v0.12.171

v0.12.171 — Apr 29, 2026

• Three new free NVIDIA-hosted models added. BlockRun refreshed the free catalog on 2026-04-29 with three additions, all wired into ClawRouter as free/-prefixed entries:
  • free/deepseek-v4-pro — 1.6T MoE / 49B active, 1M context, MMLU-Pro 87.5, GPQA 90.1, SWE-bench 80.6, LiveCodeBench 93.5. NIM ~150 tok/s on Blackwell. Strongest free reasoning model.
  • free/deepseek-v4-flash — 284B / 13B active MoE, 1M context, ~5x faster than v4-pro. Strong on chat/summarization (MMLU-Pro 86.2). Weaker factual recall (SimpleQA 34% vs Pro's 58%) — pick v4-pro for fact-heavy agentic loops.
  • free/nemotron-3-nano-omni-30b-a3b-reasoning — 31B / 3.2B active MoE, 256K context. First vision-capable free model in the catalog. Accepts text, images, video (up to 2min), audio (up to 1hr). ChartQA 90.3, DocVQA 95.6, MMMU 70.8.
• free/deepseek-v3.2 phased out in favor of free/deepseek-v4-pro (strict-superset replacement: same family, larger context, higher benchmarks). Removed from BLOCKRUN_MODELS, FREE_MODELS set, top-models.json picker, README pricing table, and SKILL.md model list. Aliases kept and redirected: nvidia/deepseek-v3.2, free/deepseek-v3.2, and deepseek-free now all resolve to free/deepseek-v4-pro so existing pins continue to work and silently get the upgrade.
• gpt-oss-120b / gpt-oss-20b deliberately kept as defaults despite BlockRun's 2026-04-28 retirement (available:false server-side). Heavy user demand outweighs the source-of-truth alignment for these specific IDs — free / nvidia / gpt-120b / gpt-20b aliases all still resolve to free/gpt-oss-120b (or 20b), FREE_MODEL constant still points at free/gpt-oss-120b, and ecoTiers.SIMPLE primary stays unchanged. ClawRouter's existing fallback-chain logic handles any 400 ("Model not available") from BlockRun by trying the next chain entry, so failures degrade gracefully rather than break user workflows.
• New shorthand aliases for the additions: deepseek-v4-pro, deepseek-v4-flash, v4-pro, v4-flash, nemotron-omni, nano-omni, vision-free — chosen to mirror BlockRun's bare-name aliases at route.ts:639-640 plus a vision-free discovery shortcut for the new vision-capable model.
• ecoTiers.SIMPLE fallback chain extended with the three new free models (mistral-small, deepseek-v4-flash, qwen3-next) inserted before the paid Gemini fallbacks, so eco-profile users get more all-free chain depth before paid models kick in. Primary is unchanged (free/gpt-oss-120b).
• Provider routing safety note. BlockRun's NVIDIA_MODEL_MAP in src/lib/ai-providers.ts:2094-2111 does NOT have explicit entries for the 3 new models, but callOpenAICompatible falls through to the bare model name (modelMap[k] || k), so ClawRouter sending nvidia/deepseek-v4-pro reaches NVIDIA NIM as bare deepseek-v4-pro — which is what NIM expects. Documented in the BLOCKRUN_MODELS comment block in src/models.ts. If BlockRun later adds explicit map entries with different upstream names, this side needs no change.
• Net free-model count: 8 → 10 (8 originals + 3 added - 1 phased out). README badge, tagline, "Quick Start" sections, and SKILL.md description all updated to reflect "10 free NVIDIA models". Pricing table in README adds three new rows in benchmark order.
• Test fixtures. src/router/strategy.test.ts MODEL_PRICING map gains entries for the 3 new free models. No assertion changes anywhere else — gpt-oss-120b stays the asserted default in src/exclude-models.test.ts, src/models.test.ts, test/fallback.ts, and test/integration/exclude-models.test.ts.

