Skip to content

feat(agent): have sub-agents stream their responses to keep the tunnel alive - #597

Closed
OisinKyne wants to merge 1 commit into
mainfrom
oisin/streaming
Closed

feat(agent): have sub-agents stream their responses to keep the tunnel alive#597
OisinKyne wants to merge 1 commit into
mainfrom
oisin/streaming

Conversation

@OisinKyne

Copy link
Copy Markdown
Contributor

Problem to be solved

Cloudflare tunnels timeout on 100s of no data. Streaming an agents response may resolve that, but it also may put more complexity on the client side caller, we should consider if we need to improve paybe buy.py pay command to handle it if we haven't already. (I think its dealt with at the verifier side)

Summary

obol-stack/:

  • internal/x402/verifier.go — statusRecorder.Flush() + Hijack()
    so SSE chunks actually stream (the bug from the previous
    turn).
  • internal/x402/verifier_test.go —
    TestVerifier_HandleProxy_StreamsSSEChunks (over-the-wire pacing
    assertion) + TestVerifier_HandleProxy_NonStreamingResponse.
  • CLAUDE.md:
    • Added agent to obol sell and a new obol buy inference row
      in the CLI tree.
    • Added an "Agent-backed offers" paragraph and a "Prefer
      stream: true for OpenAI-compatible paid endpoints" block in the
      Monetize Subsystem, explicitly calling out the Cloudflare
      tunnel ~100s idle window and citing the test as the regression
      sentinel.
    • Added obol sell agent and obol buy inference to the CLI
      flag-reference line.

@bussyjd

bussyjd commented Jun 5, 2026

Copy link
Copy Markdown
Contributor

Superseded by the consolidated integration PR #598, which merges this branch together with the other PRs ≥ #593 into integration/pr-593-plus for a single reviewable diff. Review hardening fixes ride in #599 (targets the integration branch). Closing in favour of the consolidated review — reopen if you'd prefer to land this one standalone.

@bussyjd bussyjd closed this Jun 5, 2026
bussyjd added a commit that referenced this pull request Jun 5, 2026
…sell-agent facts

Terse rewrite of project CLAUDE.md (42725 -> 41797 bytes) corrected against the live codebase. Preserves all invariants, the 14 pitfalls, and flag warnings; adds #597's stream:true / statusRecorder.Flush guidance and agent-backed-offer (port 8642) facts so the compressed doc loses nothing rc12 ships.
OisinKyne added a commit that referenced this pull request Jun 5, 2026
…ning) (#600)

* Work on openapi spec for services

* Improve 402 html page

* Update storefronts

* feat(buy-x402): add --set-default so the agent self-adopts paid/<model> as primary

After a persistent inference buy publishes paid/<remote-model> in LiteLLM, the
agent adopts it as its own primary chat model in-pod via native
'hermes config set model.default' (atomic write, per-request re-read, no restart,
no host CLI, no new RBAC). Includes a LiteLLM /v1/models existence guard, an
auto-refill safety warning, and a PyYAML fallback writer.

Validated by a design+adversarial workflow and a live CLI smoke against a running
obol-agent: buy --set-default flips config.yaml model.default to paid/AEON-7/... and
the next agent chat settled via the x402-buyer pool (spent 0->1) with no restart;
rollback verified.

* Have agent stream responses to keep the tunnel alive

* security(x402): SRI-pin the Scalar bundle on the public /api page

The /api OpenAPI reference is served over the public tunnel and pulls the
@scalar/api-reference bundle from jsdelivr. The integrity hash was left empty
in phase 1, so the browser executed whatever the CDN returned, unverified.

Populate scalarBundleSRI with the sha384 of the pinned 1.34.0 bundle so a
tampered CDN response is blocked. Comment updated to stress the hash must be
re-derived in lockstep with every scalarBundleVersion bump.

* fix(buy-x402): run --set-default existence guard before auto-refill warning

The 'paid/<model> not selectable in LiteLLM' guard ran *after* the
no-auto-refill WARNING. A model that LiteLLM would refuse still printed a
scary 'every chat turn fails when the pool empties' warning describing a
primary-model failure mode that cannot occur when the default was never
switched. Reorder so we refuse first and only warn when we are actually
about to adopt the model.

* security(x402): sanitize ServiceOffer-sourced tokens in 402 copy-paste commands

spec.model.name and metadata.name flow from the ServiceOffer CR into
copy-pasteable 'obol buy inference ...' commands rendered on the public 402
page. A hostile or fat-fingered offer could smuggle shell metacharacters into
a command a reader might paste. Add sanitizeDisplayToken at the render
boundary: CR-sourced tokens must match the model-id/k8s-name charset
(^[A-Za-z0-9._:/-]+$) or collapse to the existing safe placeholder. Real ids
like qwen3.5:9b and anthropic/claude-3-5-sonnet-latest pass through unchanged.

* docs(claude): compress and correct CLAUDE.md; fold in #597 streaming/sell-agent facts

Terse rewrite of project CLAUDE.md (42725 -> 41797 bytes) corrected against the live codebase. Preserves all invariants, the 14 pitfalls, and flag warnings; adds #597's stream:true / statusRecorder.Flush guidance and agent-backed-offer (port 8642) facts so the compressed doc loses nothing rc12 ships.

---------

Co-authored-by: Oisín Kyne <oisin@obol.tech>
Co-authored-by: bussyjd <jd@obol.tech>
Co-authored-by: bussyjd <bussyjd@users.noreply.github.com>
@OisinKyne
OisinKyne deleted the oisin/streaming branch July 1, 2026 12:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants