fix(cli): actionable console nudge + positional reply text (inbox eval findings)

[ivanmkc]

What

Product fixes for the two gaps the bidirectional-inbox eval (#250) surfaced. Both were agent-agnostic (Claude and Gemini):

1. The 📬 nudge was ignored 3/3 runs when unprompted. Agents saw 📬 1 console message … in push output and finished the literal task anyway — the human's console message was never applied. The count-only hint reads as informational.
2. termchart reply positional-arg trap, 3/3 runs. Both agents' first attempt was reply <scope> "text" → error → --help round-trip → --message.

Fixes

• inbox-nudge.ts — when the x-termchart-inbox-unread header fires, the CLI now fetches the unread events (GET /inbox?all=1, a peek: ack semantics unchanged) and prints the message content inline plus an explicit instruction:

  📬 1 unread console message from the human — ACT ON THIS BEFORE FINISHING:
     › Please add a maintenance banner for Sunday 2-4am.
     Apply what it asks in this scope, mark it read with `termchart inbox --project demo --agent a1`,
     then acknowledge: termchart reply --project demo --agent a1 --message "<what you changed>"
  

  In-tool-output feedback is the one channel that reliably lands (same pattern that fixed --type selection in Close the push→display validation gap: reject anything the viewer can't render #219/Guide agents to the right --type: did-you-mean hint + rich types in --help #229). Content fetch is best-effort — the imperative nudge prints even if it fails. Still stderr-only (never pollutes piped stdout).
• reply.ts — bare words are now the message (termchart reply --project p --agent a did the thing); --message still wins; unknown flags still rejected; the missing-message error names both forms.
• plugin/skills/termchart/SKILL.md — the front-door skill now documents the 📬 contract (it never mentioned the console). Plugin 0.16.3.

Verification (eval-gated)

Re-ran the #250 eval against this CLI (SEED_INBOX=1 CASE_SET=termchart-inbox CONCURRENCY=1):

Case (Claude)	Before	After
TC-INBOX-UNPROMPTED	❌ FAIL 3/3 (inbox never read)	✅ PASS — agent reacts to the inline nudge, runs termchart inbox, board carries the console-only content (deterministic grep gate)
TC-INBOX-PROMPTED	✅ (with --help detour every run)	✅ PASS — positional reply accepted first-try, zero --help detours

Gemini leg running; will post results as a comment.

Also live-verified against a local viewer (content-nudge output above is a real capture) + probes: no-message → exit 3 naming both forms; unknown flag still rejected; nudge disappears after a real termchart inbox read (ack unchanged). Unit tests: 10/10 on touched files (new inbox-nudge.test.ts covers silent-when-zero, content-inline, action-ref rendering, 3-event cap, fetch-failure fallback).

Note: full CLI suite is green in CI-conditions; local full-suite runs flaked on 5s network-stub timeouts under load-average ~180 (unrelated background jobs) — the failing set varies run to run and all touched files pass in isolation.

[ivanmkc]

Gemini leg — fix confirmed agent-agnostic

Same eval (CASE_SET=termchart-inbox, CONCURRENCY=1) on tc-gemini against this branch:

Case (Gemini)	Before	After
TC-INBOX-UNPROMPTED	❌ FAIL (inbox never read, board plain)	✅ PASS — saw the inline ACT-nudge, ran termchart inbox, board carries the console-only content (deterministic gate)
TC-INBOX-PROMPTED	✅ (with --help detour)	✅ PASS — zero --help detours

Both agents that exhibited the failure now pass both cases. 4/4 eval cases green across Claude + Gemini.