---

v0.12.170 — Apr 29, 2026

• Bare kimi / moonshot aliases now resolve to Kimi K2.6. BlockRun hid Kimi K2.5 from its public model UI on 2026-04-28 (commit bfbdedf) and now features K2.6 as the Moonshot flagship. ClawRouter's local alias map followed the old direction and still pointed kimi and moonshot at K2.5, which created a quiet drift from the source-of-truth registry: agents asking for "kimi" got the previous-gen model while BlockRun's homepage advertised K2.6. The aliases now resolve to moonshot/kimi-k2.6 and a new bare kimi-k2 alias is added for the same target. Users who explicitly pinned kimi-k2.5 continue to get K2.5 — the explicit pin is preserved as a cost-stability opt-in ($0.60/$3.00 vs K2.6's $0.95/$4.00). NVIDIA-hosted K2.5 (retired 2026-04-21) still redirects to moonshot/kimi-k2.5.
• Routing tier primaries deliberately unchanged. autoTiers.MEDIUM and agenticTiers.MEDIUM continue to anchor on moonshot/kimi-k2.5. Promoting them to K2.6 would silently raise per-call cost +58% on input / +33% on output for every default user — that's a separate decision tracked outside this release, ideally with measured retention/IQ data on K2.6 vs K2.5. premiumTiers.SIMPLE was already moonshot/kimi-k2.6 and is unchanged. Net effect: behavior shift is opt-in via the kimi alias / kimi-k2 shorthand, not forced through default routing.
• Doc and test fixture refresh. README's profile-overview table now shows kimi-k2.6 in the PREMIUM column (matching docs/routing-profiles.md and src/router/config.ts:1134). src/router/strategy.test.ts gains a K2.6 pricing fixture so cost-calc tests stay honest if K2.6 ever appears in test scenarios. src/proxy.models-endpoint.test.ts now asserts both kimi-k2.6 and moonshot/kimi-k2.6 are discoverable through the /models endpoint. test/fallback.ts's "Unknown model" example list leads with moonshot/kimi-k2.6.

v0.12.169

v0.12.169 — Apr 28, 2026

• Synthesize structured tool_calls from XML/text formats some models emit in content. Earlier tool-call hardening (v0.12.165, v0.12.166) handled the case where upstream returned a structured tool_calls array (or signaled finish_reason: "tool_calls") and the model also leaked planning prose into content. This release closes a third gap where upstream returns no structured tool calls at all and the model's actual tool invocations live as XML/text inside content — typical when a downstream client (OpenClaw is the visible offender) prompt-engineers tool instructions instead of sending a structured tools[] schema, so the model dutifully honors the prompt format and emits the call as text. Two formats observed in the wild are now recognized and converted to OpenAI-shaped tool_calls:
  • OpenClaw-style — <tool_call>NAME<arg_key>K1</arg_key><arg_value>V1</arg_value>...<arg_key>Kn</arg_key><arg_value>Vn</arg_value></tool_call>. Requires at least one arg_key/arg_value pair so prose like <tool_call>name</tool_call> in documentation does not mis-fire. Surfaced via a real ClawRouter→OpenClaw session where the agent emitted six identical <tool_call>web_search<arg_key>...</arg_key>... blocks in 60 seconds, none executed, then hallucinated "I need a Brave API key" as the failure explanation.
  • Anthropic-style — <function_calls><invoke name="NAME"><parameter name="K">V</parameter>...</invoke></function_calls>. Reproduction confirmed Moonshot Kimi K2.6 emits this format when given prompt-engineered tool instructions without a structured tools[] schema.
  • Values are best-effort coerced via JSON.parse so <arg_value>5</arg_value> becomes 5 (number) and <arg_value>true</arg_value> becomes true (boolean); strings that don't parse stay as strings. Synthesized IDs are OpenAI-shaped (call_<base64url>).
  • Wired into both response paths: the SSE conversion path (src/proxy.ts:5081+) and the non-streaming JSON path (src/proxy.ts:5325+). When extraction succeeds, content is blanked, message.tool_calls is populated, and finish_reason flips to "tool_calls" — matching exactly the shape downstream tool executors already handle from the v0.12.165/166 paths.
  • New module src/textual-tool-calls.ts plus src/textual-tool-calls.test.ts (13 unit tests) and four new integration tests in src/proxy.tool-forwarding.test.ts covering OpenClaw format / non-streaming, OpenClaw format / SSE, Anthropic format / non-streaming, and a negative test (plain prose passes through unchanged with finish_reason: "stop").
• /model picker allowlist now lives in src/top-models.json (single source of truth, loaded by src/top-models.ts). Previously injectModelsConfig() in src/index.ts carried a literal array that drifted from the install scripts' TOP_MODELS (which carry their own copies in scripts/reinstall.sh + scripts/update.sh). The JSON file is the version anyone actually edits going forward; both runtime (src/index.ts) and the test suite (src/top-models.test.ts) read from it. The install scripts still carry their own embedded copies because they run before npm dependencies are resolved — but now there's one canonical list to copy from when adding a new model.
• Alias adds. br-sonnet → anthropic/claude-sonnet-4.6 (matching the existing br- partner shorthand pattern), and gpt5 now resolves to openai/gpt-5.5 instead of openai/gpt-5.4 (following v0.12.167's GPT-5.5 promotion as BlockRun's newest visible flagship).

