Add the kimi CLI as a /kickoff worker agent - #51
Merged
Conversation
kimi-code has no positional launch prompt, no initial-prompt env var, and piped stdin only prefills the input box without submitting (and would steal the TUI's tty). `-p` is the only way in, but it is mutually exclusive with both --auto and -y and exits after one answer. It does run tools unattended though, and `kimi -c` inherits its full session history — so the worker launches in two phases: `-p` seeds and works the task one-shot, then `exec kimi -c --auto` takes over as the interactive autonomous session with that history. All probed live against kimi-code 0.31.1. - Registry: --kimi -> kimi:kimi-code/k3-256k (commit,pr). The model id must be the QUALIFIED alias; the bare name aborts at startup like an unknown model. - Readiness is model-aware (grok precedent) via `kimi provider list --json`, bounded. Auth probes credentials/, not the same-named oauth/ dir — that file stays 0 bytes even when logged in. `kimi doctor` only validates config syntax. - A well-formed listing offering no models means unavailable; a document without the `models` section is drift and falls back to trusting auth. - Values ride as "$1"/"$2" positionals, never spliced into the script text: `-p` swallows the next token, so a concatenated argv is one reordering away from silently eating a flag. - Tests assert the exact word list plus that structure, and execute the resolved argv against a logging stub to check what each phase received. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SkXCG4uaXDEyf3ciKYKu2i
Completes the kimi worker surface started in the registry commit. - /continue: the reopen caveat now names `kimi -c` alongside `codex resume --last` / `grok -c`. Reopen still always sends `claude -c` (the worker isn't persisted per task), so the caveat stays inline. - /kickoff: --kimi in the flag list, kimi in the non-claude announce class, and the manual launch block spells out the two-phase form. Called out that shell-quoting is load-bearing there: one argv word is a script carrying `;`, `$` and quotes, so an unquoted render would run the `;` in the user's own shell and expand $1/$2 there. - Lifecycle verified rather than assumed: agent_name comes from the registry's `name=` and agent_status from herdr's pane hooks, so nothing reads argv[0] and the `sh -c` wrapper (which execs into kimi) leaves tab teardown and state detection unchanged. - Docs: both READMEs, marketplace description, CHANGELOG, and the kickoff-agent-selection knowledge entry, which records the launch-shape constraints and why grok's "empty listing = inconclusive" rule had to be re-derived for JSON. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SkXCG4uaXDEyf3ciKYKu2i
Owner
Author
🐝 Swarm review (local ensemble) · reviewed at e534812PR #51: Add the kimi CLI as a /kickoff worker agent
Finding 1 (das einzige critical) wurde in der Session live widerlegt: der Mechanismus lautete 'kimi -c continues the globally most-recent session', aber -c ist cwd-scoped (kimi --help: 'for the working directory'). Gegenprobe: in einem älteren Session-Verzeichnis resumierte -c dessen eigene Session, obwohl die global jüngste Session woanders lag. Zwei Worktrees = zwei cwds, das Parallel-Kickoff-Szenario existiert nicht. Berechtigter Restpunkt: kein Test deckt einen fehlschlagenden Seed ab. 3 weitere Findings wurden vom Verifier refuted, darunter die Behauptung, herdr sehe wegen des sh -c-Wrappers agent=sh — durch einen echten herdr-Probe-Pane widerlegt (agent=claude bereits während der Pre-exec-Phase). Findings 2, 4, 5, 9, 12, 13, 17 sind dasselbe Versäumnis: Skill-Bodies wurden nachgezogen, die Aktivierungsflächen (Frontmatter-Descriptions, Argument-Tabellen) und /adopt nicht. Lokaler Swarm-Review (Claude-Lenses + codex + grok), keine Backend-Fehler, keine Redactions. Local mixture-of-agents review (Claude lenses + codex + grok) run from the author's machine — not a hosted bot. Verdicts (✅/🟨/❌) are the runner's own assessment. |
Follow-up to the swarm review on PR #51: the first pass updated skill bodies but missed the activation surfaces and /adopt entirely, while README claimed /adopt takes the same worker flags. - /kickoff: frontmatter description + flag table; /adopt: both selector enumerations + the manual-launch block, incl. the load-bearing `sh -c` quoting caution; /close: worker-degradation prose reads "non-claude" - herdr-launch.sh usage header, plugin.json description, CLAUDE.md plugin line (v1.11.x), and the herdr-kickoff-automation knowledge entry, which documented only the codex/grok `-m` argv - /kickoff's non-claude announce rule now tests "not `claude:`" rather than a codex:/grok:/kimi: allowlist that needs a hand-edit per new CLI - Drop the dead `-z` arm in the kimi probe: empty output already fails the `"models"` grep and lands in the same branch Not applied (kept in the review for the record): the `critical` finding claimed `kimi -c` resumes the globally most-recent session. It does not — `-c` is cwd-scoped (verified live: in an older session's directory `-c` resumed that directory's session while the newest session lived elsewhere), so two worktrees cannot cross-resume. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SkXCG4uaXDEyf3ciKYKu2i
