[diegosouzapw]

Closes #14575

Root cause

GitHub Copilot's new claude-opus-5.5 id is absent from OmniRoute's curated GitHub Copilot
catalog. GithubExecutor.buildUrl() already routes any claude-named model (registered or
not) to Copilot's Anthropic-native /v1/messages endpoint via an unconditional /claude/i name
match. But getModelTargetFormat("gh", model) only recognized registered claude ids, so an
untagged id like claude-opus-5.5 fell through to the provider's base openai format. The
request body stayed OpenAI-shaped (tool_choice as a string, tools as
{type:"function",...}) while being dispatched to the Anthropic-native endpoint, which rejects
it with tool_choice: Input should be an object and tools.0: Input tag 'function' found....

Fix

open-sse/config/providerModels.ts::getModelTargetFormat(): after the curated-catalog lookup
misses, added a heuristic scoped to the gh/ghe-copilot aliases — any model id containing
claude (case-insensitive) resolves targetFormat: "claude", mirroring the routing decision
buildUrl() already makes unconditionally. This is the same pattern already shipped for the
openai "-pro" heuristic (#5842). Curated catalog entries still return before this heuristic
is reached, so no existing routing changes for registered models.

Regression test

tests/unit/github-copilot-claude-untagged-model-14575.test.ts (new):

• RED (pre-fix): resolveChatCoreTargetFormat({provider:"github", resolvedModel:"claude-opus-5.5", sourceFormat:"openai", ...}).targetFormat resolved to "openai" while GithubExecutor.buildUrl("claude-opus-5.5", ...) returned the Anthropic-native messagesUrl — assertion 'openai' !== 'claude' failed.
• GREEN (post-fix): targetFormat resolves to "claude", matching the endpoint buildUrl() already routes to. Also added a control test for a curated model (no regression) and a parity test for the ghe-copilot alias.

Existing tests

Two pre-existing tests encoded the OLD (buggy) premise that an unregistered claude-named id on
gh would stay on the /chat/completions path (getModelTargetFormat("gh", ...) === null).
With this fix that premise no longer holds — any claude-named id is now native, matching
buildUrl()'s own routing — so those tests were realigned to the corrected behavior (not
weakened; assertions changed to match what is now the correct contract):

• tests/unit/t27-github-copilot-response-format.test.ts — "T27: Claude + response_format=json_object injects system instruction..." renamed and now asserts response_format passes through untouched for an untagged claude id (native path skips the /chat/completions-only workaround), consistent with the already-existing github-copilot-claude-native-messages.test.ts coverage for registered ids.
• tests/unit/executor-github.test.ts — two tests updated similarly: (1) reasoning-field stripping still applies for claude models, but response_format is no longer injected as a system message (native path); (2) the Anthropic-shape content-part sanitization test now uses a non-claude model id (gpt-4o) to exercise the /chat/completions-only sanitization path, since that path is now unreachable for any claude-named id.

All three files pass after the change. Also ran and confirmed green (no regressions):
provider-registry-github-copilot-targetformat.test.ts, ghe-copilot-targetformat-parity.test.ts,
github-copilot-custom-model-target-format.test.ts, openai-responses-only-models-5842.test.ts,
provider-models-target-format-scoping.test.ts, copilot-claude-always-v1-messages.test.ts,
github-copilot-claude-native-messages.test.ts, chatcore-target-format.test.ts,
custom-model-target-format.test.ts, provider-models-config.test.ts.

Gates run

• node --import tsx/esm --test tests/unit/github-copilot-claude-untagged-model-14575.test.ts — RED then GREEN (4/4 pass post-fix)
• npm run typecheck:core — 1 pre-existing error in src/lib/services/cliproxyAccountHealth.ts (untouched by this PR; already fixed on the current origin/release/v3.8.51 tip past this branch's base commit beec76d5 — inherited, not introduced by this diff)
• node scripts/check/check-open-sse-typecheck.mjs — 2 pre-existing errors in open-sse/executors/auggie.ts (untouched by this PR; same situation — already fixed on the current tip past this branch's base)
• npx eslint --suppressions-location config/quality/eslint-suppressions.json open-sse/config/providerModels.ts tests/unit/github-copilot-claude-untagged-model-14575.test.ts — clean
• node scripts/check/check-file-size.mjs — no violations on files this diff touches (10 pre-existing violations on unrelated frozen files)
• node scripts/check/check-complexity-ratchets.mjs --base-ref origin/release/v3.8.51 — OK (0 changed source files in ratchet scope)
• node scripts/check/check-mutation-test-coverage.mjs --strict — pre-existing drift on 9 unrelated modules (none touched by this diff; providerModels.ts is not a stryker-mutated module)
• node scripts/check/check-changelog-integrity.mjs — OK
• npm run check:provider-consistency — OK (276 registry entries, 0 registry-only exceptions)

⚠️ base-red inherited: the cliproxyAccountHealth.ts and auggie.ts typecheck errors above are
pre-existing on this branch's cut point (beec76d5) and already fixed on the current
origin/release/v3.8.51 tip by later, unrelated commits — not introduced by this diff.

Plan-file: _tasks/pipeline/bugs/2-implementing/14575-fix-github-copilot-tool-choice-error-in-claude-opus-5-5-on-git.plan.md