v0.12.168

v0.12.168 — Apr 25, 2026

• Propagate openai/gpt-5.5 everywhere it should appear. v0.12.167 added the model to BLOCKRUN_MODELS, the gpt-5.5 alias, and the install-script TOP_MODELS allowlist — but every other place ClawRouter advertises a flagship still pointed at gpt-5.4. This release closes the gap so 5.5 is a first-class citizen across routing, the picker, marketing, and the OpenClaw skill page.
  • src/router/config.ts — three fallback-chain insertions, no primary changes. openai/gpt-5.5 slots in immediately before openai/gpt-5.4 in auto.COMPLEX.fallback, premiumTiers.COMPLEX.fallback, and agenticTiers.COMPLEX.fallback. Both stay reachable; 5.5 gets preference when the chain reaches OpenAI. Comments updated so 5.5 is "newest flagship — 1M+ ctx, native agent + computer use" and 5.4 is "previous flagship — benchmarked at 6,213ms, IQ 57". Tier primaries are unchanged: promoting 5.5 to a primary slot needs measured latency/IQ data, which we don't have yet — that's a separate decision tracked outside this release.
  • src/index.ts — /model picker allowlist updated. src/index.ts carries its own copy of TOP_MODELS (separate from the install scripts' identical-but-distinct list — both populate the OpenClaw allowlist depending on install path). Added openai/gpt-5.5 and anthropic/claude-opus-4.5 (also missed in v0.12.167's BLOCKRUN_MODELS add for opus-4.5), and replaced the now-deprecated minimax/minimax-m2.5 with minimax/minimax-m2.7 so the picker matches the deprecation we landed yesterday.
  • README.md — Premium Models pricing table. Added the openai/gpt-5.5 row at $5.00/$30.00 per 1M tokens (~$0.0175 per 0.5K-in-0.5K-out request), 1M context, full feature set. Placed between claude-opus-4.6 ($0.0150) and o1 ($0.0375) so the table stays sorted by approximate $/request.
  • skills/clawrouter/SKILL.md — model list line. The "55+ models including..." line now leads gpt-5.5, gpt-5.4, ... and includes claude-opus-4.5 alongside 4.7/4.6.
• Files deliberately not touched: docs/smart-llm-router-14-dimension-classifier.md and docs/llm-router-benchmark-46-models-sub-1ms-routing.md are frozen benchmark archives — adding 5.5 to a benchmark table without measured numbers would falsify the document. The posts/*.md marketing content is similarly point-in-time. Those will be refreshed if/when 5.5 gets benchmarked.

