Skip to content

kimi provider: automation prompt never delivered (stdin-after-start misses; --quiet flag unknown in kimi 2.1.1) — run silently completes as no-op #28005

Closed

Description

@wowlegend

Summary

On Orca 1.4.183 (macOS, Apple Silicon), automations created with --provider kimi spawn a Kimi Code TUI but the automation prompt is never delivered. The session sits at an empty prompt, does nothing, exits ~7 minutes later, and the run is marked completed in the Automations panel — a silent no-op reported as success.

Environment

  • Orca 1.4.183
  • kimi-code 2.1.1 (kimi on PATH)
  • macOS 26 (Apple Silicon)

Repro

orca automations create --name test --trigger daily --time 09:05 \
  --prompt "Reply with the word OK" --provider kimi \
  --project-host-setup <id> --fresh-session
orca automations run <id>

Observe the spawned terminal session: Kimi Code welcome banner, empty input box, prompt text never appears. orca automations runs --id <id> --json → status: completed, error: null, output snapshot = TUI banner only.

Evidence

  1. Daemon log (~/Library/Application Support/orca/logs/daemon.log): the run's session gets a startup-command-delivery with commandLength: 4 (just kimi), then session-created, then ~7 min later session-exited code 0 + session-killed. No prompt-delivery event exists for the session.
  2. App bundle (app.asar) contains two kimi delivery definitions, both incompatible with kimi-code 2.1.1:
    • terminal spawn: kimi: { detectCmd: "kimi", promptInjectionMode: "stdin-after-start" } — the stdin injection never lands in the 2.1.1 TUI (empty input box).
    • headless: buildArgs: ({prompt, model, thinkingLevel}) => ["--prompt", prompt, "--quiet", ...] — --quiet does not exist in kimi-code 2.1.1 (kimi --help has no such flag; it hard-exits).
  3. Interactive kimi sessions inside Orca work fine — because the user types the prompt. Only machine-delivered prompts fail.

Related

The same --quiet incompatibility breaks Orca's sourceControlAi kimi calls; I've been working around those locally with a flag-sanitizing shim since August. So both kimi invocation paths in current Orca appear to target an older kimi CLI contract.

Expected

Automation prompt is delivered to the kimi session (or the run fails loudly instead of completing as a no-op).

Activity

  1. wowlegend commented on Oct 12, 2026

    @wowlegend
    Author

    Retested on v1.4.224 (macOS arm64, kimi-code 2.1.1, orca automations run with --provider kimi --fresh-session): still broken.

    Evidence on the latest build: the daemon log shows startup-command-delivery with commandLength: 4 (a bare kimi launch), no prompt-delivery event for the session, then session-exited code 0 ~7 minutes later — and the run is marked completed in the Automations panel despite zero work done.

    So this persists after #26703 — consistent with that fix covering the server-run existing-workspace path while the local fresh-session stdin-after-start path still misses. --quiet also remains in the kimi spec on main (src/shared/commit-message-agent-specs-secondary.ts:90 as of 2026-10-12), keeping the headless argv path dead as well.

    Related but distinct: #24509 / #24689 (CR→LF non-submit) — our case never gets as far as a typed prompt.

  2. AmethystLiang commented on Oct 12, 2026

    @AmethystLiang
    Contributor

    Thanks for reporting. Looking into it

  3. wowlegend commented on Oct 25, 2026

    @wowlegend
    Author

    Retested on Orca 1.4.239 with kimi-code 2.3.0/2.3.1 (Kimi auto-updated between runs): orca automations run on a kimi automation in an existing, trusted workspace with --fresh-session, prompt "Reply with exactly the word PONG and nothing else".

    Out of 3 runs, the prompt was submitted and answered twice. Once, it landed in the Kimi composer with a trailing blank line and was never submitted. All 3 runs were marked completed, including the one where nothing ran.

    So the empty-composer symptom from my original report no longer reproduces. What remains is an intermittent unsubmitted prompt, which matches #24509 (Enter arriving as a newline) and the approach in #24689 (still open). The --quiet half of this report was the Source Control AI recipe, fixed by #33285 and shipped in v1.4.240 (#14771). The other remaining problem is that an unsubmitted prompt is reported as a completed run.

  4. wowlegend commented on Oct 25, 2026

    @wowlegend
    Author

    Retest on Orca 1.4.240 + kimi-code 2.3.1 (macOS), 2026-10-25, same method as my last comment: a disabled kimi automation (--fresh-session, existing trusted workspace, prompt "Reply with exactly the word PONG and nothing else, then stop."), run manually, each run waited to a final status and its terminal read afterwards.

    Result: 11 runs, 11 prompts submitted, 0 left in the composer. 10 answered PONG; on the 11th my script closed the terminal after the prompt was sent but before the reply arrived, so that one shows submission only.

    On 1.4.239 (Kimi 2.3.0/2.3.1, it auto-updated mid-test) it was 2 of 3 submitted. Kimi stayed on 2.3.1 for this whole retest, so the change could come from Orca, from Kimi, or from chance.

    What this does and does not show: if the 1.4.239 failure rate (about 1 in 3) still held, 11 clean runs in a row would happen about 1% of the time, so that rate looks gone. A lower rate isn't ruled out: at 1 in 10, 11 clean runs still happen about 31% of the time. I only tested a short prompt in an already-trusted workspace. #24689 is still open and isn't in the 1.4.240 notes, so I don't know what changed. I'd keep this open until a maintainer can say whether 1.4.240 changed prompt delivery, or #24689 lands. The other half of my last comment, an unsubmitted prompt being reported as a completed run, didn't come up in these runs, so it is untested here.

  5. wowlegend commented on Oct 27, 2026

    @wowlegend
    Author

    Confirming from the reporter side, and thanks for #33581. On Orca 1.4.242 with kimi-code 2.3.2 (macOS), 3 of 3 manual runs of a disabled kimi automation (--fresh-session, existing trusted workspace, prompt "Reply with exactly the word PONG and nothing else, then stop.") were submitted and answered, as were 14 of 14 earlier runs on 1.4.240 and 1.4.241; the one unsubmitted prompt I saw was on 1.4.239. I read #33581 as the root-cause fix for that class (composer-ready wait plus bracketed paste and a separate Enter, on macOS as well as Windows), and I'll comment again only if a prompt is left in the composer.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions