fix: opencode zen gateway compat (muse-spark 1.3 via opencode-go sub)

[Blankeos]

Makes muse-spark-1.3-contributor usable on crabcode through the opencode-go subscription gateway (https://opencode.ai/zen/go/v1). Three request-path bugs, all gateway-specific — no behavior change for first-party providers (OpenAI/Anthropic/xAI mint unique ids and versionless base URLs, so the new code paths are no-ops for them).

What broke

Session xcx605ghifzhcdkqvil5wwwy died on step 1 with:

Duplicate function_call_output for call_id 'call_1'. Each function_call must have exactly one matching function_call_output

plus two latent gateway issues found while pairing on the session: a /v1/v1/responses 404 and a hollow catalog entry from the Failed to deserialize catalog extension model {provider_id}/{model_id} log line.

Changes

• Responses duplicate call_ids (src/aisdk/providers/openai.rs): the zen relay mints sequential per-request ids (call_1, call_2, … restarting every request), so after two tool turns the replayed history repeats ids and the Responses validator 400s. New dedupe_responses_call_ids pass in build_openai_messages (single choke point for HTTP + websocket bodies) rewrites repeats to {id}_dupN on both the function_call and its matching output (occurrence-indexed pairing), dropping stale response item.ids on rewritten calls. First occurrence untouched.
• Duplicated version segment (src/aisdk/providers/{mod,openai,anthropic,compatible}.rs): shared base_url_has_version_segment helper; default Responses path becomes /responses (not /v1/responses) and Anthropic /messages (not /v1/messages) when the base URL already carries /vN. compatible.rs now reuses the helper instead of its local copy.
• Catalog extension patch-fragments (src/model/extensions/mod.rs): a {"attachment": true}-style fragment for a model absent from models.dev no longer fabricates a hollow entry with wrong defaults (tool_call=false, no limits/cost); it's skipped with a descriptive log (also fixes the unformatted {provider_id}/{model_id} log line) until models.dev carries the model or the spec is completed.

Verification

• New tests: dedupes_repeated_call_ids_across_turns_for_responses_input, default_responses_path_does_not_duplicate_v1_in_base_url, messages_url_does_not_duplicate_version_segment, detects_version_segment_in_base_url, catalog_extensions_patch_fragment_for_unknown_model_is_skipped
• cargo test --bin crabcode aisdk::providers → 88 passed; model::extensions → 19 passed
• cargo fmt clean, scripts/check-aisdk-boundary.sh ok
• Cross-checked against .devrefs: neither opencode (packages/llm/src/protocols/openai-responses.ts) nor grok-build (xai-grok-sampling-types/.../responses.rs) needs this because first-party backends mint globally-unique ids — this is purely a relay-compat shim.

Follow-ups (not in scope)

• Stored history still holds duplicate ids; UI pairing (src/llm/client.rs has_call) drops the 2nd call_1. Consider uniquifying on ingest.
• Same per-request-id relay over the Chat Completions path would still send dupes; dedupe there if it surfaces.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[cloudflare-workers-and-pages]

Deploying crabcode with    Cloudflare Pages

Latest commit:	b8a552d
Status:	✅  Deploy successful!
Preview URL:	https://bbc13d59.crabcode.pages.dev
Branch Preview URL:	https://fix-opencode-zen-gateway-com.crabcode.pages.dev

View logs

[Blankeos]

PR Review — fix/opencode-zen-gateway-compat -> main

PR: #46 — "fix: opencode zen gateway compat (muse-spark 1.3 via opencode-go sub)"
URL: #46
Scope: 6 files, +288/−37 (1 commit b8a552d)

Merge confidence: 4/5

Suggested PR comment body

Gateway-compat shim for muse-spark via opencode-go relay: dedupes per-request Responses call_ids (call_1…), stops doubling /v1 on versioned base URLs (Responses + Anthropic + compatible), and skips incomplete catalog patch-fragments with a clearer log. Scoped to gateway paths; first-party providers unaffected. Tests pass (88 providers + 19 extensions), fmt + aisdk boundary clean. One nit: revert the stray _plans/__TODOS.md hunk before merge.

What I checked (read-only)

• gh pr view, git diff main...HEAD, git status
• cargo fmt --check → clean (exit 0)
• scripts/check-aisdk-boundary.sh → ok
• cargo test --bin crabcode aisdk::providers → 88 passed, 0 failed ✅
• cargo test --bin crabcode model::extensions → 19 passed, 0 failed ✅
• Read src/aisdk/providers/mod.rs:20, src/aisdk/providers/openai.rs:198, src/aisdk/providers/openai.rs:2355, src/aisdk/providers/anthropic.rs:742, src/aisdk/providers/compatible.rs:119, src/model/extensions/mod.rs:304, src/model/discovery.rs:64

Changes reviewed

1. Responses call_id dedupe (src/aisdk/providers/openai.rs:2355): occurrence-indexed rewrite to {id}_dupN on repeats, drops stale id on rewritten calls, wired into build_openai_messages (src/aisdk/providers/openai.rs:2344) so HTTP + websocket share it. First occurrence untouched. Correct fix for relay-minted sequential ids.
2. Version-segment helper (src/aisdk/providers/mod.rs:20): shared base_url_has_version_segment, reused by openai/anthropic/compatible; removes local dup in compatible.rs. Default paths now /responses / /messages on versioned bases, /v1/... otherwise.
3. Catalog fragment skip (src/model/extensions/mod.rs:304): patch-fragment for unknown model logs and skips instead of inserting. Also fixes unformatted {provider_id}/{model_id} log line. Nuance: src/model/discovery.rs:64 requires id/name (no default), so a bare {"attachment": true} already failed deserialize and skipped under the old let-Ok-else — this change mostly makes the skip explicit with a better log. A fragment with id/name but nothing else would still parse (defaults: tool_call=false, no cost/limits) and still insert — the "hollow entry" hole isn't fully closed.

Regressions?

No regressions found. Risk points (low):

• Dedupe pairing counts calls and outputs in separate maps, so it assumes occurrence-k call pairs with occurrence-k output. If history ever drops an output (the known has_call drop noted in the PR follow-ups), counts can skew and a renamed call could mismatch its output. Current replay path is fine; just don't rely on this if ingest starts dropping items.
• Version helper only matches /vN where N is digits followed by / or end — multi-digit (/v10, /v12) is missed by the bytes[i+3] check. No current provider uses that, so nit only.
• Custom responses_path configs bypass the version logic (explicit path wins) — intended, no regression.

Migrations?

None needed. No persistence/schema changes: no prefs.rs / auth.rs / mcp/credentials.rs / models.dev cache format changes, no SQL, no config contract change. Drop-in code fix.

Checklist before merge

• Revert stray hunk: _plans/__TODOS.md thinking-effort TODO (unrelated to this PR)
• Optional: confirm no custom responses_path users depend on the old /v1/v1 behavior (grep configs — default path only affects empty responses_path)
• Optional: full cargo test --bin crabcode (scoped suites pass; full suite not run here)
• cargo fmt / boundary already clean — no action

Confidence bumps

• Revert the _plans/__TODOS.md hunk → 4.5/5 (removes the only merge-hygiene issue; everything else is verified).
• Plus either (a) a guard/test for the diverged call/output-count edge in dedupe_responses_call_ids, or (b) tightening merge_catalog to reject fragments missing more than id/name (or documenting that id+name fragments are accepted) → 5/5.