v0.12.167

v0.12.167 — Apr 24, 2026

• Realign the model registry to BlockRun source-of-truth. Audit found three drifts where ClawRouter's BLOCKRUN_MODELS table didn't match what blockrun/src/lib/models.ts actually exposes. The server is the source of truth for which models exist and what they cost; the proxy's local view should mirror that 1:1 so cost estimation, the /model picker, and routing tier selection all see the same world the server does.
  • Add openai/gpt-5.5. BlockRun's newest visible OpenAI flagship — first fully retrained base since GPT-4.5, 1M+ context, 128K output, native agent + computer use. Pricing $5/$30 per 1M tokens. Added to BLOCKRUN_MODELS, the gpt-5.5 alias, and the TOP_MODELS allowlist in both install scripts. Routing tiers in src/router/config.ts continue to anchor on gpt-5.4 because that's what's benchmarked; users can pin 5.5 explicitly. Routing change is a separate decision.
  • Add anthropic/claude-opus-4.5 as a distinct model. Previously ClawRouter's MODEL_ALIASES silently rewrote anthropic/claude-opus-4.5 to 4.7, making 4.5 unreachable through ClawRouter even though BlockRun lists it as a separate visible model with its own pricing and 200K context (vs 4.6/4.7's 1M). Removed the alias, added 4.5 to BLOCKRUN_MODELS with its real 200K/32K shape, and added an anthropic/claude-opus-4-5 (dashed) alias for the slug variant. Test in src/models.test.ts was codifying the old upgrade-to-4.7 behavior — flipped to assert the pin is preserved end-to-end.
  • Mark minimax/minimax-m2.5 deprecated → fallback minimax/minimax-m2.7. BlockRun retired m2.5 entirely (only m2.7 is in their MODELS table). ClawRouter still listed both; m2.5 now flips to deprecated: true with the m2.7 fallback so existing pins keep working.
  • scripts/reinstall.sh + scripts/update.sh: drop minimax/minimax-m2.5 from the TOP_MODELS picker allowlist (still reachable, just hidden from the picker) and add openai/gpt-5.5 + anthropic/claude-opus-4.5.

v0.12.166

v0.12.166 — Apr 24, 2026

• Tool-call planning prose suppressed even when finish_reason is the only signal (thanks @0xCheetah1, #162). Follow-up to v0.12.165's #161 fix. Live Telegram/OpenClaw testing caught one more shape the planning-prose leak could wriggle through: some upstreams (Moonshot Kimi K2.6 again) mark a turn with finish_reason: "tool_calls" without exposing message.tool_calls / delta.tool_calls at the same inspection point. The #161 gate (toolCalls.length > 0) saw no array and let the prose through. The gate is now endsWithToolCalls || toolCalls.length > 0 — applied consistently across the non-streaming JSON path and the SSE emission path, plus the finish-reason override in the SSE terminal chunk. Two new regression tests in src/proxy.tool-forwarding.test.ts — one per response shape — lock the behavior in: a response with finish_reason: "tool_calls" and no tool_calls array has its content blanked and the tool_calls finish_reason preserved. User-visible impact: fewer "I should look up X before replying" preambles sneaking into agent chat surfaces for turns that are supposed to be pure tool invocations.

