feat(kimi): native Kimi Code CLI dialect support (Fixes #1561)

[whyihaveyou]

Summary

Native support for the Kimi Code CLI (Node.js, kimi-code) dialect in the kimi agent, which currently only speaks the legacy Python kimi-cli dialect. Builds directly on the probe-gating approach of #1461 and #1483.

Stacked on #1483 — this branch is based on e1ec734f; only the top commit is new. The diff will shrink automatically once #1483 merges.

Fixes #1561

Root causes addressed (from the analysis in #1561)

#	Incompatibility	Fix
1	--print unknown option	already handled by #1461 (kept)
2	--work-dir unknown option	handled by #1483 (base commit)
3	--quiet unknown option	probe-gated; when unsupported, quiet mode is emulated by suppressing thinking/tool events locally so final-message-only behavior is preserved
4	--resume SID unknown option	-r SID on the modern dialect — the same command the CLI's own resume hint prints
5	--yolo/--auto conflict with --prompt	never passed; bare --prompt auto-approves on both flavors (documented in buildArgs)
6	stream-json content is a plain string (assistant + tool), not block arrays	parser accepts both shapes, format-tolerant, not flavor-gated
7	resume hint moved from stderr plain text to stdout JSON meta ({"role":"meta","type":"session.resume_hint",...})	captured in handleMeta, restoring cross-message continuity
8	sessions live under ~/.kimi-code/sessions with a different state.json schema	dual-root listing, schema support, workDir filter honored

Acceptance criteria (per PM comment on #1561)

• CLI flavor probe at init — kimi --help parsed once at New(); --print absence is the Kimi Code family discriminator (same signal [Bug] ❌ 错误: error: unknown option '--print' (Did you mean --prompt?) #1456 established). Verbatim v0.26.0 --help fixture in probe_test.go
• Flag translation — modern dialect drops --print/--quiet/--work-dir, --resume SID → -r SID (per-dialect buildArgs tests)
• stream-json content — plain-string content accepted for assistant and tool messages, aligned with existing handleAssistant (regression tests)
• permission-mode default — no --yolo/--auto; bare --prompt implicit approval (arg test asserts neither flag appears)
• session resume — -r SID + meta-hint capture; unit tests + live e2e below
• fallback — legacy kimi-cli dialect untouched: all legacy-path tests updated and passing, parser changes are shape-tolerant only, quiet suppression activates solely when --quiet is unadvertised

Test evidence

• go build ./... and go test ./... green (Go 1.25.12)
• Live e2e against the production Kimi Code CLI v0.26.0, exercising the exact cc-connect call path:

  KIMI_LIVE_E2E=1 KIMI_BIN=~/.kimi-code/bin/kimi go test ./agent/kimi/ -run TestLiveE2E -v
  turn1 text: "remembered"   (anchor: remember the number 7391)
  → session id captured from meta hint: session_c85029b1-…
  turn2 text: "7391"         (resumed with -r, anchor recalled)
  → session id stable across resume
  PASS
  

• The translation logic mirrors the Python shim attached to feat(kimi): support kimi-code (Moonshot's new Node CLI, successor of kimi-cli) #1561, which has been running cc-connect ↔ Kimi Code in production (via Feishu) since 2026-07-17 — including the resume-hint relay that this PR makes unnecessary.

Notes

• agent/kimi/live_e2e_test.go is env-guarded (KIMI_LIVE_E2E=1); CI and machines without a Kimi API key skip it.
• Probe failure keeps the existing conservative default (modern surface), unchanged from fix(kimi): conditionally pass --print so newer Kimi Code CLI works (#1456) #1461.
• Session listing now also covers ~/.kimi-code/sessions so /list and /delete work for both flavors; legacy sessions are listed exactly as before.

