fix(xai-responses): strip enum values containing '/' from tool schemas by Slimydog21 · Pull Request #28021 · NousResearch/hermes-agent

Slimydog21 · 2026-05-18T13:55:10Z

Summary

xAI's /v1/responses and /v1/chat/completions reject tool schemas whose enum values contain a forward slash with a generic HTTP 400 Invalid arguments passed to the model. — emitted as the first SSE frame on the streaming path, surfacing as _StreamErrorEvent and then retry-exhaustion in run_agent.py. The schema compiler trips on the / character regardless of where it appears in the enum value.

This is hit most commonly by MCP-derived tools whose enum lists HuggingFace model IDs (Qwen/Qwen3.5-0.8B, openai/gpt-oss-20b) or owner/name environment identifiers — e.g. the prime-cli lab MCP server's mcp_prime_lab_train_model tool.

Adds a strip_slash_enum sanitizer that mirrors the existing strip_pattern_and_format and is wired up at the same is_xai_responses call site in agent/chat_completion_helpers.py.

Reproduction

Captured the live failure by spying on the request body sent to https://api.x.ai/v1/responses, isolating the single offending tool from a 40-tool request, and bisecting the schema keyword against the live API:

all 40 tools                                       -> 400
each tool individually except mcp_prime_lab_train_model -> 200
mcp_prime_lab_train_model alone                    -> 400
^ same tool with enum: ["foo","bar"]               -> 200
^ same tool with enum: ["Qwen/...","openai/..."]   -> 400

Replayed the captured 78 KB body against the live API with the new sanitizer applied: ✅ 200, real response.

Design notes

Drops the whole enum keyword when any value contains /. Keeping a partial enum (slashless values only) would still 400 because xAI's failure is all-or-nothing on the keyword. The field description still reaches the model, so the prompting hint is preserved.
Walks both schema shapes. Handles OpenAI-format ({"function": {"parameters": …}}) and Responses-format ({"name": …, "parameters": …}) tool entries.
Ignores non-string enum values (integer/boolean enums can't contain /).
Idempotent — second pass is a no-op.

Test plan

9 new unit tests covering: HF-style IDs, slashless enums preserved, partial-match → whole-enum stripped, Responses-format tools, anyOf recursion, idempotency, empty/None inputs, non-string enum values.
Full tests/tools/test_schema_sanitizer.py suite passes (40/40).
End-to-end: hermes chat -q '…' with provider: xai-oauth, model: grok-4.3, and the prime_lab MCP server enabled — fails on main with Invalid arguments, succeeds on this branch.

Files changed

tools/schema_sanitizer.py — new strip_slash_enum(tools) -> (tools, count) function.
agent/chat_completion_helpers.py — chains the new sanitizer alongside strip_pattern_and_format in the is_xai_responses branch.
tests/tools/test_schema_sanitizer.py — 9 new tests.

xAI's /v1/responses and /v1/chat/completions reject tool schemas whose enum values contain a forward slash with a generic HTTP 400 "Invalid arguments passed to the model." before any token is emitted — the schema compiler appears to trip on the '/' character regardless of where it appears. Most commonly hit by MCP-derived tools whose enum lists HuggingFace IDs ("Qwen/Qwen3.5-0.8B", "openai/gpt-oss-20b") or owner/name environment identifiers. Reproduces against the live API; verified by isolating the single offending tool from a 40-tool request body and bisecting the schema keyword: enum: ["foo", "bar"] -> 200 enum: ["Qwen/...", "openai/..."] -> 400 Mirrors the existing strip_pattern_and_format sanitizer wired up at the codex_responses code path for is_xai_responses. The new strip_slash_enum walks tool parameters and drops the entire enum keyword when any value contains '/' — keeping it partial would still 400 since xAI's failure is all-or-nothing on the enum. The field description still reaches the model so the prompting hint is preserved. Adds 9 unit tests covering: HF-style IDs, slashless enums (preserved), partial-match (whole-enum stripped), Responses-format (no `function` wrapper), anyOf recursion, idempotency, empty/None inputs, and non-string enum values.