v0.12.165

v0.12.165 — Apr 24, 2026

• Tool-call planning prose no longer leaks to chat surfaces (thanks @0xCheetah1, #161). Some OpenAI-compatible providers — Moonshot's Kimi K2.6 was the visible offender through OpenClaw Telegram — return { content: "The user wants the current time. I should call get_current_time with Chicago.", tool_calls: [...] }. Tool execution only needs tool_calls; the content field is internal planning that the upstream should have hidden behind a <think> tag but didn't. ClawRouter now suppresses content whenever tool_calls.length > 0, in both the non-streaming JSON response path and the SSE-conversion path that clients like OpenClaw hit with stream: true. Tool execution is unaffected; only the user-visible planning prose goes away. Covered by two regression tests in src/proxy.tool-forwarding.test.ts (one per response shape).
• Plugin restart loop killed. injectModelsConfig() in src/index.ts writes ClawRouter-owned keys into ~/.openclaw/openclaw.json on every plugin load. OpenClaw's config watcher has a catch-all rule — any change with no matching plugin-declared prefix triggers a full gateway restart — so mcp.servers.blockrun writes kept ping-ponging the gateway. The plugin definition now exposes reload: { noopPrefixes: ["mcp.servers.blockrun"] } (new optional field on OpenClawPluginDefinition) to tell OpenClaw's loader that ClawRouter self-manages that prefix. Silently ignored on OpenClaw runtimes that predate the reload field.
• Dedup + response cache now isolate streaming and non-streaming callers. Discovered while adding the SSE regression test for the tool-call fix: a stream: true request that followed an identical-body stream: false request was getting content-type: application/json instead of text/event-stream. Two compounding bugs. ClawRouter rewrites parsed.stream = false before the upstream call (BlockRun API doesn't support streaming), and both RequestDeduplicator.hash(body) and ResponseCache.generateKey(body) ran AFTER that rewrite — so a stream:true and stream:false request hashed identically. Worse, response-cache.ts's normalizeForCache explicitly stripped stream from the key with the comment "we handle streaming separately" (it never did). Fix: (1) prefix both dedupKey and cacheKey in src/proxy.ts with the original isStreaming intent ("sse:" vs "json:"), so the two shapes never share a cache slot; (2) stop stripping stream in normalizeForCache. Latent bug — real-world impact was small because the exact scenario (identical body, different stream flag, within 30s/10min TTL) is rare in practice — but a correctness bug nonetheless. Regression test added (isolates dedup cache between streaming and non-streaming requests with identical bodies); the existing response-cache.test.ts expectation was inverted (it was codifying the broken behavior).

v0.12.164

v0.12.164 — Apr 23, 2026

• Video generation switched to async submit + poll (tracks BlockRun server commit 654cd35). The server-side /v1/videos/generations endpoint no longer blocks for the full 60–180s upstream generation — POST now returns 202 { id, poll_url } in ~3–20s, and a separate GET on the poll_url (same x-payment header) returns 202 while the job is queued/in_progress and 200 with the final video on completion. Server settles only on the first completed poll, so upstream failure or caller disconnect = zero USDC charged. ClawRouter's proxy handler in src/proxy.ts now collapses this back into a single blocking POST for the client: submit upstream, poll the poll_url every 5s (initial 3s grace) up to a 5-min deadline, then backup + serve locally as before. Legacy sync-shaped server responses still work — the handler checks for poll_url before switching to the poll loop. Client-side timeouts bumped: buildVideoGenerationProvider.timeoutMs 200s → 330s; /videogen slash 200s → 330s; both sit above the 5-min internal poll deadline so the last data[0].url finishes streaming back. User-facing impact: same blocking POST as before, but Cloudflare's 100s edge timeout no longer kills long-running Seedance 2.0 jobs.

• Image/video plumbing parity — four exposure surfaces now match the backend. The BlockRun server has supported 8 image models (DALL-E 3, GPT Image 1, Nano Banana / Pro, Flux 1.1 Pro, Grok Imagine / Pro, CogView-4) and 4 video models (Grok Imagine, Seedance 1.5 Pro / 2.0 Fast / 2.0) since v0.12.162, but the ClawRouter client exposed them inconsistently:

