feat(virtualmodels): disable_reasoning flag for redirects (Kimi Code K2.6 via virtual models)

[weselben]

Prototype C of 4 for #27 — a virtual-models based approach. Compare: prototype A (provider-level static map + KIMICODE_MODEL_MAP=off toggle, PR #30), prototype B (provider-level map, always-on, PR #29), prototype D (alternative architecture).

Approach

A new disable_reasoning flag on virtual model redirects lets an operator make any redirect serve the non-thinking variant of its target:

virtual_models:
  - source: kimi-k2.6
    target: kimicode/kimi-for-coding
    disable_reasoning: true
  - source: kimi-k2.7-code
    target: kimicode/kimi-for-coding
  - source: kimi-k2.7-code-highspeed
    target: kimicode/kimi-for-coding-highspeed
  - source: kimi-k3
    target: kimicode/k3
  - source: kimi-k3-256k
    target: kimicode/k3-256k

When a request hits a flagged redirect, rewriteChatRequest strips the typed Reasoning field and the flat reasoning_effort extra, then sets thinking.type=disabled on the forwarded copy — so Kimi Code routes the request to the K2.6 behavior of the same kimi-for-coding endpoint. Providers without a thinking toggle ignore the unknown extra (Postel's law).

What's wired

• config.VirtualModelConfig.DisableReasoning bool (yaml + json)
• virtualmodels.VirtualModel.DisableReasoning bool + View.DisableReasoning bool
• virtualmodels.Service.DisableReasoningForSource(source string) bool (lookup by source)
• virtualmodels.Service.ResolveRedirectForUserPath (returns full Resolution so request rewrites can consult the matched redirect's flags)
• rewriteChatRequest applies the flag; signature of the two internal resolve helpers now returns the matched source too (batch preparer updated for the new signature)
• SQL store: new disable_reasoning boolean column with a backfill migration (mirrors the slowdown pattern)
• Mongo store: bson field, round-tripped in upsert/decode
• Admin View carries the flag

Trade-offs vs A/B

Aspect	A/B (provider-level map)	C (this)
Default-on	yes	operator-only — must declare the aliases
Metadata for canonicals	yes (auto-resolved from registry bare entries)	no — pricing/modes must come from a registered upstream model; if the target isn't in the registry, metadata is missing
Thinking-pin per alias	yes	yes (this flag)
K2.6 routing	yes (canonical kimi-k2.6 → thinking off)	yes (disable_reasoning: true redirect to kimicode/kimi-for-coding)
Operator-editable without code	no	yes
Reuses existing primitives	new provider code	existing virtual models
Token/cost tracking	records client-requested canonical name	records the post-redirect resolved target — see caveat below

Caveat: usage recording

Today internal/gateway/usage.go records the resolved target model (kimicode/kimi-for-coding), not the requested virtual-model source. For the K2.6 case this means usage rows will show kimi-for-coding rather than kimi-k2.6, which is misleading. A small follow-up — adding a precedence slot for the original request model when a virtual-model alias rewrote it — would close this gap. The PR documents the divergence; fixing it is out of scope here so this stays a minimal, reviewable diff.

Tests

• go build ./... — pass
• go test ./internal/virtualmodels/... ./config/... — pass (4 new unit tests + carried-over suite):
  • TestRewriteChatRequest_DisableReasoningStripsControls — flag strips Reasoning + reasoning_effort, sets thinking.type=disabled, preserves other extras, does not mutate the caller's request
  • TestRewriteChatRequest_DisableReasoningUnsetKeepsControls — unflagged redirect leaves reasoning alone
  • TestService_DisableReasoningForSource — flag lookup by source, missing/empty handled
  • TestConfigModel_MapsDisableReasoning — config → service mapping round-trips the flag and marks Managed

Refs #27

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}