[omartelo]

Closes #577.

README.md called all eight providers first-class; docs/ceilings.md recorded what each one actually reaches. A contributor document is not where a stranger picks a binary, so the spread now lives beside the provider list, in what the user sees.

The table

New ## Provider support section in both READMEs, rows = readings, columns = providers:

What you get	Claude Code	Codex	Antigravity	opencode	oh-my-pi	Crush	Cursor CLI	Kiro CLI
Context window in the footer	yes	yes	no	no	no	no	no	yes
Cost in the footer	yes	yes	no	yes	yes	yes	no	credits
How much of your plan is left	yes	yes	no	no	no	no	no	no
Spinner while a turn runs, ring when it ends	yes	yes	yes	yes	yes	no	no	yes
Bell when the agent is blocked on you	yes	yes	no	yes	no	no	no	no
Machine kept out of idle sleep while a turn runs	yes	yes	yes	yes	yes	no	no	yes
Review tab's Last turn, with the agent's recap	yes	yes	yes	yes	yes	no	no	yes
Search the conversation from the palette	yes	yes	yes	yes	yes	yes	no	yes
Fork a conversation into a new worktree	yes	yes	no	yes	no	no	no	no

It is read off usageSourceFor, contextUsageFor, quota.Plans, providers.SupportsFork and docs/hooks/session-state.md, not off the issue's own table, which is wrong in one row: Antigravity has palette search and the last-turn recap, because transcriptReaderFor reads its JSONL. Cursor CLI is the only provider with neither.

"first-class" is spent rather than repeated. Every provider is spawned, resumed, sandboxed and reachable from other sessions the same way; what differs is what lich can read back, because it reads what each CLI writes down.

The Review tab stops hiding it

On a Crush or Cursor CLI session, {switchable && <SourceRow/>} meant the Working tree / Last turn control simply never appeared, which reads the same as a bug. The strip is now drawn with Last turn dead under a line naming why, the idiom SessionForkItem already uses for a card that cannot fork.

A session whose provider does report is untouched, including one that has not reported yet: its switch is seconds away, and turnUnavailableReason deliberately stays silent there rather than blaming a provider for half a second. The reason goes on its own line because the dock resizes down to MIN_REM, where a sentence sharing the row with the toggle group has nowhere to go.

Paperwork

• docs/ceilings.md: the "neither the switch nor the recap band is ever drawn" clause is gone, replaced by the trap it leaves behind, that turnSwitchable reads the session's own reports while turnUnavailableReason is a hand-written list, so the two disagree until the list moves with the contract.
• Rule 5 in CLAUDE.md and docs/adding-a-provider.md now point at the table: a provider added later answers every row, or ships a column that reads as a provider giving you nothing.
• CHANGELOG.md under [Unreleased].

Out of scope

No new transcript reader. Antigravity already has one and what it lacks is a window and a cost its CLI never writes; Cursor files a chat as a content-addressed blob store whose order lives in a protobuf index, and neither schema is a contract anybody promised to keep.

Test plan

• biome ci ., vitest run, tsc --noEmit all exit 0 (run as the direct binaries; pnpm exec masks the exit code here)
• gofmt -l ., go vet ./..., go test ./... clean
• turnUnavailableReason covered four ways: names the provider on the two that report no turn, says what the absence costs, stays silent for a provider that reports, stays silent for a shell and for no active session
• disabled confirmed a first-class base-ui Toggle prop that renders a natively disabled button, so toggleVariants' disabled:opacity-50 applies
• Eyeball the dead switch on a real Crush card at MIN_REM and at full screen