[omartelo]

Two failures with the same shape: a session row outlives what it points at, and
the spawn goes ahead anyway. Both end with the user looking at a terminal that
explains nothing.

The resume that cannot happen

Reopening a session that had run Claude Code always raised the resume prompt.
The decision read the row alone (resumableSession, sessions.ts):
kind === "claude" plus a non-empty providerSessionId was taken as proof the
conversation still existed. It is not — Claude Code prunes its own transcripts,
so an old enough id outlives what it points at. Accepting the prompt spawned
claude --resume <id> against nothing, and Claude's own error landed in the PTY
in place of a session.

It was a documented shortcut, taken back when a resume only ever followed a
restart (terminal.go, Start):

An id Claude no longer knows fails in the PTY like any other bad invocation —
the user sees Claude's own error.

Parked worktree sessions age, so it stopped being rare.

The check was already written: claudeTranscriptPath locates a conversation's
transcript by id for the context-window readout, honouring CLAUDE_CONFIG_DIR.
ResumeAvailable asks it for existence alone. The prompt is now raised only for
a conversation still on disk; when it is gone the session starts fresh and says
so. Offering a choice that cannot be honoured was the bug.

The checkout that is gone

A worktree removed outside lich leaves its session behind, pointing at a
directory that no longer exists. Terminal.Start rejected, and nothing caught
it (TerminalView.tsx) — the terminal mounted empty, in silence.

Nothing recovers that session, so the gate settles the directory first and
closes it. Closes, not deletes. An absent path is not proof of a permanent
one: a mount that was not up yet answers exactly the same, and a deleted row
takes the provider conversation with it. Parking costs the same and loses
nothing — re-creating the worktree picks the session back up through
ReopenWorktreeSession, which already keys on the path.

WorkdirMissing is deliberately asymmetric: only a directory provably absent
(or a path that is not a directory) answers true. An empty cwd, a permission
error, a mount answering badly all fall through to the spawn, because closing a
session on a check that never answered is the worse trade.

The floor under both

Every other spawn failure — a provider binary that is not installed, a wrong
path in Settings › Providers, an exhausted fd table — reaches the frontend as
one opaque string from startPTY, indistinguishable from each other. They all
used to fail into the same blank terminal. They now surface as a toast, with the
session left in place: all of them are worth another try once the cause is
fixed, which is exactly why none of them may close a card.

Shape

The decision moved to lib/session/spawn-gate.ts, with its two backend checks
injected. The host's effect was already at the size limit, and the frontend gate
cannot render — injection is what puts the four outcomes, and both fail-safe
directions, under a test.

Test plan

• gofmt -l ., go vet ./..., go test ./... (529 pass)
• biome check (12 warnings, all the standing a11y backlog), vitest run (501 pass), tsc --noEmit + vite build
• Cross-compile loop for linux/darwin/windows
• TestResumeAvailable: transcript on disk, pruned transcript, empty id, shell session, another provider
• TestWorkdirMissing: existing directory, empty cwd, deleted checkout, path that is a file
• spawn-gate.test.ts: all four verdicts, plus each check failing (workdir spawns anyway, resume still asks)
• By hand in task dev — the frontend gate is node-only, so the render path is unverified:
  • remove a session's transcript from ~/.claude/projects/ → no prompt, a toast, a working fresh session
  • transcript present → prompt still resumes
  • rm -rf a worktree behind an open session → toast, card closes, re-creating the worktree brings it back
  • point Settings › Providers at a binary that does not exist → toast naming the failure, card stays