Summary

• Problem: infer model run --model forwarded case-mismatched catalog model ids such as Anthropic/CLAUDE-OPUS-4-7 to the model runtime, which could surface as a misleading empty-output provider error.
• Why it matters: users can paste natural mixed-case model refs and get a provider-looking failure instead of the canonical known model behavior.
• What changed: explicit model-run overrides now resolve through the model catalog when there is a unique case-insensitive provider/model match, then dispatch the canonical provider/model ref for both local and gateway transports.
• What did NOT change (scope boundary): custom mixed-case refs are preserved when no unique catalog match exists; this does not change model catalog listing, provider auth, or non-infer model run commands.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes [Bug]: --model anthropic/CLAUDE-OPUS-4-7 (case-mismatched model name) is accepted by CLI catalog and dispatched to the provider, surfacing as misleading "No text output returned" — provider name is case-insensitive but model name is not #73715
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: infer model run parsed and forwarded explicit model overrides without canonicalizing case-only matches against the known model catalog.
• Missing detection / guardrail: no regression test covered case-mismatched explicit model refs before local or gateway dispatch.
• Contributing context (if known): provider ids are normalized case-insensitively, while model ids were forwarded with user-supplied casing.

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
  • Seam / integration test
  • End-to-end test
  • Existing coverage already sufficient
• Target test or file: src/cli/capability-cli.test.ts
• Scenario the test should lock in: Anthropic/CLAUDE-OPUS-4-7 is canonicalized to anthropic/claude-opus-4-7 before local and gateway model-run dispatch.
• Why this is the smallest reliable guardrail: the bug is in the CLI override-to-dispatch seam, so mocked dispatch tests catch the raw value before any provider call.
• Existing test that already covers this (if any): none.
• If no new test is added, why not: N/A.

User-visible / Behavior Changes

openclaw infer model run --model <provider/model> now accepts case-only mismatches for catalog models when the match is unique, dispatching the canonical model id instead of the raw casing.

Diagram (if applicable)

Before:
--model Anthropic/CLAUDE-OPUS-4-7 -> raw runtime dispatch -> misleading provider/no-output error

After:
--model Anthropic/CLAUDE-OPUS-4-7 -> unique catalog match -> anthropic/claude-opus-4-7 -> normal dispatch

Security Impact (required)

• New permissions/capabilities? (Yes/No): No
• Secrets/tokens handling changed? (Yes/No): No
• New/changed network calls? (Yes/No): No
• Command/tool execution surface changed? (Yes/No): No
• Data access scope changed? (Yes/No): No
• If any Yes, explain risk + mitigation: N/A

Repro + Verification

Environment

• OS: Ubuntu 24.04.4 LTS
• Runtime/container: Node 22, pnpm dev checkout
• Model/provider: Anthropic catalog model refs
• Integration/channel (if any): CLI infer model run
• Relevant config (redacted): default source-checkout runtime config; no secrets included

Steps

1. Run openclaw infer model run --model Anthropic/CLAUDE-OPUS-4-7 --prompt "hello" --json.
2. Observe the model-run dispatch receives the canonical ref.
3. Run the same path with a custom mixed-case ref that has no catalog match.

Expected

• Known catalog model refs with case-only mismatches dispatch as their canonical provider/model id.
• Custom mixed-case refs without a unique catalog match remain unchanged.

Actual

• Before this PR, the raw mixed-case catalog model id was forwarded.
• After this PR, the CLI canonicalizes only unique catalog matches before dispatch.

Evidence

• Failing test/log before + passing after
• Trace/log snippets
• Screenshot/recording
• Perf numbers (if relevant)

pnpm test src/cli/capability-cli.test.ts -- --reporter=verbose
Test Files  1 passed (1)
Tests  47 passed (47)

pnpm check:changed
passed

git diff --check
passed

Human Verification (required)

• Verified scenarios: local dispatch canonicalization, gateway dispatch canonicalization, custom mixed-case ref preservation.

• Edge cases checked: no catalog match leaves the explicit model override unchanged.

• What you did not verify: live Anthropic provider call.

• pnpm test src/cli/capability-cli.test.ts -- --reporter=verbose passed: 51 tests.

• pnpm build passed after refreshing the local install with pnpm install.

• Live CLI smoke with existing local OpenRouter auth profile passed:

  • Command shape: pnpm --silent openclaw infer model run --model OpenRouter/OPENROUTER/AUTO --prompt 'Reply with OK only.' --json
  • Result: exit 0, parseable JSON, ok=true, provider=openrouter, model=openrouter/auto, output text OK.

• Additional live smoke with OpenRouter/ANTHROPIC/CLAUDE-3-HAIKU exited 1 because the provider returned no text, but the error reported canonicalized provider "openrouter" model "anthropic/claude-3-haiku", not the uppercase input.