  • buildImageGenerationProvider in src/index.ts only advertised 4 image models. OpenClaw's native image picker couldn't see Flux, Grok Imagine (×2), or CogView-4 — the only way to hit them was raw curl with an explicit model field. The models array now lists all 8; defaultModel switched from openai/gpt-image-1 to google/nano-banana (cheapest general-purpose default); capabilities.geometry.sizes adds CogView-4's 512x512, 768x768, 768x1344, 1344x768, 1440x1440 sizes; capabilities.edit.enabled flipped to true so OpenClaw's edit UI surfaces gpt-image-1's /v1/images/image2image path.
  • MODEL_ALIASES in src/models.ts had zero image/video shortcuts. All 140+ aliases were LLM chat models. Added 17 new aliases so resolveModelAlias("dalle") → openai/dall-e-3, "flux" → black-forest/flux-1.1-pro, "seedance" → bytedance/seedance-1.5-pro, plus banana, banana-pro, nano-banana-pro, gpt-image, flux-pro, grok-imagine / -pro, grok-video, cogview, seedance-1.5, seedance-2, seedance-2-fast.
  • /imagegen and /videogen slash commands now actually exist. README documented /imagegen a dog dancing on the beach as if it worked, but no such command was ever registered — it was silent drift from the aspirational README. Both commands now register via api.registerCommand, accept --model=<alias>, --size=WxH, --n=<int>, --duration=<5|8|10> flags (parsed by a shared parseGenArgs helper), resolve aliases through resolveModelAlias, POST to the proxy's /v1/images/generations and /v1/videos/generations endpoints, and return inline markdown (![image](http://localhost:8402/images/...)) or video URLs. 402 responses surface as "top up with /wallet" hints; video timeout is 200s to cover upstream polling. /img2img remains README-only for now — will land in a follow-up.
  • Partner framework now includes image/video as LLM-callable tools. Added three new PartnerServiceDefinition entries in src/partners/registry.ts — image_generation, image_edit, video_generation — so the existing buildPartnerTools → api.registerTool pipeline surfaces them as blockrun_image_generation, blockrun_image_edit, blockrun_video_generation tools. Agents can now tool-call image/video from chat without the skill layer guessing at raw HTTP shapes.

• Dropped the Twitter/X user-lookup partner. We no longer run X data as a product surface. Removed x_users_lookup from PARTNER_SERVICES, deleted the skills/x-api/ skill directory, and stripped x| from the /v1/(?:x|partner|pm|...)/ paid-route regex in src/proxy.ts (so /v1/x/* no longer short-circuits to the partner proxy — it now falls through to the usual chat-completion path or 404s cleanly). Server-side /v1/x/* endpoints are still live at blockrun.ai/api for any existing integrations; only the client wiring is retired.

• /partners + clawrouter partners CLI output compressed ~4×. Previously 6 lines per service (name, full agent-facing description, tool name, method, pricing block, blank) × 17 services ≈ 100 lines of wall-of-text, which is what @vicky was calling out as "读不了" (unreadable). PartnerServiceDefinition gained two fields — category ("Prediction markets" / "Market data" / "Image & Video") and shortDescription (≤ 40 chars) — driving a new grouped, column-aligned one-liner per tool. The long description field stays intact for the LLM-facing JSON Schema (agents still see "Call this ONLY when..." guidance). Output is now ~25 lines, one screen.

v0.12.163

v0.12.163 — Apr 23, 2026

• README leads with the free tier. Post-v0.12.160 the product story changed — 8 NVIDIA models free forever, no wallet required to start — but the README still opened "fund your wallet" as step 2 of Quick Start and buried the free tier in a single line at the bottom. Rewrites so the free tier is the hook, not a footnote: hero tagline adds "8 models free, no crypto required. No signup. No API key. No credit card." plus a 🆓 shields.io badge; the "Why ClawRouter exists" list opens with "Starts at $0"; the comparison-vs-others table adds a "Free tier" row showing ClawRouter's "8 models, no signup" against OpenRouter's rate limits and LiteLLM/Martian/Portkey's "no"; Quick Start gets a "No wallet? 8 models work free out of the box" callout and reframes step 2 as optional; routing-profiles table adds /model free at 100% savings; the Costs section lists the current 8 free model IDs by name (was a stale 11-model list referencing the retired Nemotron Ultra / Mistral Large / Devstral). This release is README-only — code is identical to v0.12.162 — version bump exists so the updated marketing reaches the npmjs.com package page and the clawhub marketplace listing.

v0.12.162

v0.12.162 — Apr 23, 2026

• ByteDance Seedance video models wired into the client. BlockRun server has exposed three Seedance models since late April — bytedance/seedance-1.5-pro ($0.03/sec), bytedance/seedance-2.0-fast ($0.15/sec, ~60–80s gen time), and bytedance/seedance-2.0 Pro ($0.30/sec) — all 720p, text-to-video + image-to-video, 5s default and up to 10s. The /v1/videos/generations proxy passthrough in src/proxy.ts already forwarded any model value untouched, so actual USDC charges were always correct (server dictates the amount in its 402 response and payment-preauth.ts caches the server-sent PaymentRequired, not a local estimate — charges never depended on ClawRouter's local pricing table). Three client-side gaps were fixed anyway:
  • Usage telemetry was wrong for Seedance. estimateVideoCost in src/proxy.ts only knew xai/grok-imagine-video, so every Seedance request logged $0.42/clip to logUsage regardless of what the user was actually billed — skewing /usage output, savings %, and journal cost fields. VIDEO_PRICING now carries all four models at real server rates.
  • OpenClaw's native video UI only saw one model. buildVideoGenerationProvider in src/index.ts advertised models: ["xai/grok-imagine-video"], so users of the UI picker couldn't pick Seedance at all; the only path was raw curl with an explicit model field. The models array now lists all four, and provider capabilities widen to maxDurationSeconds: 10 / supportedDurationSeconds: [5, 8, 10] to cover both vendors' ranges (server still validates per-model maxDurationSeconds, so invalid combos return a clean 400).
  • README docs only mentioned Grok. Video-generation section now lists all four models in the table, swaps the curl example to bytedance/seedance-2.0-fast (sweet-spot price/quality), and makes the upstream-polling note vendor-neutral instead of xAI-specific.
• Docs: fixed proxy port in free-models guide. Thanks to @Bortlesboat (#160) for catching 4402 → 8402 typos in docs/11-free-ai-models-zero-cost-blockrun.md. The rest of the repo, src/config.ts (DEFAULT_PORT = 8402), and all other docs have always said 8402; that one guide was sending new users at the wrong local port.

v0.12.159

• Market data tools — BlockRun gateway now exposes realtime and historical market data; ClawRouter wires them into OpenClaw as 6 first-class agent tools so the model stops scraping finance sites. Paid ($0.001 via x402, same wallet as LLM calls): blockrun_stock_price and blockrun_stock_history across 12 global equity markets (US, HK, JP, KR, UK, DE, FR, NL, IE, LU, CN, CA). Free (no x402 charge): blockrun_stock_list (ticker lookup / company-name search), blockrun_crypto_price (BTC-USD, ETH-USD, SOL-USD, …), blockrun_fx_price (EUR-USD, GBP-USD, JPY-USD, …), blockrun_commodity_price (XAU-USD gold, XAG-USD silver, XPT-USD platinum). Tool schemas advertise market codes, session hints (pre/post/on), and bar resolutions (1/5/15/60/240/D/W/M). Path routing extended: the partner-proxy whitelist in src/proxy.ts now matches /v1/(?:x|partner|pm|exa|modal|stocks|usstock|crypto|fx|commodity)/, routing all new paths through proxyPaidApiRequest (payFetch handles 402 when present, passes through 200 for free categories). Tool definitions added in src/partners/registry.ts; skills/clawrouter/SKILL.md gains a "Built-in Agent Tools" section listing market data + X intelligence + Polymarket alongside the LLM router.