integration/v0.13.0-rc4: x402 settle-guard, /v1 discovery fix, hermes context config, flow de-flakes, dead-code sweep#749
Merged
Conversation
Renders gateway.api_server.max_concurrent_runs into the agent's config.yaml. Nil omits the gateway block (Hermes internal default, 10); 0 disables the in-process cap so edge concurrency middleware governs instead. Existing agents render byte-identical. Part of #742. Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
If the buyer's request context is already canceled when the upstream handler succeeds, do not settle — the buyer would be debited for a response nobody receives (proven on-chain against hyperliquid-analyst, see #742). Reports failure reason client_disconnected; the interceptor hijacks the write path so nothing reaches the dead socket. Fixes the propagated-cancel leg of #742. Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
The first chat completion after stack-up intermittently fails with empty output (curl -f hides the response; observed across three release-smoke runs while steps 3-7 through the same port-forward and key always pass). Retry up to 3x and surface the HTTP status so a real failure is diagnosable instead of an empty string. Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
Local-server auto-discovery registered OpenAI-compatible endpoints (vLLM on :8000) with a bare api_base. LiteLLM's OpenAI provider does not append /v1 (CLAUDE.md pitfall 6), so every request through the discovered entry hit POST /chat/completions and 404'd. When an explicit 'obol model setup custom --endpoint .../v1' entry shared the same model group, requests coin-flipped between the good and bad deployment — root cause of the intermittent release-smoke flow-11 step[43] 404 and the flow-03 step[2] flake (latent since v0.10.0, 69f25bf). Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
Zero callers anywhere (including tests) — flagged by the /v1 pipeline audit's dead-code sweep alongside WarnAndStripV1Suffix (removed in #745). Other deadcode-tool hits in internal/{buy,inference} are test-covered exported API and were left alone. Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
…urrent-runs', 'origin/fix/flow03-first-request-retry' and 'origin/fix/discovery-api-base-v1' into integration/v0.13.0-rc4
Two-vendor (GPT + Grok) audited deletions, intersection-only: - inference: dead ProbeEndpoint/ScanLocalEndpoints/DetectServerType/ ParseModelsResponse wrappers (live path is the *Context variants used by model discovery); tests retargeted to live functions - erc8004: superseded NewClient/Register/RegisterDetailed/SetAgentURI (+ETH/RegisterWithOpts/SubmitRegister); tests retargeted to the WithOpts variants - x402: ResolveAssetInfo/BuildV1Requirement/TokenSupportedOnChain (production uses ResolveAssetInfoForPayment/BuildV2Requirement/ ResolveToken); test coverage migrated, not dropped - hermes: hermesExecArgs legacy test adapter + dead List/Skills/ ResolveInstance; agentruntime/dns/stack/schemas/serviceoffercontroller zero-reference symbols; dead flow-lib shell functions - flow-11: fix stale diag capture (deploy/litellm -c x402-buyer -> deploy/x402-buyer, stale since the b68fc34 split) Held back by committee split or product decision: RouteRulesForOffer (golden-test seam), erc8004 SetMetadata, all TEE/CoCo + enclave + OpenClaw capabilities (kept intentionally, see #748). Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
- pitfall 6 generalized: EVERY LiteLLM openai/ api_base must include /v1 (LiteLLM posts <api_base>/chat/completions verbatim); buyer-side tooling uses the opposite convention; discovery fixed in #745 - new pitfall 22: poisoned model group (two deployments, one bare api_base -> intermittent 404, no retry; GET /model/info vs CM is the decisive check) - new pitfall 23: buyer disconnect != no charge (hermes non-streaming path completes + settles after abort; keep runs short via spec.maxTurns; maxConcurrentRuns semantics; 4xx never settled) - skill: lessons table updated (poisoned model group, parserless QA endpoint -> relaunch with publisher-exact launcher), version 3.2.0 Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
…gration/v0.13.0-rc4
…d agents Hermes agents cannot discover the real context window through LiteLLM (its /models response strips max_model_len) and fall back to assuming 65536 tokens — then request max_tokens=65536, which any backend with a smaller window rejects. Hermes misclassifies that as context exhaustion, compresses a near-empty conversation, and dies with "Context length exceeded (21 tokens). Cannot compress further." — the exact failure of flow-13/14 step 45 (agent readiness preflight), reproduced and fixed standalone against hermes-agent v2026.7.1 + vLLM. Hermes also hard-requires >=64K context, so the QA endpoint itself must serve >=65536 (spark1 vLLM relaunched with --max-model-len 65536). - generateConfig: honor OBOL_LLM_CONTEXT_LENGTH / OBOL_LLM_MAX_TOKENS env knobs -> model.context_length / model.max_tokens (hermes config resolution step 0, overrides its probing) - release-smoke: default OBOL_LLM_CONTEXT_LENGTH=65536, OBOL_LLM_MAX_TOKENS=8192 whenever the OBOL flows are enabled Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
…s 100->2048) Reasoning models (qwen3 family behind the vLLM reasoning parser) emit reasoning_content before the tool call; at max_tokens=100 the budget is sometimes exhausted mid-reasoning and the response comes back finish_reason=length with no tool_calls — an intermittent FAIL on an otherwise healthy endpoint (passed run5, failed run6, temperature=0 notwithstanding). 2048 verified live: finish_reason=stop, tool call returned. Claude-Session: https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk
This was referenced Jul 14, 2026
OisinKyne
approved these changes
Jul 14, 2026
OisinKyne
left a comment
OisinKyne
left a comment
Contributor
There was a problem hiding this comment.
Seems like a surprising amount of deletions but i'll trust you on it
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
What changed: v0.13.0-rc4 integration train. Rolls up the four post-rc4 fix PRs (#743 x402 settle-guard for disconnected buyers +
spec.maxConcurrentRuns, #744 flow-03 first-completion de-flake, #745 discovery/v1api_base fix ending the poisoned-model-group 404s, #747 committee-verified dead-code sweep ~630 LOC + docs refresh) plus two integration-branch fixes found by the release smoke itself: explicit hermesmodel.context_length/max_tokensrendering (hermes ≥64K requirement invisible through LiteLLM) and flow-03 tool-call reasoning headroom.Why it matters: closes the two chronic release-smoke failures (flow-11 poisoned model group, flow-03 flake), makes buyer-disconnect settlements impossible, and produces the first fully green 18-flow release smoke including the two OBOL flows (13/14) that had never passed.
Risk level: medium
Commit under test: 8f9b58e
Base branch: main
Scope
Validation
Unit tests:
Release smoke:
Flow tests:
Live Chain Evidence
Network: base-sepolia (eip155:84532)
RPC/provider: Alchemy (key redacted)
Facilitator: https://x402.gcp.obol.tech
Contracts and tokens:
Wallet roles:
Transaction receipts:
Known non-gating warning: flow-14
setMetadatareverts (custom error 0x7e273289) when metadata is already set on the live registry from a prior QA run; registration + settlement unaffected.Runtime Evidence
Superseded PRs
Content already on main (patch-equivalent, landed via the earlier rc4 train): #726, #728, #729, #730, #731, #732, #733.
Rolled into this integration branch: #743, #744, #745, #747.
All will be closed as superseded with a reference to this PR.
Related issues: #742 (zombie settlement, fixed), #746 (model-group newtypes follow-up), #748 (dormant-capability subtree decisions — TEE/OpenClaw kept).
https://claude.ai/code/session_014YjPMViNrZ7zBVgUQzwEKk