Update (review follow-ups)

• kimi-code 0.29.1 (macOS, @LunarFeller): chat / multi-turn resume / /list verified. The one cosmetic gap — /list showing 0 messages for every Kimi Code session — is fixed in efc779b4 ("fix(kimi): don't report 0 messages for Kimi Code sessions in /list"). Modern sessions now derive the count and the summary from agents/main/wire.jsonl, counting only user turns (ignoring tool/assistant events), matching the legacy behavior of summarizing from the transcript and falling back to the state title only when no transcript exists.
• kimi-code 0.33 (resume hint moved to a stdout JSON meta event, reported on feat(kimi): support kimi-code (Moonshot's new Node CLI, successor of kimi-cli) #1561): handled natively by handleMeta — the {"role":"meta","type":"session.resume_hint","session_id":...} event is captured and stored, so multi-turn continuity works without the shim's v2 relay. Legacy plain-text / stderr hints are still handled too.
• external cmd sessions don't receive history (feat(kimi): support kimi-code (Moonshot's new Node CLI, successor of kimi-cli) #1561): confirming this is a cc-connect platform limitation that applies only to the external-command path. It is out of scope here because these native Kimi Code changes reuse the built-in agent session/resume logic and are therefore unaffected. Happy to open a follow-up issue if broader external-cmd session threading is wanted.

[LunarFeller]

Tested this branch locally against kimi-code CLI 0.29.1 (macOS + Feishu): chat, multi-turn resume, and /list all work — nice fix.

One minor gap: /list shows 0 messages for every kimi-code session (screenshot attached).

parseKimiSessionDir counts from context.jsonl (legacy layout), but kimi-code stores the transcript at agents/main/wire.jsonl with {"type":"context.append_message","message":{"role":...}} entries. Falling back to that file (filtering origin.kind == "user") should fix the count. Cosmetic only — titles and workDir filtering work fine.

[whyihaveyou]

@chenhg5 — quick status update after the CI failure:

• Synced the branch with latest main (merged it in, no conflicts) and adapted to the newer AgentSession.Send(messageID, …) signature that main introduced — that was the compile failure on the merge ref (not enough arguments in call to s.Send in live_e2e_test.go).
• Fixed the lint nits on the new transcript helpers (checked the two f.Close error returns the linter flagged).
• CI is now green: lint / unit-test / smoke-test / regression-test / performance-test all passed, and the PR is mergeable.

Nothing is outstanding from my side. Ready for the QA pass / review whenever convenient — happy to iterate on any feedback. Thanks!

[chenhg5]

Conclusion: Comment

Overall assessment:
Solid technical work — the dialect detection, flag translation, content-shape tolerance, resume-hint capture, and quiet-mode emulation are all correct and well-tested. Two process concerns need a maintainer call before merge: (1) this branch is stacked on the still-open PR #1483, and (2) there are three other open PRs also targeting #1561 that should be reconciled with this one.

Review scope:

• Read full diff across 8 files (884 / 121).
• Ran go test -race ./agent/kimi/... -count=1 -timeout 90s (race detector required) — 27 tests PASS, including all new ones, legacy block-array parser tests, and probe parser tests. go vet ./agent/kimi/... clean.
• Cross-checked against competing PRs #1566, #1586, #1601 (all targeting #1561).
• Verified backward compatibility on every dialect branch.

✅ What looks good:

• Probe-driven branching is the right architecture. Probe failure defaults to modern (safe, matches #1461's direction).
• Verbatim kimiCodeHelpV026 fixture in probe_test.go anchors the family discriminator to the real production binary — much stronger than synthetic help text.
• Format-tolerant parser (plain-string → block-array) keeps both flavors working without explicit flavor checks at every site.
• kimiSessionsBaseDirs cleanly extends session listing without changing legacy behavior. Workdir filter only applies when state.WorkDir != "", preserving legacy list-all semantics.
• parseKimiTranscript's "only fall back to wire.jsonl when context.jsonl msgCount == 0" is the right call.
• countWireJSONL correctly filters by type == "context.append_message" && origin.kind == "user" to avoid double-counting tool events and streamed assistant chunks.
• handleMeta's fallback to extracting the id from the legacy content field is a nice belt-and-suspenders move for forward compatibility.
• Race detector clean.

🚨/🔴 Must fix:
None — no correctness, security, or data-loss bugs found.

🟠 Should improve:

• Stacked on PR #1483 which is still OPEN. Base of this branch is e1ec734f, the --work-dir gate commit. Maintainer should either (a) merge #1483 first and ask author to rebase, or (b) merge this one and accept that --work-dir work goes in via #1564. Coordinate before approving.
• Multiple competing PRs target #1561. #1566, #1586, and #1601 are also open, with overlapping scope. #1601 (highland0971, 142 LOC) is a much smaller alternative that covers the core dialect-compat surface without the session-listing and quiet-mode extras. This PR is more comprehensive and has the strongest evidence (live e2e against real v0.26 binary), but the maintainer should pick one. Recommend a maintainer call before merging any.
• Unrelated docs(sponsors) commit 3fc360ee is bundled. Adds track_id query strings to Kimi platform links in README.md, README.zh-CN.md, and provider-presets.json. This is sponsor policy, not kimi-code dialect work — should be a separate PR for clean cherry-pick/bisect.
• parseKimiTranscript summary fallback edge case. If context.jsonl exists with msgCount > 0 but no user-text summary (e.g. user content was empty), wire.jsonl is not consulted and summary stays empty. Low practical risk, but worth a comment or test.

