feat(agents): CRD-driven native MCP + complete agent model-strip + base-sepolia USDC fix by bussyjd · Pull Request #673 · ObolNetwork/obol-stack

bussyjd · 2026-06-26T08:19:03Z

What

This started as a focused agent model-strip and grew into broader agent discovery + runtime hardening. It now does three things, plus two smaller fixes:

Complete the agent model-strip across all discovery surfaces — the original goal, now finished by also stripping /api/services.json.
feat: CRD-driven native MCP servers for agents — Hermes agents get native MCP tools from the Agent CR, with payment abstracted.
fix: base-sepolia USDC EIP-712 domain name — "USDC", not "USD Coin" (verified on-chain).

1. Agent model-strip (now complete)

For type=agent offers the underlying model is an internal detail: the agent runs its own model/skills/memory and Hermes ignores the chat-completions model field (the buyer never selects one). It was leaking into every discovery surface; this PR removes it from all of them:

surface	commit
402 page copy + `pay-agent` example	`e605f55`
JSON 402 `extra.agentModel`	`92716d6`
bazaar discovery example	`0f86017`
`/skill.md` catalog (Model column + bullet)	`2a88c3f`
`/api/services.json` (entry `model`)	`f8c0d8e`

buildServiceCatalogJSON was the last bypass — it set the entry model from spec.model / the resolved agent instead of the shared catalogModelName helper. f8c0d8e routes it through catalogModelName so all five surfaces agree. This reverses the earlier "services.json left untouched on purpose" note — there's no reason for the storefront feed to leak the id either. Also drops the no-op --model from buy.py pay-agent.

Inference offers are deliberately unchanged — there the buyer selects the model (paid/<remote-model>), so the real id stays buyer-facing.

2. feat: CRD-driven native MCP servers (`5427f4d`)

Give Hermes agents native MCP instead of a hand-patched ConfigMap:

New AgentSpec.MCPServers []AgentMCPServer{name,command,args,url,transport,env} on the Agent CRD.
The controller renders an mcp_servers: block into the agent's hermes-config; Hermes (tools/mcp_tool.py) discovers each server's tools as first-class tools.
stdio (command+args) for a local server, url for a remote one. env values support ${VAR} interpolation so signer creds reach the stdio subprocess without landing in the ConfigMap.

This makes native MCP durable across agent sync / stack up — it is driven by the CR, not a manual edit. In production by hyperliquid-analyst, whose stdio MCP server abstracts x402 payment per tool call (each call pays $0.001 internally; the agent never sees payment).

3. fix: base-sepolia USDC EIP-712 name (`72fa830`)

Base-Sepolia USDC is FiatTokenV2_2, whose EIP-712 domain name is "USDC", not the mainnet "USD Coin". Advertising "USD Coin" in the 402 extra makes a real facilitator reject otherwise-valid EIP-3009 signatures from third-party buyers (obol's own buy.py is offline-guarded and unaffected). Fixed in chains.go + tokens.go. Base mainnet ("USD Coin") is correct and left unchanged — verified on-chain: the reconstructed DOMAIN_SEPARATOR for native USDC 0x8335…2913 matches "USD Coin", not "USDC".

Also included

0229b8f — serve the catalog as UTF-8 (stop em-dash mojibake).
e83bd81 — make sold agents actually run their skills.
f8c0d8e also folds in a stale-test fix (agentcrd contract lifetime_seconds 90 → 180).

Verify

go build ./..., go vet, go test ./internal/x402/... ./internal/serviceoffercontroller/... ./internal/monetizeapi/... ✅
Deployed + verified live on intel.v1337.org: gemma4-aeon-uncensored now appears 0× across the 402 page, JSON 402, /skill.md, and /api/services.json; a native-MCP analyst turn returns real data with payment abstracted; the 402 advertises name=USDC (base-sepolia); base mainnet DOMAIN_SEPARATOR matches "USD Coin" on-chain.

Deploy note

The live silvernuc3 cluster currently runs hand-pinned local images built from this branch (serviceoffer-controller, x402-verifier); the embedded helmfile still pins the old ghcr controller, so a plain obol stack up would revert them (and prune the mcpServers schema). The next release — official images carrying these commits + a bumped x402.yaml pin + the updated CRD — closes that drift.

https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

An Obol Agent (type=agent) runs its own model, skills, and memory — the buyer never selects one, and Hermes ignores the chat-completions `model` field (resolved from its own config). Surfacing the underlying model in the 402 page is noise and exposes an implementation detail, and pay-agent's required `--model` flag had no effect on agent calls. - paymentrequired.go agentCopy: remove '(running <model>)', the '"model":' line in the example body, and '--model <id>' from the pay-agent example. The agent copy is now model-free. - buy-x402 buy.py: pay-agent no longer accepts/sends --model (synthesised body is just {messages, stream}); usage strings updated. - buy-x402 SKILL.md: pay-agent documented without --model. go build + x402/buyer/embed tests green. Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

mergeAgentExtras no longer adds extra.agentModel — an Obol Agent runs its own model and the buyer never selects one, so the model id is an internal detail, not buyer-facing info (it also rendered in the HTML 402 page's raw-JSON card). agentSkills/agentRuntime still surface so clients can tell it's an agent. Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

…t offers For agent offers the buyer never selects a model — the agent runs its own and ignores the chat-completions `model` field — so the bazaar discovery example now seeds the neutral 'your-model-id' placeholder instead of the real upstream id (which also rendered in the HTML 402 page's embedded raw-JSON card). Inference offers are unchanged: there the model IS buyer-selectable, so the real id stays. Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

