box/docs/DIGEST-PROTOCOL.md

Digest Protocol Runbook

Box is the main surface. All operator work goes through Box (box.muse-dev.online). The web UI, box CLI, and agents share the same API endpoints. No UI-only powers.

How the main loop sends digests and how agents close them.

Status: spec (2026-10-04). Implemented in self_main_loop.py + response-harvester.py; no new machinery — the protocol wires digests into the existing dm.py --expect-reply / followups.json / harvester response path.

---

1. Why this protocol exists

Main-loop digests ([main-loop] N new in <agent> main chat: ...) were being sent with no response contract: no stable ID, no declared timeout, no nudge/escalation policy, and no way to distinguish "seen and closed" from "ignored". Across 43 digests there were zero RESULT replies and zero measurable outcomes — not because agents ignored them, but because nothing defined how to answer.

This protocol closes that wiring gap without inventing a new system:

• digest composer → dm.py send --expect-reply (followup record, timeout/nudges declared at send time)
• agent replies in-band with a verb: [ACK|CLAIM|RESULT|DECLINE|NO-ACTION <id>]
• response-harvester.py matches by job_id and records the outcome in followups.json
• the sweeper nudges/escalates only unresolved followups, per the declared policy

2. Digest classes

Not every digest deserves a reply demand (notification fatigue; and the standing rule that the human only sees real decisions and escalations). The composer already marks [?] (question) and [!] (operator-needed / urgent).

Class	Marker	Response required?	Followup record?	Measured by
ACTIONABLE	carries [?] or [!]	yes	yes	delivered → acked → closed
INFORMATIONAL	neither	no	no	delivered only

Informational digests are sent exactly as today. They are excluded from every reply-rate metric — an informational digest is measured as delivered, full stop. Judging informational FYI traffic by answer rate is structurally misleading.

3. Digest ID

Every ACTIONABLE digest gets a stable ID:

ml-<agent>-<YYYYMMDD-HHMMSS>

Example: ml-646-20261004-213000.

The ID is embedded visibly in the digest as [JOB ml-...]. This marker survives extract_trailing_tags (only known canonical keys are stripped) and is picked up by dm_send's job_id regex (\[JOB\s+([A-Za-z0-9_-]+)\]) anywhere in the message, so it lands in tags["job_id"] → the followup record. No new dm.py flags are needed for the ID.

Replies match on this ID, so attribution games with [from:X] can never break resolution.

4. Reply verbs

Printed in the contract footer of every ACTIONABLE digest (Section 7).

Verb	Syntax	Meaning	Followup effect
ACK	[ACK <id>]	Seen, noted	status acknowledged — stops nudges, not closed
CLAIM	[CLAIM <id>]	I'm handling it	status acknowledged — stops nudges, not closed
RESULT	[RESULT <id>] <result text>	Done, result attached	status resolved, outcome RESULT — closed
DECLINE	[DECLINE <id>]	Won't act (reason optional)	status resolved, outcome DECLINE — closed
NO-ACTION	[NO-ACTION <id>]	Reviewed, nothing needed	status resolved, outcome NO-ACTION — closed

Semantics:

• ACK vs CLAIM: both acknowledge. CLAIM signals ownership; ACK is just "seen". Use CLAIM when you're picking up the work so the human knows who's on it.
• NO-ACTION is a real answer, not a dodge. It means "I reviewed this and there is genuinely nothing to do." It closes the digest so it stops being counted as open.
• DECLINE should carry a reason when one exists ([DECLINE ml-646-...] blocked: chromebox flapping), but a bare DECLINE is still a valid close.
• RESULT closes only when the result text is present. An empty RESULT is a malformed reply — the harvester treats it as unrecognized.

Nudge interaction: ACK/CLAIM suppress further nudges but leave the followup open until a closing verb arrives. This matches the real workflow — "I'm on it" shouldn't trigger nags, but the loop isn't closed until there's a result.

5. Timeout policy

Declared at send time (standing rule: policy is part of the send, not decided later).

Digest class	Timeout	Nudges
normal (marked [?])	60 min	2
urgent (marked [!])	30 min	2

Implemented as --reply-timeout {3600|1800} --reply-nudges 2 on the dm.py send invocation for ACTIONABLE digests. The timeout and nudge count are stored in the followup record; the sweeper honors them.

6. Nudge / escalation policy

• The sweeper nudges only followups with status == "pending". Anything acknowledged or resolved is never re-nudged (ACKed digests getting nagged trains agents to ignore the protocol — verify this when wiring the sweeper).
• After the declared nudges are exhausted and the timeout passes, the followup is escalated, not retried forever:
  • Phase 1: the sweeper's existing escalation path (DM to the escalate target declared at send time via --reply-escalate).
  • Phase 2: route to the registered main-loop brain sidechat for triage. The brain is registered in job-sidechats.json as main-loop brain and is name-resolvable, so no hardcoded UUIDs are needed.
• stale = still pending past timeout + nudges → escalated.

Digest routing must use sidechat names (646 tasks, main-loop brain) via the existing name resolution — never hardcoded thread UUIDs (see the stale-UUID lesson in AGENTS.md).