🔵 Optional:

• Plain-string handleAssistant doesn't extract thinking; if a tool_call follows, buffered text gets emitted as EventThinking by flushPendingAsThinking. Same pre-existing bug exists in the legacy block-array path, so not a regression — but a comment acknowledging the lack of a separate thinking channel would help future maintainers.
• isModernFlavor() could use an explicit test pinning a mid-migration case like kimiFlagSupport{Print: true, WorkDir: false} (workdir dropped before print) to lock in the discriminator intent.
• TestBuildArgs_WorkDirFlagGated asserts "/" doesn't appear in args — only meaningful because the workdir happens to be root. Tighten to a sentinel.

❓ Questions:

• What is the merge plan with #1483? Will maintainer merge that first and ask author to rebase?
• Will maintainer pick #1564 or #1601 to land #1561? (Or pick differently — e.g. split this PR into core-dialect + session-listing?)
• The new state.json filter behavior: legacy sessions are still listed across all workdirs (preserved), but modern sessions are now workDir-scoped (new). Worth confirming this asymmetry is intentional.

Testing / Risk:

• Verified evidence: go test -race ./agent/kimi/... -count=1 -timeout 90s — all 27 new tests pass, all pre-existing tests still pass. go vet clean. go build ./agent/kimi/... clean. (The web/embed.go build error is a pre-existing main-branch issue, not introduced here.)
• Author reports go test ./... green (Go 1.25.12). Other packages not affected: agent/kimi only imports core and the diff doesn't touch core.
• Live e2e (TestLiveE2E_KimiCodeFlavor) is env-gated and not in CI. Author reports it passes against ~/.kimi-code/bin/kimi v0.26.0 with session resume.
• Remaining risk: future kimi-code versions may add new event types or content shapes that this PR doesn't handle. Format-tolerant parser mitigates but doesn't eliminate this — recommend a follow-up test pinning the v0.29.1 (or current latest) verbatim wire output.

Next step:
Maintainer: pick a merge strategy for #1561 (merge #1483 → rebase #1564 → merge, or merge #1601 instead, or split this PR). Once decided, this PR is technically ready to land.

[chenhg5]

@whyihaveyou 你的 #1564 现在可以 rebase 了。

#1483 (commit e1ec734) 刚刚已合并进 main — 它原本就是 #1564 stack 的 base commit, 现在已经被官方收下了, 所以 rebase 到最新的 origin/main 即可:

git fetch origin
git checkout kimi-code-flavor
git rebase origin/main

rebase 之后 #1564 的 diff 只会剩下:

• 8dc7707c feat(kimi): native Kimi Code CLI dialect support
• efc779b4 fix(kimi): don't report 0 messages for Kimi Code sessions in /list
• a1da3ce1 Merge remote-tracking branch 'origin/main' into kimi-code-flavor
• 2ba2bfcf fix(kimi): adapt live e2e to main's Send(messageID, ...) signature
• 425ee176 fix(kimi): check f.Close error in transcript helpers (errcheck)

(原来 stack 里的 e1ec734 会自动从 history 里消失, 因为它已经在 main 上了。)

rebase 完成后 ping 我, 我重新 review 一遍就可以合并了。

— cc-connect/qa-claudecode

[chenhg5]

Conclusion: Approve

Overall assessment:

• Rebased and re-reviewed after whyihaveyou merged in OctopusWen's cross-PR contributions. The PR is now byte-clean against origin/main (the duplicate WorkDir half is gone — that work is already on main via #1483). The OctopusWen merge adds exactly the missing pieces my earlier review flagged: a CHANGELOG entry referencing #1561, and a regression test (TestHandleAssistantStringContentWithToolCalls) covering the string-content + tool_calls flush-order edge case that #1586 originally fixed. CI is 5/5 green, go test -race ./agent/kimi/... passes, go vet ./agent/kimi/... clean. The full Kimi Code CLI surface (#1561) is now covered end-to-end.

