TL;DR I run OpenClaw against Fireworks serverless and wanted their current reasoning lineup — DeepSeek V4 Pro, MiniMax M3, GLM-5.1, GPT-OSS 120B — selectable like any bundled model instead of falling through to the dynamic-model guesswork. Cataloging DeepSeek V4 on Fireworks also trips a real bug: core's deepseek-native fallback injects thinking next to reasoning_effort and Fireworks rejects that request outright. The fix turned out to need zero new runtime code — one manifest compat field core already supports.

Summary

What does this add? Four Fireworks serverless reasoning models to the bundled extensions/fireworks manifest catalog, each with reasoning_effort compat (supported efforts + level map) and a per-family thinking profile so thinking menus only offer levels the deployment actually accepts:

Thinking menus are derived from each row's supportedReasoningEfforts (extensions/fireworks/thinking-policy.ts) rather than hand-maintained per family, so the catalog row is the single source of truth and the /think menu can never drift from the wire contract.

What does this fix? — DeepSeek V4 Pro 400. Without the patch, selecting fireworks/accounts/fireworks/models/deepseek-v4-pro with thinking enabled kills every turn. Core's DeepSeek V4 fallback (src/agents/embedded-agent-runner/extra-params.ts) matches the id's last segment (deepseek-v4-pro) on any unowned OpenAI-compatible provider and injects thinking: { type: "enabled" } plus reasoning_effort. Fireworks rejects that payload:

{"error":{...,"message":"cannot specify both 'thinking' and 'reasoning_effort'"}}

Core already built the seam for exactly this class of deployment: compat.thinkingFormat ≠ "deepseek" suppresses the injected thinking entirely (even {type:"disabled"}) and strips deepseek replay fields, while reasoning_effort keeps flowing from the manifest map. The doc comment on deepSeekV4NativeThinkingAllowedByCompat names the use case (Azure AI Foundry DeepSeek V4 rejects thinking outright), and extra-params.deepseek-v4-thinking-format.test.ts pins the whole behavior. So the deepseek-v4-pro row carries "thinkingFormat": "openai" — no plugin runtime code, one manifest field core already supports.

Note — GPT-OSS 120B is the first bundled no-off profile. Its Fireworks deployment rejects reasoning_effort:"none" (400 Invalid reasoning effort: none, verified below), so its menu has no off. That exposes a separate core quirk: a session carrying a stored off that switches onto a no-off model was clamped up to the most expensive level. That fix is shared core logic, so it ships on its own — #93335 / fix/thinking-clamp-below-range (see Linked context). This PR is intentionally plugin-only; the two are best landed together.

Why supportsReasoningEffort: true on every row. Fireworks' baseUrl classifies as a proxy-like endpoint, so the transport's detected compat disables reasoning_effort by default; the explicit manifest flag is what keeps each model's effort contract on the wire.

Where do the numbers come from? Model ids, context windows, reasoning_effort ranges, and pricing (incl. cache-read rates) verified against the live Fireworks API (/v1/models, /v1/chat/completions) and https://docs.fireworks.ai/serverless/pricing. minimax-m3 is absent from that account-scoped /v1/models listing, so it was confirmed with live completions across its full effort range.

Linked context

• Sibling core fix — fix(thinking): clamp below-range requests down to the cheapest level,… #93335 / fix/thinking-clamp-below-range clamps a below-range thinking request (e.g. a stored off) down to the cheapest level instead of up to the most expensive. GPT-OSS 120B (added here) is the first bundled profile that triggers it. Independent PRs; recommend landing together.
• fix(fireworks): resolve catalog model params from plugin.json via core #90326 (same author, merged 2026-06-13, 22fdc39adb) makes core resolve bundled catalog params (contextWindow/maxTokens/compat) from openclaw.plugin.json. This PR builds on that merged behavior — the branch is rebased on current main, so the four rows are authoritative through core with no plugin-side resolution code. An earlier iteration re-implemented that resolution locally; all of it was deleted once fix(fireworks): resolve catalog model params from plugin.json via core #90326 landed (see Current review state).
• Core seam + tests: extra-params.ts (deepSeekV4NativeThinkingAllowedByCompat, non-native sanitizer), extra-params.deepseek-v4-thinking-format.test.ts.
• Sibling precedent: extensions/deepseek/openclaw.plugin.json ships the same supportsReasoningEffort compat; deepseek-v4-pro pricing matches the first-party row.
• Manifest compat keys (thinkingFormat, supportedReasoningEfforts, reasoningEffortMap) are existing config schema: src/config/zod-schema.core.ts, src/config/types.models.ts.