Slimydog21 · 2026-05-18T13:58:38Z

I am a vibe coder and code may be bad; have mercy with me.

Just found out that my PrimeIntellect stuff had some conflicts with the XAI integration.

Hope this is useful to someone ...

alt-glitch · 2026-05-18T14:12:35Z

Duplicate of #27907 — same fix (strip slash-containing enum values from tool schemas for xAI Responses). #27907 was triaged earlier and covers the same code paths.

Slimydog21 · 2026-05-18T14:17:34Z

Duplicate of #27907 — same fix (strip slash-containing enum values from tool schemas for xAI Responses). #27907 was triaged earlier and covers the same code paths.

Do I close this PR then? New to this ...

@Slimydog21

xAI's /v1/responses and /v1/chat/completions endpoints reject tool schemas whose enum values contain a forward slash with a generic HTTP 400 'Invalid arguments passed to the model.' before any token is emitted — the schema compiler trips on the '/' character regardless of where it appears. Most commonly hit by MCP-derived tools whose enum lists HuggingFace model IDs ('Qwen/Qwen3.5-0.8B', 'openai/gpt-oss-20b') or owner/name environment identifiers. Mirrors the existing strip_pattern_and_format sanitizer (PR for #27197). The new strip_slash_enum walks tool parameters and drops the entire enum keyword when any value contains '/' — keeping it partial would still 400 since xAI's failure is all-or-nothing on the enum. The field description still reaches the model so the prompting hint is preserved. Wired in at both code paths for parity: - agent/chat_completion_helpers.py (main agent xAI Responses path) - agent/auxiliary_client.py (aux client xAI Responses path, matching the same parity guarantee 2fae8fb established for pattern/format) Salvaged from #28021 by @Slimydog21 — contributor's branch was severely stale (would have reverted ~5000 LOC across azure/kanban/i18n); fix re-applied surgically on current main with their sanitizer + 9 tests preserved verbatim. Author noreply email used (original was a Mac hostname leak).

teknium1 · 2026-05-18T17:37:40Z

Salvaged via #28122 — same severely-stale-branch situation as #27911/#28003 (would have reverted ~5000 LOC unrelated). Sanitizer + 9 tests preserved verbatim, plus wired in the auxiliary_client.py parity path the verdict noted was missing. You remain the commit author (used noreply email — your original commit leaked a Mac hostname). Thanks!

@Slimydog21

xAI's /v1/responses and /v1/chat/completions endpoints reject tool schemas whose enum values contain a forward slash with a generic HTTP 400 'Invalid arguments passed to the model.' before any token is emitted — the schema compiler trips on the '/' character regardless of where it appears. Most commonly hit by MCP-derived tools whose enum lists HuggingFace model IDs ('Qwen/Qwen3.5-0.8B', 'openai/gpt-oss-20b') or owner/name environment identifiers. Mirrors the existing strip_pattern_and_format sanitizer (PR for NousResearch#27197). The new strip_slash_enum walks tool parameters and drops the entire enum keyword when any value contains '/' — keeping it partial would still 400 since xAI's failure is all-or-nothing on the enum. The field description still reaches the model so the prompting hint is preserved. Wired in at both code paths for parity: - agent/chat_completion_helpers.py (main agent xAI Responses path) - agent/auxiliary_client.py (aux client xAI Responses path, matching the same parity guarantee 2fae8fb established for pattern/format) Salvaged from NousResearch#28021 by @Slimydog21 — contributor's branch was severely stale (would have reverted ~5000 LOC across azure/kanban/i18n); fix re-applied surgically on current main with their sanitizer + 9 tests preserved verbatim. Author noreply email used (original was a Mac hostname leak).

