mail CLI: add a seat-specific --self/--to selector + document the skill-owned consume path (multi-seat repo ergonomics)

[satur8d]

Problem

The totem mail CLI under-exposes two things that exist in the mechanism but not the front-door, making multi-seat repos awkward to operate by hand. Surfaced by lc-codex (2026-06-19) during an lc blind-round reply, sanity-checked by strategy-claude (the doctrine is correct; this is CLI exposure/docs, not a doctrine change).

1. No seat-specific self selector

totem mail --json resolves selfAgents.source = "dirs" → every agent dir in the repo is treated as "self." From liquid-city that returned ["lc-agy", "lc-claude", "lc-codex", "lc-gemini"] and listed mail addressed to lc-claude, not just the querying seat. Fine for a repo-level dashboard, surprising when the operator asks one seat to "check for a dispatch."

The seat-specific mechanism already exists — TOTEM_SELF_AGENT env (the SessionStart hooks call pollMail({ env: { TOTEM_SELF_AGENT: <seat> } }), and mmnto-ai/totem-strategy:scripts/poll-cohort-mail.mjs takes --to <seat>) — but totem mail surfaces no --self <seat> / --to <seat> flag.

2. No consume command, and no docs pointer saying so

totem mail --help exposes show/send/reply but no consume/mark-processed op. This is intended — processed markers are written by the consuming tooling (/signoff, the handoff CLI) at consumption time, holding the single-writer invariant (ADR-106 Amendment 1 §A1.3). But the CLI doesn't say it's skill-owned, so a seat has to guess (lc-codex correctly avoided hand-writing a marker, but only after uncertainty).

Asks

• Add a seat selector to totem mail (--self <seat> / --to <seat>), surfacing the existing TOTEM_SELF_AGENT resolution as a flag.
• totem mail --help note: processed markers are written by /signoff (read/send/reply only here).
• (Optional, same area) fold the known ECL-no-derived-handled-state gap — an inbound "unread" poll doesn't subtract sent replies, so handled-state isn't derivable (unread = inbound − handled).

Not in scope

The ECL doctrine itself — it is correct (ADR-106 §3 glob discovery; Amendment 1 §A1.3 skill-owned markers). This issue is @mmnto/cli ergonomics + docs only.

Provenance

lc-codex dispatch liquid-city:.totem/orchestration/lc-codex/outbox/2026-06-19T2147Z-strategy-claude-mail-tool-self-resolution-gap.md; strategy-claude sanity-check reply 2026-06-19T2155Z-lc-codex-mail-tool-self-resolution-sanity-check.md. Doctrine: ADR-106 §3 + Amendment 1 §A1.3.

[satur8d]

Corroboration — lc-claude independently re-hit both asks (2026-06-29)

A second LC session surfaced the same two gaps this issue already tracks, a different vantage and ~10 days after the lc-codex original — so the signal is recurring and multi-seat, not one-off:

• Ask 1 (seat selector). totem mail from liquid-city listed all 4 LC seats' unread (~309 today); the running seat's actionable mail required manual narrowing. Two LC agents (lc-codex on 2026-06-19, lc-claude on 2026-06-29) hit this independently.
• Optional fold-in (false-unread). Re-confirmed as the dominant pain: mail is never marked read, so every session opens to hundreds of "unread," and genuinely-new dispatches have to be found by Z-stamp > last-journal. The processed/ handled-marks already exist (ADR-106 Amendment 1 §A1.3) — the unread count just doesn't derive from them (unread = inbound − handled, Tenet 20). With the seat selector but without this, --to <seat> still shows all-historical-for-that-seat, so the two asks are complementary; the derived-handled piece is what makes the count a usable signal. Recommend promoting it from "optional" to a first-class ask.

Item 3 (recipient-validator self-agent set) from the same lc-claude note maps to #2017 (agent-list portability) — and lc-claude reports his 2026-06-29 sends to lc-agy/lc-codex went through clean, so it may already be fixed in the 1.81/1.82 line. Flagging for a quick verify against #2017's hardcoded-list audit; close there if confirmed.

Provenance: liquid-city:.totem/orchestration/lc-claude/outbox/2026-06-29T1734Z-strategy-claude-lc-forward-note-totem-mail-dx.md.

[satur8d]

Extension: lc-codex 2026-06-30 — escalate from opt-in selector to seat-scoped default + count semantics

A second operator-surfaced instance, sharper than the 2026-06-19 one this issue was filed on. Provenance: lc-codex dispatch liquid-city:.totem/orchestration/lc-codex/outbox/2026-06-30T2221Z-strategy-claude-strategy-claude-mail-scope-counts.md.

New evidence — the over-broad self-resolution is now worse than §1 recorded. At lc-codex session-start, totem mail reported 594 unread; totem mail --json after a reply showed 595, broken down by <repo> -> <recipient>:

312  totem -> strategy-claude
153  totem-strategy -> strategy-claude
34   liquid-city -> lc-claude
23   totem-strategy -> lc-claude
19   liquid-city -> lc-codex
17   liquid-city -> lc-agy
17   liquid-city -> lc-gemini
10   liquid-city -> strategy-claude
9    totem -> lc-claude
1    totem-strategy -> lc-codex

So from an lc-codex seat the count is dominated by mail addressed to strategy-claude (312 + 153 + 10 = 465 of 594) — i.e. the self-set/scan now reaches beyond the local repo's seats entirely (the 2026-06-19 finding was LC-seats-only: [lc-agy, lc-claude, lc-codex, lc-gemini]). For a seat-scoped startup this is actively misleading: the agent reads/triages outside its mandate and reports a backlog number the operator reasonably hears as local.

Sharpened asks (additive to the §"Asks" checklist above — the opt-in selector is necessary but not sufficient):

• Make seat-scope the default when totem mail is run from a seat (resolved self agent + current-repo scope), not just an opt-in --self. Broader views move behind explicit switches: --all-agents / --workspace-wide (and the existing --self <seat> / --to <seat> / a --repo <repo> filter). A workspace-wide scan is a coordinator/dashboard view, not a seat startup default.
• Human output distinguishes three numbers, not one headline: (a) new since last signoff / last processed for this seat; (b) total unarchived addressed to this seat; (c) workspace-wide total — only when requested.
• --json carries scope metadata: grouped counts by repo and to, plus the resolved active agent and the scope actually used.

Relationship to sender-side GC (mmnto-ai/totem-strategy#700, interim shipped in mmnto-ai/totem-strategy#791). The retention prune shrinks the bulk of that 594 (once totem-claude prunes the 312 totem -> strategy-claude pile and strategy-claude prunes its own), but it is related, not sufficient — even against a clean corpus a seat must not count other seats' mail as its own. GC fixes corpus size; this issue fixes read-side scope semantics. The two are the sender-side / read-side twin of one missing ECL lifecycle.

Doctrine companion (strategy lane): a short doctrine/ecl-discipline.md § 5 note — workspace-wide mail scans are coordinator-mode only; a seat defaults to its own scope — pointing here as the CLI tracking issue. PR forthcoming; this issue stays the canonical home for the CLI behavior change.