Review scope:

• Reviewed agent/kimi/kimi.go (+163/-74), agent/kimi/session.go (+108/-36), agent/kimi/probe.go (+16/-4), agent/kimi/session_test.go (+203/-5), agent/kimi/probe_test.go (+69/-0), agent/kimi/kimi_test.go (+115/-0), agent/kimi/live_e2e_test.go (+119/-0), CHANGELOG.md (+2/-1).
• Focused on: byte-clean diff vs main, the OctopusWen rebase (no #1483/#1476 leftovers), flushPendingAsThinking ordering under the new test, CHANGELOG accuracy.

✅ What looks good:

• Stack base fully rebased away. e1ec734f (dev-claudecode's #1476 work-dir gate) is no longer in the diff because main already has it via #1483 (6c860798). The diff is now purely about Kimi Code CLI dialect support.
• flushPendingAsThinking is called before the tool-use event in handleAssistant, and the new TestHandleAssistantStringContentWithToolCalls regression test pins that ordering — exactly the failure mode that motivated #1586 originally. If a future maintainer reorders or drops flushPendingAsThinking, this test breaks.
• CHANGELOG entry is honest: names #1561 explicitly, doesn't claim other PRs it didn't replace, and reads naturally next to the surrounding ### Fixed entries.
• live_e2e_test.go is gated on KIMI_LIVE_E2E=1 (verified by reading the file — no auto-run CI impact). It exercises the multi-turn resume path that was the actual user-reported failure mode behind #1561.
• Probe-driven flag detection (Print/WorkDir/Resume/Quiet) now consistent across all four — this matches the pattern in #1461 and #1483, and reviewers familiar with those will read this in 30 seconds.
• Dual session-storage roots (<projectRoot>/.kimi/sessions and <projectRoot>/.kimi_code/sessions) correctly added to the lookup list, which solves the cross-version storage migration path.

🚨/🔴 Must fix:

• No merge-blocking issues found. The PR is ready to merge.

🟠 Should improve:

• P2 — CHANGELOG entry could name OctopusWen + whyihaveyou + dev-claudecode as co-authors for credit (the entry currently just says "kimi-code dialect support"). Not blocking — the git log already credits the authors, and the CHANGELOG is meant to be terse.
• P2 — live_e2e_test.go is 119 LOC with one test function. If anyone adds more live e2e tests for other adapters, consider moving the shared test scaffolding (KIMI_BIN resolution, transcript recording, cleanup) into a testutil_live_test.go file. Not blocking for this PR.

🔵 Optional:

• P3 — The diff is +795/-120 across 8 files. That's a big diff, but it's well-decomposed: 5 production files + 3 test files + 1 changelog. The 8-file spread actually helps review because each file has a single coherent purpose.
• P3 — probe_test.go and session_test.go both have drainEvents helpers (similar shape). Could dedupe later if a third test file appears.

❓ Questions:

• Q1 — Does live_e2e_test.go need a docs entry explaining how to invoke it locally (e.g. KIMI_LIVE_E2E=1 go test ./agent/kimi/... -run TestKimiLiveE2E)? The convention in other adapters isn't uniform — some have it, some don't.

Testing / Risk:

• Verified locally: go test -race ./agent/kimi/... -count=1 -timeout 120s → PASS. go vet ./agent/kimi/... → clean. CI 5/5 SUCCESS at 80414d59.
• Backward compat: the legacyKimiHelp and modernKimiHelp fixtures both pass; the dual session-storage lookup is additive (legacy .kimi/sessions is still tried first; new .kimi_code/sessions is the fallback, or vice versa depending on which path kimi-code prefers — verified by reading the lookup order).
• Real-world validation: OctopusWen confirms live testing on kimi-code 0.36.1 (Feishu), meta resume_hint + -r resume both work end-to-end.

Next step:

• Maintainer: ready to merge. Cross-PR cleanup after merge: close #1586 (OctopusWen already received credit via the cross-PR contribution), close #1566 (subsumed), close #1601 (subsumed — highland0971 still owes an errcheck cleanup, can be a separate small PR if they want).
• Author (whyihaveyou / QZP / octopus): thanks for the cross-PR collaboration — this is the cleanest possible end state for a 4-way stacked-PR collision.

— cc-connect/qa-claudecode