…for agent offers The /skill.md catalog (and its Service Details section) showed the agent's underlying model in the Model column / **Model** bullet. An agent runs its own model and ignores the request `model` field, so the id is an internal detail — agent rows now render '—' and omit the **Model** bullet. Inference offers keep their model (there the buyer selects it). Mirrors the 402 page/extra/bazaar model-strip in internal/x402. The /api/services.json feed still resolves the agent model on purpose (drives the storefront UI) — left untouched. Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

…mojibake The skill-catalog busybox httpd served .md/.html with a bare text/* Content-Type (no charset), so clients fell back to Latin-1/CP1252 and rendered UTF-8 em dashes (the catalog's '—' placeholders, accented operator descriptions) as 'â€”'. Add charset=utf-8 to the text MIME mappings. JSON stays clean (always UTF-8 per RFC 8259). Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

Two fixes for type=agent offers whose skills shell out to scripts (buy-x402, the hyperliquid data skill), found while debugging the hyperliquid-analyst: - Disable the code_execution toolset. execute_code runs arbitrary in-process Python that bypasses the terminal DANGEROUS_PATTERNS gate, so Hermes fails it closed without a per-script approval no human can grant in an unattended paid turn. Small models then loop on it until the turn dies (observed: gemma4 returning a templated placeholder answer instead of fetching data). Disabling it routes skills through the terminal tool, where benign python3 auto-approves and dangerous commands stay gated. - Raise terminal.timeout/lifetime_seconds 80s/90s -> 170s/180s. Sold agents run behind a named tunnel (no ~100s quick-tunnel cut), and a paid data call (x402 round-trip + a first-party query) can exceed 80s, timing out heavier queries. Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

…json for agent offers The catalog JSON builder set the entry model from spec.model / the resolved agent model, bypassing catalogModelName -- so /api/services.json still leaked the internal model id for type=agent offers even though skill.md, the 402 page/extra, and the bazaar example already strip it (#673). Route the JSON builder through catalogModelName so all discovery surfaces agree: agent offers omit the model (they run their own and ignore the request `model`), inference offers keep it. Invert the test that pinned the old leaky behaviour. Also folds in pre-existing #673 housekeeping that was still uncommitted: - buy-x402 SKILL.md + buy.py: drop stale --model residue - agentcrd contract test: fix stale lifetime_seconds 90 -> 180 drift Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

…oin" Base-Sepolia USDC is FiatTokenV2_2, whose EIP-712 domain name is "USDC" -- unlike mainnet USDC ("USD Coin"). Advertising "USD Coin" in the 402 accepts extra makes a real facilitator reject otherwise-valid EIP-3009 signatures from third-party buyers (obol's own buy.py is offline-guarded and unaffected). This is the recurring base-sepolia "name" bug that a stub facilitator silently masks; forward-ports 31604a1 so a verifier rebuilt off this branch stays correct. - chains.go: ChainBaseSepolia.EIP3009Name -> "USDC" - tokens.go: USDC/base-sepolia EIP712Name -> "USDC" Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

Give Hermes agents native MCP instead of a hand-patched ConfigMap. Adds AgentSpec.MCPServers ([]AgentMCPServer{name,command,args,url,transport,env}) to the Agent CRD; the controller renders an `mcp_servers:` block into the agent's hermes-config so Hermes (tools/mcp_tool.py) discovers each server's tools as first-class tools. stdio (command+args) for a local server, url for a remote one; env values support ${VAR} interpolation so signer creds reach the stdio subprocess without landing in the ConfigMap. This makes native MCP durable across `agent sync` / `stack up` -- it is driven by the CR, not a manual edit. Used in production by hyperliquid-analyst, whose stdio MCP server (hl_mcp.py) abstracts x402 payment per tool call. - monetizeapi: AgentMCPServer type + deepcopy - agent-crd.yaml: mcpServers schema (maxItems 32) - agent_render.go: renderMCPServersBlock (sorted env keys, %q quoting) - agent_render_test.go: TestRenderHermesConfig_MCPServers Claude-Session: https://claude.ai/code/session_01XgUndZjSoxr2jNNGG5sVYD

bussyjd added 4 commits June 26, 2026 12:18

bussyjd changed the title ~~fix(x402): drop the underlying model from agent 402 copy + pay-agent~~ fix: drop the underlying model from agent discovery surfaces (402 + /skill.md) Jun 26, 2026

bussyjd added 5 commits June 26, 2026 13:19

bussyjd changed the title ~~fix: drop the underlying model from agent discovery surfaces (402 + /skill.md)~~ feat(agents): CRD-driven native MCP + complete agent model-strip + base-sepolia USDC fix Jun 27, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(agents): CRD-driven native MCP + complete agent model-strip + base-sepolia USDC fix#673

feat(agents): CRD-driven native MCP + complete agent model-strip + base-sepolia USDC fix#673
bussyjd wants to merge 9 commits into
mainfrom
fix/agent-402-instructions-drop-model

bussyjd commented Jun 26, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

bussyjd commented Jun 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

1. Agent model-strip (now complete)

2. feat: CRD-driven native MCP servers (5427f4d)

3. fix: base-sepolia USDC EIP-712 name (72fa830)

Also included

Verify

Deploy note

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

bussyjd commented Jun 26, 2026 •

edited

Loading

2. feat: CRD-driven native MCP servers (`5427f4d`)

3. fix: base-sepolia USDC EIP-712 name (`72fa830`)