@Slimydog21

xAI's /v1/responses and /v1/chat/completions endpoints reject tool schemas whose enum values contain a forward slash with a generic HTTP 400 'Invalid arguments passed to the model.' before any token is emitted — the schema compiler trips on the '/' character regardless of where it appears. Most commonly hit by MCP-derived tools whose enum lists HuggingFace model IDs ('Qwen/Qwen3.5-0.8B', 'openai/gpt-oss-20b') or owner/name environment identifiers. Mirrors the existing strip_pattern_and_format sanitizer (PR for NousResearch#27197). The new strip_slash_enum walks tool parameters and drops the entire enum keyword when any value contains '/' — keeping it partial would still 400 since xAI's failure is all-or-nothing on the enum. The field description still reaches the model so the prompting hint is preserved. Wired in at both code paths for parity: - agent/chat_completion_helpers.py (main agent xAI Responses path) - agent/auxiliary_client.py (aux client xAI Responses path, matching the same parity guarantee 2fae8fb established for pattern/format) Salvaged from NousResearch#28021 by @Slimydog21 — contributor's branch was severely stale (would have reverted ~5000 LOC across azure/kanban/i18n); fix re-applied surgically on current main with their sanitizer + 9 tests preserved verbatim. Author noreply email used (original was a Mac hostname leak).

@Slimydog21

xAI's /v1/responses and /v1/chat/completions endpoints reject tool schemas whose enum values contain a forward slash with a generic HTTP 400 'Invalid arguments passed to the model.' before any token is emitted — the schema compiler trips on the '/' character regardless of where it appears. Most commonly hit by MCP-derived tools whose enum lists HuggingFace model IDs ('Qwen/Qwen3.5-0.8B', 'openai/gpt-oss-20b') or owner/name environment identifiers. Mirrors the existing strip_pattern_and_format sanitizer (PR for NousResearch#27197). The new strip_slash_enum walks tool parameters and drops the entire enum keyword when any value contains '/' — keeping it partial would still 400 since xAI's failure is all-or-nothing on the enum. The field description still reaches the model so the prompting hint is preserved. Wired in at both code paths for parity: - agent/chat_completion_helpers.py (main agent xAI Responses path) - agent/auxiliary_client.py (aux client xAI Responses path, matching the same parity guarantee 2fae8fb established for pattern/format) Salvaged from NousResearch#28021 by @Slimydog21 — contributor's branch was severely stale (would have reverted ~5000 LOC across azure/kanban/i18n); fix re-applied surgically on current main with their sanitizer + 9 tests preserved verbatim. Author noreply email used (original was a Mac hostname leak).

Slimydog21 closed this May 18, 2026

Slimydog21 reopened this May 18, 2026

alt-glitch added type/bug Something isn't working provider/xai xAI (Grok) comp/agent Core agent loop, run_agent.py, prompt builder tool/mcp MCP client and OAuth P3 Low — cosmetic, nice to have labels May 18, 2026

teknium1 mentioned this pull request May 18, 2026

fix(xai-responses): strip enum values containing '/' from tool schemas #28122

Merged

teknium1 closed this in #28122 May 18, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(xai-responses): strip enum values containing '/' from tool schemas#28021

fix(xai-responses): strip enum values containing '/' from tool schemas#28021
Slimydog21 wants to merge 1 commit into
NousResearch:mainfrom
Slimydog21:slimydog/fix-xai-responses-slash-enum

Slimydog21 commented May 18, 2026

Uh oh!

Slimydog21 commented May 18, 2026

Uh oh!

alt-glitch commented May 18, 2026

Uh oh!

Slimydog21 commented May 18, 2026

Uh oh!

teknium1 commented May 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

Slimydog21 commented May 18, 2026

Summary

Reproduction

Design notes

Test plan

Files changed

Uh oh!

Slimydog21 commented May 18, 2026

Uh oh!

alt-glitch commented May 18, 2026

Uh oh!

Slimydog21 commented May 18, 2026

Uh oh!

teknium1 commented May 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants