[backnotprop]

Why

After a context compaction, an agent kept a thread id from its own summary and sent seven messages into the wrong Inbox thread (the misfiled-thread AAR, 2026-10-27). The Inbox filed each one where it was told. Separately, the owner asked that every rule an agent needs reaches it with no setup, and that installed notes update by themselves. Then, from a meta-analysis: every Inbox text decided chat-or-Inbox by whether the person "is here", which an agent cannot observe. An agent raised an Inbox decision again as a blocking question in the chat once the person typed there.

What changed

• The session's open threads, from the Inbox. After a compaction (Claude Code, Pi) and on a resume (Claude Code), the session gets its open threads (inboxSessionThreadsText): one header line, at most 5 rows, then where the rest is (More: read_thread with no id …; search …). This comes from a read-only peek (_meta inbox-peek; health reports peek: true). There is no list from an older Inbox, and no error after 1.5 s.
• reply_to checks.
  • INBOX_REPLY_TO_CHECK is in the list header.
  • The reply_to field and the wake note say "never an id from memory".
  • When reply_to/thread names a thread with person messages the session has not read, send_message's result adds a "You last wrote in this thread at …" line (not_read_since).
• Chat or Inbox (new rule; replaces "when the person isn't here"). A decision in the Inbox stays there. When the person writes in the chat, the agent says what waits and links the thread (redundancy is fine). It never also asks it in the chat with a blocking question tool (AskUserQuestion or similar). The same wording is used in every channel. A chat question is only for something new the person raised in the chat that needs no pictures or files, or when they ask. This is in the agent tool's lead and the server instructions (INBOX_CHANNEL_RULE), the reminder (INBOX_REMINDER_TEXT, _NO_WAKE, so after a compaction too), the thread list header (INBOX_SESSION_THREADS_CHANNEL), plannotator inbox --help ("Chat or Inbox"), the skill and note v3.
• Every rule through channels that update by themselves:
  • the wake note, read_thread and search leads;
  • the asked_by / include_resolved descriptions;
  • the Recommended: line in body;
  • the server instructions.
    These all fit within the 2,048-character caps: instructions 2,042, agent tool 1,910 (about 2,029 with the Board's card line). The Pi tool cost was re-measured: 9,298 bytes. --help also gains the reply_to check: take the id from the Inbox, never from memory.
• Note v3 and auto-refresh. refreshInboxNotes runs at every Inbox start, so an update reaches installed notes when the Inbox restarts onto it. It works deterministically on the existing marker rules:
  • It finds exactly one start line before exactly one end line; anything else is damaged and left untouched.
  • It replaces the block only when the block is byte for byte an earlier Plannotator version and the one its record names.
  • It never touches a block the person edited or removed, a file with no record, a read-only file, a file with other hard links (a dotfiles copy), or a file that changed (bytes or mode) between the read and the write.
  • It keeps CRLF, the BOM and the file mode. A symlink is written at its target and stays a link. The write is atomic. A shared file is written once.
  • agent-notes.json is written before the file and put back if the file's write fails. If the records can't be written, the note isn't written either.
  • Every version's note text is literal. The SHA-256 of the v1, v2 and v3 blocks is pinned in a test, so a released note can't change without a new version.
  • Tests cover every case: packages/server/inbox-agent-notes-refresh.test.ts.
• send_message's result hints are built as a list, so this PR and the recommended-marker PR merge cleanly in either order (verified both ways).

Audit: which channel carries each rule, per host

Hosts:

• CC = Claude Code (tool on by default).
• Pi and OC2 = OpenCode 2: the tool comes from the Settings switch.
• OC1 = OpenCode 1: the tool, with no wake.
• Codex/others = MCP server after the one-time connect step.

Rule	Tool / field text (every request)	Server instructions	Session-start reminder	After compaction / resume	Wake (reply as a turn)	Send result
Chat or Inbox	lead: CC, Pi, OC1, OC2	Codex/others	CC, Pi, OC1, OC2	reminder + thread header: CC (both), Pi (compaction)	n/a	—
reply_to from the Inbox, never memory	reply_to field: all hosts; wake note: CC, Pi, OC2	all MCP hosts	—	header check: CC, Pi	instruction line: CC, Pi, OC2, Codex	not-read line: all hosts
Recommended: line, never in a label	body field: CC, Pi, OC1, OC2; send_message guide: Codex/others	—	—	—	—	nudge (other PR): all hosts, server-side
Name a thread the person started	thread field: all	—	—	—	INBOX_NAME_THREAD_LINE: CC, Pi, OC2, Codex	namingLine; name_this_thread in read_thread / wait_for_reply: all
:::question format	body / send_message description: all	pointer: all MCP	—	—	—	—
Find threads (list, search)	read_thread / search leads, asked_by, include_resolved: all	all MCP	—	"More:" line: CC, Pi	closing line: CC, Pi, OC2, Codex	—

Where a host has no automatic channel:

• OpenCode 1, OC2 and Codex: no compaction or resume list. They get the rules from the field descriptions on every request.
• OpenCode 1: no wake.
• Codex: has no tool until the one-time connect step. Whether Codex surfaces MCP server instructions is unverified.
• The note: opt-in, and carries nothing that is not also above.

Live checks still owed

• The guarded SessionStart "skipped" line. The harness can't tell it apart, because next.trace lists non-matching links as bypassed.
• The env-reset ordering: whether the compact command hook runs before the next turn.step/turn.start.
• Whether session.compact reaches the mod under the real guard.
• The first Inbox start on this build refreshes an installed v1 note to v3. Seen in tests and the real-start test; not yet watched on a real machine.

Tests

• New: inbox-agent-notes-refresh.test.ts, agent-link-session-threads.test.ts, session-threads-text.test.ts; register / mod cases.
• The refresh tests also cover: hard links, an unwritable agent-notes.json, the mode changing, a record naming v1 over a v2 block, and the pinned hashes.
• The reminder tests compare against the exported constants and check what each line must carry.
• Typecheck is clean, and builds ran in order.
• The only failures under full load were timing or port flakes (mod/inbox.test.ts real processes, live-proxy-node-port), which pass alone.
• The mod harness passes 62/62.

Approved by the owner in the Inbox, 2026-10-27