fix(kimi): gate --work-dir flag on probe, like --print (#1476)

[chenhg5]

Fixes #1476

问题

Kimi Code CLI builds newer than what #1456 covered dropped the --work-dir
flag. Passing it now produces:

error: unknown option --work-dir

The reporter (#1476, v1.4.1, Kimi Code agent) hits this on every spawn
because session.go:97-99 always appended the flag whenever a non-default
workDir was set. The #1461 / #1456 fix gated --print the same way; this
PR closes the gap for --work-dir.

改动

Follows the exact same probe-based pattern as #1456 / PR #1461 (referenced
for both users and reviewers who already validated that approach).

agent/kimi/probe.go

• kimiFlagSupport gains WorkDir bool.
• probeKimiFlags populates it from the existing parseKimiHelpFlags map:

  support := kimiFlagSupport{
      Print:   flags["--print"],
      WorkDir: flags["--work-dir"],
  }

• Doc comment extended to reference both [Bug] ❌ 错误: error: unknown option '--print' (Did you mean --prompt?) #1456 and [Bug] ❌ 错误: error: unknown option '--work-dir' #1476.

agent/kimi/session.go

• One-line gate:

  if ks.workDir != "" && ks.flagSupport.WorkDir {
      args = append(args, "--work-dir", ks.workDir)
  }

• The agent still runs in the correct directory because exec.Command.Dir
  is set on the spawned *exec.Cmd (downstream of this line) — same as in
  the fix(kimi): conditionally pass --print so newer Kimi Code CLI works (#1456) #1461 fix.

agent/kimi/probe_test.go

• TestParseKimiHelpFlags_LegacyAdvertisesPrint now also asserts
  flags["--work-dir"] is true (legacy CLI supports it).
• New constant modernKimiHelpWithoutWorkDir mirrors the Kimi Code CLI
  build the reporter hits (no --print, no --work-dir).
• New TestParseKimiHelpFlags_ModernWithoutWorkDir asserts neither flag
  is detected on the deeper-no-work-dir surface.
• Existing TestParseKimiHelpFlags_ModernHidesPrint left intact (it
  covers the in-between surface where --print is gone but --work-dir
  still exists — useful regression coverage).
• New TestKimiFlagSupport_LegacyHelpSetsWorkDir +
  TestKimiFlagSupport_ModernWithoutWorkDir exercise the full probe-
  mapping, not just parseKimiHelpFlags, since that's what
  buildArgs actually consumes.

agent/kimi/session_test.go

• TestBuildArgs_WorkDirFlagGated — workDir=/, flagSupport{WorkDir:false}
  → --work-dir MUST NOT appear, AND the value / MUST NOT leak into args
  (catches a partial-gate bug where only the flag name is dropped).
• TestBuildArgs_WorkDirFlagEmitted — workDir=/, flagSupport{WorkDir:true}
  → --work-dir / MUST appear (legacy CLI keeps non-default dir support).

触发条件 / trade-off

• Why reuse parseKimiHelpFlags instead of adding a new parser?
  --work-dir already appears in the help-text of both legacy and the
  in-between build the existing modernKimiHelp constant models. The
  existing parser picks it up the same way it picked up --print — no
  new parser logic, no new layout assumptions.
• Why a separate modernKimiHelpWithoutWorkDir instead of editing
  modernKimiHelp? The modernKimiHelp constant is anchored in the
  [Bug] ❌ 错误: error: unknown option '--print' (Did you mean --prompt?) #1456 fix history; editing it would silently invalidate that test's
  assertion contract. Two-realistic-surfaces is the cleaner modeling.
• Risk for users on default dirs? None — cmd.Dir is set
  unconditionally, so whether the CLI sees --work-dir $X or just takes
  cwd from cmd.Dir, the agent runs in the same directory. The only
  practical difference is users on non-default dirs whose modern CLI
  previously crashed: they now get clean startup (cwd fallback).
• Risk for users with legacy kimi-cli? None — they advertise
  --work-dir in help, the probe sees it, buildArgs still emits it.

验证

• gofmt -l agent/kimi/ — clean
• go vet ./agent/kimi/... — clean
• go test -count=1 ./agent/kimi/... — all pass (4 prior + 6 new =
  10 tests in the probe/session suites)

不做什么

• 不改 Kimi Code CLI 本身
• 不改其他 agent adapter (Claude / Codex / Gemini / Cursor / OpenCode
  etc. each have their own flag handling; out of scope)
• 不改 probe 探测逻辑本身 (沿用现有 kimi --help 解析 + 短超时
  fallback to modern surface on probe failure)
• 不重构 buildArgs / kimiFlagSupport — 仅加一个 bool 字段

关联

• 同 template 参考: PR fix(kimi): conditionally pass --print so newer Kimi Code CLI works (#1456) #1461 (--print gating, [Bug] ❌ 错误: error: unknown option '--print' (Did you mean --prompt?) #1456)
• 同 reporter: issue [Feature] QQ OneBot: add group_reply_all option (parity with Feishu, Discord, Telegram) #1475 (group_reply_all for QQ platform) 是独立
  PR，本次不合并。

🤖 Generated with Claude Code

[chenhg5]

Conclusion: Comment

Overall assessment:

• Solid, minimal fix that closes the --work-dir gap for the Kimi Code CLI build (#1476) using the exact same probe-gating pattern already validated by PR #1461 / #1456 for --print. Single commit on origin/main, CI fully green, local go test -race ./agent/kimi/... passes 10/10. Mergeable as-is.

