[omartelo]

When a plan's usage limit stops a turn, lich now says so on the card and picks the turn back up once the limit resets. Gap #14 of the ADE triage (Claude Desktop week 33, cmux 0.65.0).

How each provider signals a limit (measured)

Provider	Signal	Reset time	What lich does
Claude Code 2.1.293	StopFailure with error: "rate_limit" instead of Stop; a synthetic assistant line in the transcript	quotaLimits.resetsAt on that line	reads it on every done and parks the continuation
Codex 0.144.5 (codex exec)	no hook: UserPromptSubmit fires, Stop never does; the rollout writes a token_count with a window at 100%, then task_complete with no message	rate_limits.*.resets_at	reads open Codex turns every 30s, ends the turn as interrupted, parks the continuation
opencode	session.status retry	only the next attempt	nothing: it retries by itself
oh-my-pi	auto_retry_start / auto_retry_end	no	nothing: it retries by itself
Antigravity, Crush, Cursor CLI, Kiro CLI	no failure event	no	out, documented

Claude Code 2.1.293 also ships its own autoContinueAtUsageLimit, gated per account by the server (off on the account measured). lich lives with it rather than toggling it: see below.

Design

• Signal. Claude: the plugin reports StopFailure as done (fix: report Claude Code's StopFailure as done lich-plugin#48), and noteLimit reads the limit off the transcript on every done. Codex: RunLimitWatch reads the rollout of each open Codex turn every 30s and ends a limited one through noteInterrupt, which also fixes the card spinning forever.
• Continuation. relay.ParkResume parks a scheduled prompt (the existing one, later.go) due two minutes after the reset, so it shows on the card, survives a restart and is typed like any scheduled prompt. Nothing is parked for a reset already past or more than 24h out (Claude Code's own horizon), nor over a prompt already parked.
• No double continuation. Any busy on that session before the prompt is due drops it (dropResume): the person going on by hand, Claude Code's own auto-continue (which resumes at the reset itself, hence the two-minute grace), or a relay delivery. A prompt the person parked in its place stays.
• Card. A new amber rung: Session limit · resumes in 2h, or Weekly limit · resets Mon 09:00 when nothing will resume it. Cancelling goes through the existing Scheduled… item and its Cancel it.

Ceilings (in docs/ceilings.md)

• The continuation is told apart from a person's prompt by its exact text (ponytail: in resume.go); a column on the row is the upgrade if that ever matters.
• The Codex read is inferred from the rollout's shape and was measured on codex exec, not on the TUI.
• The limit rung lives in the window only; a reload leaves the parked prompt as a plain scheduled one.

Provider support rows added to README.md and README.zh-CN.md; hook contract updated in docs/hooks/session-state.md (no new state, no fixture change).

Refactor: the card's status rung

SessionCard.tsx was already past the 800-line ceiling, and this adds one more rung to its ladder. The ladder, and the stores only it reads (relay, handoff, inbox, limit, todo, tool, waiting reason), move to SessionStatusRung.tsx; the card keeps one line and drops to ~710 lines.

render-budget.test.tsx changes because the contract did, not to buy a green run: the rung is a child the card repaints with, so it joins CARD_CHROME, and a todo event now repaints the rung alone instead of the whole card (narrower than before).

Tests

• internal/terminal/limit_test.go: transcript and rollout readers, against the lines measured on a real Claude transcript and a Codex rig (fake responses endpoint answering usage_limit_reached).
• internal/terminal/limit_service_test.go (!windows, reuses stubBins): done on a Claude limit parks the resume; the watch ends a limited Codex turn.
• internal/relay/resume_test.go: park, horizon, the person's prompt, drop on busy.
• Frontend: session-limit-store.test.ts, limit-line.test.ts.

Local gate: gofmt, go vet, go test ./..., the cross-compile loop, biome ci (exit 0), vitest with coverage (1973 passing), pnpm build. The new async tests ran 100 times clean.

Needs lichdotdev/lich-plugin#48 for Claude Code; Codex works without it.