Second swarm round on PR #51. Applied the agreed findings: - A failed seed no longer passes unnoticed. `&&` would kill the pane and a bare `;` let the error scroll away behind the TUI's first repaint, leaving a tab that looked like a working worker but never read TASK.md. It now reports and waits for a keypress before handing over. Verified live that a failed seed in a fresh worktree opens a NEW empty session, never a foreign one — `kimi -c` is scoped to the working directory. - `resolve` emits `argv_shell=`: the argv words POSIX-single-quoted by the registry. /kickoff and /adopt print it verbatim rather than re-deriving quoting from a prose rule — for kimi a mis-quote would run the `;` in the user's own interactive shell. Uses shell_quote, not `printf %q`, which on bash 3.2 emits per-character backslashes and `$'…'` for non-ASCII: correct but unreadable before pasting, and bash/zsh-only. The seed message is ASCII-only for the same reason. - The model check matches the alias in KEY position, so it can't fire on the alias appearing as a value while `models` is empty. The real document is flat-qualified (verified live), which also refutes the nested-schema half of that finding. - Tests: the previously untested failing-seed path, plus the argv_shell shape; the "prompt not spliced" check compares the whole prompt word now that the script text legitimately mentions TASK.md. - Stale pointers/enumerations: adopt + kickoff cited "/kickoff step 12" for a launch form living in 13b; manager-worker-orchestration and the knowledge index still listed the pre-kimi worker set. The selector hint is derived from REGISTRY instead of restated. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SkXCG4uaXDEyf3ciKYKu2i
Outcome of the security finding both swarm rounds raised: kimi is the only worker that acts without tool-approval prompts. Kept deliberately — the seed cannot have approvals (`-p` refuses --auto/-y and is the only way to deliver the task), and dropping --auto from phase 2 would leave the phase that does the work just as unattended. Mitigation is visibility, matching the existing announce-not-prompt precedent: - /kickoff states the unattended start whenever the worker resolves to kimi, by flag, repo default or picker alike - /adopt warns additionally: its TASK.md is summarized from another branch's commits, so it is the one path where an unattended worker acts on content the user did not write - README says plainly what the autonomy costs; the knowledge entry records the decision and what a future revisit should change (a second registry entry, not the --auto flag) Also sync the launch form in README/CHANGELOG/knowledge with the seed-failure handler added last commit — the docs still showed the bare `;` variant, which is the same doc drift the review flagged twice. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SkXCG4uaXDEyf3ciKYKu2i
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
/kickoffworker alongside claude/codex/grok:--kimi→kimi:kimi-code/k3-256k.<cli> -m <model> <prompt>launch form, so it launches in two phases — every constraint behind that shape was probed live against kimi-code 0.31.1, not inferred from docs.-mid aborts kimi at startup — a stale model would otherwise hand the user a worker tab that dies on sight.Changes
Registry + launch argv (
agent-registry.sh)--kimi | kimi | kimi-code/k3-256k | commit,pr. The model id must be the qualified alias; the bare name is rejected exactly like an unknown model.sh -c 'kimi -m "$1" -p "$2"; exec kimi -c --auto' kimi-worker <model> <bootstrap>. Rationale: kimi has no positional launch prompt (kimi "text"→ "unknown command"), no initial-prompt env var, and piped stdin only prefills the input box without submitting (and would steal the TUI's tty in a pane).-pis the only entry point but is mutually exclusive with both--autoand-yand exits after one answer. It does run tools unattended, andkimi -cinherits its full session history — so the seed works the task through once, thenexechands over to the interactive autonomous session.execre-roots the herdr pane at kimi;;(not&&) keeps phase 2 alive if the seed fails, leaving a usable tab.kimi provider list --jsonfor the model, plus an auth probe oncredentials/kimi-code.json— not the same-namedoauth/file, which stays 0 bytes even when logged in.kimi doctoronly validates config syntax and is not an auth check.modelssection is treated as format drift (trust auth); a well-formed listing offering nothing means unavailable. grok's "empty output = inconclusive" rule does not transfer to JSON, where{"models": {}}is non-empty but a real answer.Tests (
test_agent_registry.py)-pconsumes the next token, so a concatenated argv is one reordering away from silently eating a flag (kimi -p --auto "…"turns--autointo the prompt). The model and prompt therefore ride as"$1"/"$2"positionals, never spliced into the script text — asserted structurally, plus the exact word list.Skills, lifecycle, docs
/continue: the reopen caveat now nameskimi -cnext tocodex resume --last/grok -c(reopen still always sendsclaude -c; the worker isn't persisted per task)./kickoff:--kimiin the flag list, kimi in the non-claude announce class, and the manual launch block documents the two-phase form — including that shell-quoting is load-bearing there, since one argv word is a script carrying;,$and quotes.agent_namecomes from the registry'sname=andagent_statusfrom herdr's pane hooks, so nothing readsargv[0]and thesh -cwrapper leaves tab teardown and state detection unchanged.kickoff-agent-selectionknowledge entry.Readiness
test_agent_registry.pypasses;check-structure.py0 errorsplugin.json+marketplace.jsonin synckickoff-agent-selection.mdupdated (launch-shape constraints + the JSON/empty-listing distinction)Test plan
bash plugins/work-system/scripts/agent-registry.sh listshowskimi:kimi-code/k3-256kas available/kickoff <task> --kimiopens a worktree tab; phase 1 works the task, phase 2 lands in an interactiveautosession carrying that history/closeon a kimi worker tab tears it down (Scenario A) — not yet exercised live; only the code paths were verifiedagent-registry.sh resolve --kimiwith a bogus model in the registry reports unavailable instead of launching🤖 Generated with Claude Code
https://claude.ai/code/session_01SkXCG4uaXDEyf3ciKYKu2i