Review scope:

• Reviewed the full diff (5 files, +124 / -6): agent/kimi/probe.go, agent/kimi/session.go, agent/kimi/probe_test.go, agent/kimi/session_test.go, CHANGELOG.md.
• Focused on correctness (the gate and the probe), test coverage (both supported and unsupported CLI surfaces), backward compatibility, and stacking vs PR #1564.

✅ What looks good:

• Pattern consistency with #1461: if ks.workDir != "" && ks.flagSupport.WorkDir { ... } mirrors the existing --print gate at session.go:79-81 exactly. Reviewers who already approved #1461 can verify this in 30 seconds.
• exec.Command.Dir is set unconditionally at session.go:187, so omitting --work-dir on modern CLI is functionally equivalent for users on default dirs and a graceful cwd fallback for users on non-default dirs. Author's claim is verified by reading the code path.
• Two-realistic-surfaces approach in tests: kept the existing modernKimiHelp (transition build: no --print, still has --work-dir) and added a separate modernKimiHelpWithoutWorkDir fixture for the #1476 surface. This preserves the #1456 regression contract instead of mutating an anchored fixture.
• Partial-gate future bug catch: TestBuildArgs_WorkDirFlagGated asserts the value / doesn't leak into args when the gate is closed. A future refactor that drops the flag-name but keeps the value would be caught.
• Probe-failure contract preserved: zero-value kimiFlagSupport{} on probe failure remains the conservative "modern CLI" default — unchanged from #1461.
• No scope creep: only agent/kimi/ is touched (plus one CHANGELOG line); no other agent adapters, no platform, no core.
• CHANGELOG entry is informative: names the issue, references prior art, explains the cwd fallback, gives one-line legacy/modern outcomes.

🚨/🔴 Must fix:

• No merge-blocking issues found.

🟠 Should improve:

• None.

🔵 Optional:

• Add a buildArgs-level test for the in-between surface (Print=false, WorkDir=true). The parser is tested at that surface via TestParseKimiHelpFlags_ModernHidesPrint, but a TestBuildArgs_ModernWithWorkDir (or extending TestBuildArgs_PlanMode) would explicitly assert the gate fires when WorkDir=true while Print=false. Not needed for correctness today, but a 6-line addition that locks in the integration. Follow-up only.

❓ Questions:

• Maintainer decision (not for the author): this PR is the base commit of PR #1564 (feat(kimi): native Kimi Code CLI dialect support). #1564's first commit (e1ec734f) is byte-identical to this PR's HEAD, and #1564's second commit (8dc7707c) likely consumes the kimiFlagSupport.WorkDir field added here — meaning #1564 will not compile standalone without #1483 merged first. Recommendation: merge #1483 first, then ask whyihaveyou (#1564 author, first-time contributor) to rebase #1564 onto the new main so its diff reflects only the actual Kimi Code dialect work, not the stacked base.

Testing / Risk:

• Verified locally:
  • go test -race ./agent/kimi/... -count=1 -timeout 60s → ALL PASS (10/10 tests; 6 new in this PR).
  • go vet ./agent/kimi/... clean.
  • gofmt -l agent/kimi/ clean.
• Verified parser correctness manually against both fixtures: legacyKimiHelp → --work-dir:true, modernKimiHelpWithoutWorkDir → --work-dir:absent.
• CI: lint, unit-test, smoke-test, regression-test, performance-test all SUCCESS at e1ec734f (runs/28550372464).
• Remaining risk: if a future Kimi CLI build keeps --work-dir but moves it to a sub-command or aliases it differently, the parser might miss it. The existing two-realistic-surfaces fixtures cover both "in-between" and "fully removed" transitions; future moves would need a new fixture.
• Backward compatibility: zero regression risk — legacy users keep --work-dir, modern users get clean startup, in-between users keep --work-dir (parser still detects it).

Next step:

• For the maintainer: merge #1483 first to unblock #1564. Then ask whyihaveyou to rebase #1564 onto the new main for a cleaner review surface.
• For the author (chenhg5, self-PR): nothing required. The P3 follow-up (in-between-surface buildArgs test) can be a separate PR if desired.