MUL-3417: tolerate OpenClaw config file CLI mismatch

[multica-eve]

Summary

• keep openclaw config file as the preferred active-config discovery path when supported
• fall back to documented OpenClaw path env vars when that subcommand is rejected with the 2026.2-style command-shape errors
• preserve the original openclaw config file stderr if the env-var fallback also fails
• add regression coverage for execenv prep fallback, all fallback path sources, and the known unsupported-command matcher shapes

Fallback is intentionally limited to these known command-shape signals: too many arguments for 'config', Expected 0 arguments but got 1, or an unknown ... config ... file form. Other CLI errors still fail closed.

Closes MUL-3417
Fixes #4299

Testing

• go test ./internal/daemon/execenv

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
multica-docs	Ready	Preview, Comment	Jun 18, 2026 9:11am

[YOMXXX]

I investigated the backend CI failure on this PR. The failing tests are unrelated to the OpenClaw diff itself:

• server/internal/handler / TestMembershipCache_InvalidatedOnDeleteMember: setup: expected cache hit after Set
• server/internal/middleware / TestDaemonAuth_DaemonTokenCacheHit: expected cache hit but got 401

Root cause looks like Redis test isolation: the Redis-backed test helpers in multiple packages all use the same REDIS_TEST_URL DB and call FlushDB at setup/cleanup. Since go test ./... runs packages in parallel, one package can delete another package's freshly-set cache key. That matches both failures: Set succeeds, then a package-parallel FlushDB turns the immediate Get into a miss.

I could not push directly to multica-ai:agent/gpt-boy/e1bcbc5d (403 Permission to multica-ai/multica.git denied to YOMXXX), so I pushed a patch branch with the exact fix candidate:

• YOMXXX:codex/fix-4302-redis-test-isolation
• commit d263e2862 (test: isolate redis-backed package tests)

Patch shape: each Redis-backed test package selects a package-dedicated Redis DB before FlushDB (auth=10, middleware=11, handler=12, service=13), preserving per-test clean DB semantics without cross-package deletion.

Local validation:

• cd server && go test ./internal/daemon/execenv -run 'Openclaw|PrepareEnvironmentOpenclaw' -count=1\n- cd server && go test ./internal/auth ./internal/middleware ./internal/service -count=1\n- cd server && go test ./internal/handler -run 'MembershipCache|Redis(LocalSkill|Update|Model|Liveness)' -count=1\n- git diff --check\n\nNote: Docker/Redis/Postgres are not running on my local machine, so Redis/DB-backed tests that require those services are skipped locally per the existing test helpers; the CI failure itself was diagnosed from the failed backend job logs.

[YOMXXX]

Follow-up: I also opened the Redis test-isolation patch as standalone PR #4348 so it can be reviewed/merged independently of this OpenClaw PR: #4348

If #4348 lands first, rebasing/rerunning this PR should avoid the package-parallel Redis FlushDB cache miss seen in the current backend job.

[YOMXXX]

#4348 is now green (frontend, backend, and both installer checks passed; mergeStateStatus=CLEAN). That gives a clean path to unblock this PR without mixing the Redis test-infra change into the OpenClaw branch manually.