the top 10 owed-reply threads from `mxr owed --since 5 --format
json`. For each, summarize the latest inbound with `mxr summarize
<thread_id>`. End with a single recommended verb per thread: REPLY,
ARCHIVE, DELEGATE."

See also

• Pre-send safety — where the commitment candidates are extracted
• Automated follow-ups — reminders and send-later for the outbound side
• Search workflow — the is:owed-reply operator
• LLM features — configure the model used for commitment extraction
• CLI — mxr commitments, mxr owed, mxr send