Note: the harvester only monitors threads registered in job-sidechats.json. opm/heartbeat is not registered (digests there resolve via fuzzy search), so digest replies there would never resolve. Route opm's digests to the registered main-loop brain sidechat instead of heartbeat — cleaner separation (brain = triage workspace, heartbeat = heartbeats) and zero new registrations.

7. In-band contract footer

The digest budget is 600 chars, so the footer is compact. Actionable digests trim message previews from 5 to 3 to fit it:

[main-loop] [JOB ml-646-20261004-213000] 3 new in 646 main chat — reply needed:
- human: <preview> [?]
Reply: [ACK id] seen | [CLAIM id] mine | [RESULT id] done | [DECLINE id] | [NO-ACTION id]. Report back here. Box dispatches the next step; do not DM the next agent directly.

The footer is the primary contract. This doc is the reference.

8. Example exchange

# opm → 646 tasks (ACTIONABLE, --expect-reply --reply-timeout 3600 --reply-nudges 2)

[from:opm] [id:a6579a8d] [main-loop] [JOB ml-646-20261004-213000] 3 new in 646 main chat — reply needed:
- human: should we archive the stale pipe-7d2896 thread? [?]
- opm: propose yes, it's superseded by the brain channel [?]
Reply: [ACK id] seen | [CLAIM id] mine | [RESULT id] done | [DECLINE id] | [NO-ACTION id]. Report back here. Box dispatches the next step; do not DM the next agent directly.

# 646 picks it up

[from:646] [CLAIM ml-646-20261004-213000] reviewing now

→ followup status: acknowledged (nudges suppressed, loop still open)

# 646 finishes

[from:646] [RESULT ml-646-20261004-213000] archived pipe-7d2896 via muse-cli; brain channel is canonical

→ followup status: resolved, outcome: RESULT (loop closed)

A shorter close:

# pip sees an informational-with-a-question, reviews, nothing to do

[from:pip] [NO-ACTION ml-pip-20261004-214500] reviewed — nothing needed, timer stagger already handles it

→ followup status: resolved, outcome: NO-ACTION (loop closed)

9. Metrics

Source of truth: followups.json (outcome field, written by the harvester).

Metric	Definition
delivered	dm.py reported verified:true with placement confirmed for the digest message (see dm-log)
acked	outcome ∈ {ACK, CLAIM}
closed	outcome ∈ {RESULT, DECLINE, NO-ACTION}
stale	still pending past timeout + nudges → escalated
closure_rate	closed / actionable_delivered

Rules:

• Informational digests are excluded from every rate. They are measured as delivered, full stop.
• Actionable digests are measured as closed. A digest that is ACKed but never resolved counts as open until a closing verb arrives.
• delivered requires the placement confirmation, not just the verified:true flag (see the AGENTS.md placement-blindness lesson).

Reporting: self_main_loop.py status exposes digest_health() counts (delivered/acked/closed/stale), surfaced through the existing Box main-loop/status API. No new endpoint needed in phase 1.

10. Implementation notes

• dm.py — no changes. --expect-reply, job_id extraction from [JOB …], followup registration, and placement verification already exist and are tested.
• response-harvester.py — extend the reply regex to verb-aware matching: \[(ACK|CLAIM|RESULT|DECLINE|NO-ACTION)\s+([A-Za-z0-9_-]+)\]; on match, resolve by job_id and record outcome. ACK/CLAIM → acknowledged; RESULT/DECLINE/NO-ACTION → resolved + outcome.
• self_main_loop.py — compose_digest classifies actionable vs informational; actionable embeds [JOB ml-…] + footer; send_prompt passes --expect-reply --reply-timeout {3600|1800} --reply-nudges 2 for actionable only.
• Sweeper — verify it nudges only status == "pending" (no code change expected).

All changes follow review-then-commit. No new daemons, no new state files beyond the existing followups.json.

---

11. Recursive workflow discipline (box -> agent -> box)

Success = box -> agent -> box -> agent -> box (recursive). The failure mode is box -> agent : agent — work handed agent-to-agent in chat threads, with nothing ever returning to box. The tooling (box job result|status|next|chain) makes the recursive path the easy path; this section makes it the default:

• Reply in-thread. Post [RESULT <id>] <outcome> back in the digest thread (this thread). The harvester watches it; box job result <job-id> records it into job-log.jsonl.
• Box chains the next step. A [RESULT] on a chained job fires trigger_chain_next() (or box job next to dry-run, box job chain to wire); box dispatches the next step to the next agent. The agent never needs to know who goes next.
• Never DM the next agent directly. Handing work off in a chat thread bypasses box orchestration: no followup record, no chain state, no job status visibility. If you catch yourself composing a DM to hand off work, stop and post the [RESULT] instead — box routes onward.
• The footer says it. Every ACTIONABLE digest now closes with: "Report back here. Box dispatches the next step; do not DM the next agent directly." — the contract is in-band, not tribal knowledge.

Companion docs: DM-SPEC.md (DM format), THREAD-BOOKKEEPING.md (pin/archive), WARP-EGRESS-FIX.md (partition handling).