[Bug] cron-nested lane mis-resolves hyphen-prefixed provider aliases on 2026.4.29

Summary

The cron-nested lane's model resolver mis-handles aliases of the form <provider-with-hyphen>/<model> (e.g., kimi-code/kimi-for-coding), producing a stripped/different alias and reporting Unknown model. The same alias resolves correctly in the agent/embedded lane. This blocks any cron from using a provider whose name contains a hyphen.

Environment

• OpenClaw CLI: 2026.4.29 (a448042)
• Node.js: v22.22.2
• macOS: Darwin 25.3.0 (Mac Mini, Apple Silicon)
• Service shape: native LaunchAgent ai.openclaw.gateway via gateway install --force --wrapper
• Provider in question: kimi-code (Kimi For Coding subscription endpoint, https://api.kimi.com/coding/v1, api: anthropic-messages)

Repro

In ~/.openclaw/agents/main/agent/models.json, a provider kimi-code exists with a model entry:

{
  "id": "kimi-code/kimi-for-coding",
  "name": "kimi-for-coding",
  "api": "anthropic-messages",
  "reasoning": false,
  ...
}

In ~/.openclaw/openclaw.json:

{
  "agents": {
    "defaults": {
      "model":     { "primary": "kimi-code/kimi-for-coding", "fallbacks": [...] },
      "subagents": { "model": { "primary": "kimi-code/kimi-for-coding" } }
    }
  },
  "plugins": { "entries": { "active-memory": { "config": { "model": "kimi-code/kimi-for-coding" } } } }
}

Each cron job's payload.model set to "kimi-code/kimi-for-coding".

Expected behavior

kimi-code/kimi-for-coding resolves consistently across all dispatch lanes. Verified working on the agent/embedded lane:

2026-05-01T19:02:52.004 [agent/embedded] phase=attempt-dispatch totalMs=14872
  model-resolution=7030ms  attempt-dispatch=3285ms

Observed behavior

When a cron fires (lane=cron-nested), the resolver requests a stripped alias kimi/kimi-for-coding (note: the -code suffix on the provider name has been removed):

[diagnostic] lane task error: lane=cron-nested durationMs=23445
  error="FailoverError: Unknown model: kimi/kimi-for-coding"

[model-fallback/decision] decision=candidate_failed
  requested=kimi/kimi-for-coding candidate=kimi/kimi-for-coding
  reason=model_not_found next=openai-codex/gpt-5.5
  detail=Unknown model: kimi/kimi-for-coding

The string kimi/kimi-for-coding does not appear anywhere in openclaw.json, cron/jobs.json, or models.json at the time of the failure. It is being constructed by the resolver itself when it processes the kimi-code/kimi-for-coding alias.

Workarounds attempted (all failed)

1. Renamed provider kimi-code → kimicode (no hyphen) in models.json + updated all references. The cron-nested lane still produced kimi/kimi-for-coding requests. So the bug is not strictly hyphen-related — there's a deeper alias-normalization step in cron-nested that doesn't honor the modern alias.

2. Cleared payload.model from all crons so they inherit agents.defaults.model.primary. Crons still mis-resolve because they go through cron-nested.

3. Added agents.defaults.models["kimi-code/kimi-for-coding"] registry entry. No effect on cron-nested resolution.

4. Added a model entry with id: "kimi/kimi-for-coding" to the legacy kimi provider. The custom entry was stripped on next gateway bootstrap (related to issue #73607 — agents.defaults.models[].params and similar custom additions are wiped on restart). So I couldn't even register the alias the cron-nested lane was producing.

What works

Reverting cron payload.model to the legacy alias kimi/kimi-code (the only alias registered on the legacy kimi provider — its model entry has id: "kimi-code"). Cron-nested can resolve this. Slow (reasoning: true on legacy entry → 100-150 s first-token), but functional.

Companion: direct API verification

The Kimi For Coding endpoint itself is fast when called correctly. From the same machine:

curl -H "Authorization: Bearer $KEY" \
     -H "User-Agent: KimiCLI/1.5" \
     -H "Content-Type: application/json" \
     -d '{"model":"kimi-for-coding","messages":[{"role":"user","content":"hi"}],"max_tokens":20,"thinking":{"type":"disabled"}}' \
     https://api.kimi.com/coding/v1/chat/completions

Returns in 1.3 seconds. The K2.6 backend is fast when thinking: disabled is sent.

Asks

1. Make the cron-nested lane's model resolver use the same alias-resolution path as agent/embedded, so both lanes accept kimi-code/kimi-for-coding.
2. Stop stripping custom additions to models.json providers/models on bootstrap (related to #73607).
3. Document the canonical alias format expected for hyphen-prefixed providers — if there's a sanctioned shape, surface it in the model-providers docs.

Sources / supporting context

• Kimi K2.6 quickstart (official UA + thinking param requirement): https://platform.kimi.ai/docs/guide/kimi-k2-6-quickstart
• Kimi-cli reference UA: https://github.com/MoonshotAI/kimi-cli (KimiCLI/1.5)
• Related upstream/sibling issues:
  • https://github.com/openclaw/openclaw/issues/73607 (params stripping on 2026.4.26+)
  • https://github.com/openclaw/openclaw/issues/61270 (Kimi Coding tool calls regression, 2026.4.2+)
  • https://github.com/badlogic/pi-mono/issues/3585 (built-in kimi-coding provider 429 due to missing UA — pi-mono context)
  • https://github.com/NousResearch/hermes-agent/issues/5796 (same UA requirement, hermes-agent context)