v0.8.61: Add pre-release smoke harness for exec streams, model routes, and shell job UX

[Hmbown]

Why

During v0.8.60 live testing, the TUI still exposed several release-smoke failures that should be caught automatically before tagging/publishing:

• Shell/background work can still present as a blocking wait Bash / timeout-shaped UI state instead of a clear foreground-vs-background job model.
• codewhale exec --output-format stream-json can emit terminal OSC/title/progress escape sequences before the first JSON object, making the stream not clean machine-readable JSONL.
• codewhale exec flag order appears significant: codewhale exec --provider zai --model GLM-5.2 --output-format stream-json ... returned plain text, while codewhale exec --output-format stream-json --provider zai --model GLM-5.2 ... emitted stream-json.
• Earlier live testing also saw a successful codewhale exec route exit 0 with no visible final text, which belongs in the same smoke family as reasoning-only / empty-output turns.

This complements #3211, #3212, #3213, and #3166. Those issues define the runtime behavior; this issue makes sure the release process catches regressions across installed binaries and configured model routes.

Evidence from local v0.8.60 binary

Installed binaries under PATH:

codewhale 0.8.60 (652d3925ae05)
codew 0.8.60 (652d3925ae05)
codewhale-tui 0.8.60 (652d3925ae05)

Plain one-shot smoke succeeded across routes:

default_text          rc=0 out=CW_SMOKE_OK
deepseek_flash_text   rc=0 out=CW_SMOKE_OK
zai_glm52_text        rc=0 out=CW_SMOKE_OK
minimax_m3_text       rc=0 out=CW_SMOKE_OK

Stream-json route/order problem:

codewhale exec --provider zai --model GLM-5.2 --reasoning-effort high --output-format stream-json ...
# rc=0, but output was plain text: CW_SMOKE_OK

codewhale exec --output-format stream-json --provider zai --model GLM-5.2 --reasoning-effort high ...
# rc=0, emitted stream-json

Stream-json cleanliness problem:

first output bytes included terminal OSC/title/progress sequences before JSON:
\033]9;4;1\a\033]0;CodeWhale\a...{"type":"content","content":"CW"}

Desired smoke harness

Add a bounded, automatable pre-release smoke script/check that can run against the installed binary and/or target/release binary without publishing anything.

It should cover:

1. Version surface

   • codewhale --version, codew --version, codewhale-tui --version all match the intended release version and git SHA.

2. Exec text route matrix

   • Default route and selected configured providers return non-empty output for a tiny prompt.
   • Include at least DeepSeek Flash, the configured default route, and any authenticated non-DeepSeek routes discovered locally.
   • Hard timeout per route; failures include provider/model/exit/status/stdout/stderr preview.

3. Exec stream-json route matrix

   • --output-format stream-json returns valid JSONL for default and explicit provider/model routes.
   • Every non-empty stdout line must start with { and parse as JSON.
   • No terminal title/progress/OSC sequences on stdout in machine modes.
   • Final metadata/done status must be present and must not be hidden behind reasoning-only output.
   • Flag order must not change output format semantics.

4. Empty/reasoning-only turn guard

   • A route that exits 0 with no final content is a smoke failure.
   • Reasoning-only/no-action responses should be classified and surfaced as recoverable runtime errors, not success.

5. Shell job UX smoke

   • A long shell command can be launched as background/detached without blocking the main turn.
   • Foreground shell remains a deliberate barrier only when the next step needs the output.
   • The UI/status surface should say running, background, waiting for output, timed out, or cancelled clearly; do not display a confusing timeout/wait Bash state as the main bottom-line UX.
   • Cancelling a background job frees the turn and updates sidebar/job state.

Suggested implementation seams

• Add a release smoke script under scripts/release/ or scripts/smoke/ that takes --binary <path> and --version <version>.
• Keep live-provider probes opt-in or auto-discovered from configured/authenticated routes; mock-provider tests should cover CI deterministically.
• Add unit/integration tests around codewhale exec argument forwarding so output-format/provider/model flag order is stable.
• Ensure stream-json disables terminal title/progress/OSC writes on stdout; if terminal signals are needed, send them only to an interactive TTY or stderr.
• Wire the smoke script into the pre-release verification checklist, not the publish step.

Acceptance criteria

• The smoke harness fails if stream-json stdout contains non-JSON bytes before the first JSON object.
• The smoke harness fails if --output-format stream-json returns plain text for any accepted flag order.
• The smoke harness fails on rc=0 + empty output.
• The smoke harness reports provider/model route names for every failure.
• The shell-job smoke proves background work does not freeze the TUI/turn loop and that timeout/cancel states are human-readable.
• The release checklist runs this before tag/publish approval.