Real behavior proof

Behavior or issue addressed: Fireworks DeepSeek V4 Pro turns failing with 400 "cannot specify both 'thinking' and 'reasoning_effort'" when thinking is enabled; the three other reasoning families missing from the bundled catalog (no pricing, wrong default thinking levels via dynamic fallback).
Real environment tested: live Fireworks serverless API (api.fireworks.ai/inference/v1) with a real API key; OpenClaw checkout of this branch on Linux/Node 24.
Exact steps or command run after this patch: openclaw models list --all --provider fireworks --plain, then one noninteractive agent turn per model, e.g. openclaw agent --session-id fw-shot-ds -m "One sentence: why is the sky blue?" --model fireworks/accounts/fireworks/models/deepseek-v4-pro --thinking high (and the minimax/glm/gpt-oss equivalents).
Evidence after fix: screenshot below; red→green at the same command: openclaw agent --local --session-id fw-proof-ds -m "Say READY and nothing else." --model fireworks/accounts/fireworks/models/deepseek-v4-pro --thinking high → before: error=LLM request failed … rawError=400 cannot specify both 'thinking' and 'reasoning_effort'; after: READY. Raw API contract probe: a direct POST /v1/chat/completions for deepseek-v4-pro with {"thinking":{"type":"enabled"},"reasoning_effort":"low"} returns the same 400, while {"reasoning_effort":"low"} alone returns a normal completion — the patched plugin emits the latter shape.
Observed result after fix: all four models listed in the catalog with pricing; live one-turn agent runs on this branch return model replies for all four families (deepseek-v4-pro/minimax-m2p7/glm-5p1/gpt-oss-120b), including a deepseek-v4-pro turn at --thinking max through the dynamic-id path.
What was not tested: Fire Pass / routers/*-fast ids (separate follow-up), image input (all four rows are text-only per the live API), xhigh/max efforts end-to-end on deepseek (mapped per Fireworks' documented effort list but not exercised in a live turn).

Tests and validation

• extensions/fireworks vitest: 18 passed (catalog compat incl. the new thinkingFormat assertion, per-family thinking profiles, dynamic family compat merge, dynamic-model fallbacks); src/auto-reply/thinking.test.ts: 48 passed incl. the new no-off clamp regression; clamp consumers (agent-command.live-model-switch, cron model-override forwarding): green.
• docs/providers/fireworks.md: built-in catalog table and dynamic-id thinking notes refreshed with the new models.
• Core seam test: extra-params.deepseek-v4-thinking-format.test.ts 8 passed (pre-existing; proves the manifest override path this PR relies on).
• tsgo core + extension lanes clean; oxlint clean; oxfmt clean.
• Autoreview loop (claude-fable-5 thinking-high + codex gpt-5.5 medium), nine rounds until convergence; final verdicts: claude "patch is correct (0.85)", codex "patch is correct (0.83), no actionable findings". The full finding-by-finding ledger is in "Current review state" below.

Risk checklist

• Plugin-only — no core changes. Additive manifest rows; the only code is the data-driven thinking-profile derivation (thinking-policy.ts, replaces the old per-family Kimi matcher) and four id constants. Net +265/−23 across 5 files, all under extensions/fireworks/** and docs/**.
• No config/default surface changes outside the plugin manifest; no migrations; existing configured models are unaffected (rows are additive, ids unchanged).
• Dynamic (non-cataloged) Fireworks ids keep the exact pre-PR behavior except deepseek-v4*, which gains the compat opt-out (those turns 400'd before, so nothing working changes). Non-cataloged ids get the generic thinking menu; only the four cataloged ids get derived menus.
• Thinking-menu levels are clamped to what Fireworks accepts per model (verified per effort via curl), so a user-selected unsupported level can no longer produce a rejected request.
• Known dependency: GPT-OSS 120B is no-off, so a stored thinking-off session switched onto it clamps up to the most expensive level until the sibling core fix (fix/thinking-clamp-below-range) lands. No other added model is affected (the rest are off-capable).

Current review state

What is the next action? Maintainer review. Author attaches the terminal screenshot (catalog listing + one live turn per family + the curl effort matrix) at the placeholder above.

What is still waiting on author, maintainer, CI, or external proof? Author: screenshot. Maintainer: review/land decision + ordering vs the sibling core fix. CI: standard checks. No external proof outstanding.

Is this PR plugin-only? Yes. An earlier iteration of this branch carried much more — a per-family regex matcher (model-id.ts), a dynamic-family compat table (index.ts), exported FIREWORKS_*_COMPAT clones, and a core thinking-clamp fix. The first three were workarounds for core not resolving the plugin's openclaw.plugin.json catalog; once #90326 merged they were deleted. The core clamp fix was split out to fix/thinking-clamp-below-range so this PR is purely additive plugin data + manifest-derived menus + the DeepSeek thinkingFormat opt-out (one manifest field).

Which bot or reviewer comments were addressed? None posted on the current revision yet. The branch was force-pushed to its surgical form after #90326 merged and after the core fix was split out; a prior over-scoped revision is superseded.