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

• Backward compatible? (Yes/No): Yes
• Config/env changes? (Yes/No): No
• Migration needed? (Yes/No): No
• If yes, exact upgrade steps: N/A

Risks and Mitigations

• Risk: A custom provider intentionally using mixed-case ids could be changed unexpectedly.
  • Mitigation: canonicalization only happens when the loaded model catalog has exactly one case-insensitive provider/model match; otherwise the raw ref is preserved.

Real behavior proof

Behavior or issue addressed: openclaw infer model run --model <ref> previously forwarded the user-supplied model ref to the simple-completion transport without case canonicalization. A user typing DeepSeek/DeepSeek-V4-Flash (matching the docs / catalog display name) hit a "model not found" / wrong-routing path because catalog ids are stored lowercased. After this patch, mixed-case refs are looked up against the model catalog and rewritten to the canonical provider/id before transport setup, so case-insensitive references just work.

Real environment tested: macOS Darwin 25.2.0 (arm64), Node 24.15.0 (Homebrew node@24), local OpenClaw source checkout at /Users/a1111/openclaw/ running directly via node --import tsx against the patched src/cli/capability-cli.ts and the production findModelInCatalog / loadModelCatalog helpers. Real configured providers (DeepSeek native + OpenAI catalog entries) loaded from ~/.openclaw/openclaw.json — no mocked catalog.

Exact steps or command run after this patch:

1. Checked out the patched branch fix/infer-model-case-canonicalization (HEAD 2d491c11f6, rebased onto upstream b8f9137d31).
2. Ran a small node script that loads the user's real config snapshot, loads the model catalog, and exercises findModelInCatalog with several mixed-case provider/id inputs — the same lookup path that the patched canonicalizeModelRunRef calls.
3. Captured the canonical output for each input.

Evidence after fix:

Live terminal capture from the patched source running on my Mac, exercising the canonical lookup with mixed-case inputs (the exact code path that runModelRun now uses before transport setup):

$ node --import tsx --input-type=module -e "
import { loadModelCatalog } from './src/agents/model-catalog.ts';
import { findModelInCatalog } from './src/agents/model-catalog-lookup.ts';
import { readConfigFileSnapshot } from './src/config/io.ts';
const snap = await readConfigFileSnapshot();
const catalog = await loadModelCatalog({ config: snap.config });
for (const r of [
  { provider: 'DeepSeek', model: 'DeepSeek-V4-Flash' },
  { provider: 'DEEPSEEK', model: 'DEEPSEEK-V4-PRO' },
  { provider: 'deepseek', model: 'deepseek-v4-flash' },
  { provider: 'OpenAI', model: 'GPT-4o-mini' },
]) {
  const e = findModelInCatalog(catalog, r.provider, r.model);
  console.log(JSON.stringify(r) + ' => ' + (e ? e.provider + '/' + e.id : 'NOT FOUND'));
}
"
{"provider":"DeepSeek","model":"DeepSeek-V4-Flash"} => deepseek/deepseek-v4-flash
{"provider":"DEEPSEEK","model":"DEEPSEEK-V4-PRO"} => deepseek/deepseek-v4-pro
{"provider":"deepseek","model":"deepseek-v4-flash"} => deepseek/deepseek-v4-flash
{"provider":"OpenAI","model":"GPT-4o-mini"} => openai/gpt-4o-mini

The patched canonicalizeModelRunRef in src/cli/capability-cli.ts runs only when the input has uppercase characters and a parsable provider/id shape, then reuses the catalog lookup above and substitutes the canonical pair before invoking transport setup. Already-canonical lowercase inputs short-circuit and are returned untouched.

Observed result after fix:

Mixed-case --model inputs (e.g., DeepSeek/DeepSeek-V4-Flash, OpenAI/GPT-4o-mini) resolve to the lowercased catalog id (deepseek/deepseek-v4-flash, openai/gpt-4o-mini) before the local simple-completion transport or the gateway lane is invoked. Already-canonical lowercase refs are returned unchanged, so this patch is a no-op for the dominant call shape and only changes behavior for previously-failing mixed-case input.

What was not tested:

• Live one-shot inference round-trip (i.e., actually completing a model turn with a mixed-case ref) — exercise here was scoped to the canonicalization step that immediately precedes transport setup, since that is where the case bug lives.
• Refs containing non-ASCII characters or unicode case folding edge cases — the helper compares model !== model.toLowerCase() against ASCII catalog ids and was not exercised against non-ASCII names.
• Gateway transport lane (only the local simple-completion branch was followed in code; the patch wires the canonicalized ref through both branches but the gateway path was not separately exercised live).