Skip to content

v0.1.6 — Hermes Loop

Choose a tag to compare

View all tags
@ritsuKai2000 ritsuKai2000 released this 19 May 13:48
v0.1.6
579541c
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v0.1.6 — "Hermes Loop"

2026-05-19 · 48 PRs · 296 commits · 3 contributors

Hermes is live. An operator can join a SolverNet on Base Sepolia with the Hermes harness, pick a model from the dashboard catalog, and close the loop: claim → solve → deliver → get scored. The 2026-05-08 "evaluations silently broken" failure mode from v0.1.5 is demonstrably absent in v0.1.6 — verified end-to-end on real testnet with verdictCode: 1 on a real sympy fix.

But the loop closing is downstream of a much larger story. v0.1.6 ships the first end-to-end builder workflow, the Network explorer, an admission substrate that makes evals actually trustworthy, and the per-harness auth cutover that decouples claim-loop liveness from harness-specific credentials.

Highlights

🧩 Plug-in builder system — scaffold to chain in 60 seconds

The first release where a contributor can ship a SolverPlugin end-to-end without leaving the toolchain.

  • jinn create plugin (#210) — scaffolds a working plug-in package (two patterns: SolverType plug-in or runtime plug-in) modelled on the production swe-rebench-v2-runtime package. New skeleton compiles, tests pass, ready to edit.
  • jinn solver-plugins publish + revoke (#213) — packs the plug-in, uploads to IPFS, writes a plugin:<cid> record to the on-chain IdentityRegistry under the builder's agentId. Lazy Stage-1 identity bootstrap on first publish.
  • Ponder indexer for plug-ins (#214) — PluginPublication entity + plugin:* MetadataSet handler + PublishedArtifact base. Plug-ins are discoverable across the network.
  • Discovery API REST routes (#215) — five new /v1/discovery/* routes (plugin-publications, builder-artifacts, plugin-scores, launched-solvernets, claimable-tasks).
  • /build SPA page (#216, #228, #292) — the canonical operator-app surface for builders. Lists published plug-ins for a SolverNet, your own published plug-ins under your builder agentId, the shape reference, and a quickstart. Designed against DESIGN.md tokens.
  • Discovery auth gate (#227) — /v1/discovery/* routes get the same auth treatment as every other /v1/* daemon route.

A builder can today: scaffold → edit a skill → publish → see it appear on /build under their agentId → another operator installs it via jinn solver-nets add-plugin and the next task's signed envelope carries the plug-in CID in executor.plugins[].

🌐 Network explorer

The data layer becomes visible. Operators and builders can finally see what's happening across the network.

  • Full explorer surface (#181) — design, indexer schema/routes/enrichment, and the SPA. Network-level views of tasks, attempts, verdicts, operators, and plug-ins.
  • Solve-rate hero on Network view (#251) — leads with the load-bearing metric.

🔬 SWE-rebench v2 eval substrate (fufn)

The work that made today's verdictCode: 1 verification possible at all.

  • Eval admission + verdict-time substrate recheck (#234) — operators maintain a validated pool of scorable instances; the generator only posts admitted instances; the evaluator re-verifies HF row hash + Docker image digest at verdict time. Reproducible verdicts; no more silently grading against drifted substrate.

🔐 Per-harness auth + claim-loop readiness (vh74.2)

The bootstrap stops being gated on Claude auth. Each harness reports its own readiness, and claim-loop refuses to attempt with an unready harness.

  • Per-harness readiness registry (#248) — claude-code-learner, codex-code-learner, and hermes-agent each expose an isReady() snapshot via /v1/harnesses/readiness + /v1/harnesses/:name/readiness. Onboarding drops from 4 phases to 3 (no more "Sign in to Claude" step at bootstrap).

🚀 Hermes harness production-usable

The headline that named the release.

  • Hermes harness + model catalog (part of #292) — 10 canonical models in the dashboard dropdown: Anthropic Opus 4.7 (default, OpenRouter), Sonnet 4.6, Hy3 Preview, DeepSeek V4 Pro / Flash, Gemini 3.1 Flash Lite, Kimi K2.6, Owl Alpha, MiniMax M2.7, Hermes 4 405B.
  • Harness-aware PluginPicker (part of #292) — claude-code-learner is no longer force-included as a default plug-in when the operator picks Hermes (Hermes has its own learning loop).
  • Provider routing (#298) — <org>/<model> model ids auto-route to OpenRouter, so the catalog actually works without operators manually editing their hermes config.yaml.
  • MCP launcher path resolution (#299) — the jinn-client MCP server starts correctly in dev-layout installs, unblocking solution submission.

🏗 Bootstrap reliability (u34i / h74p / hjex.4 / 3nc5 / k1ng)

  • Bootstrap reliability stack (#262, #275, #279, #237, #238, #255, #257) — gate+transfer parity, faucet/gate single source, getCode retry, panel auto-continue, harness-readiness holder, fresh-Safe race retry, setAgentWallet revert reason surfacing, Stage 1 bindResult narrowing, Docker stage copies docs.
  • One-shot funding + Tenderly default RPC (#292) — operators send 0.020 ETH once; daemon doesn't re-prompt. Tenderly gateway replaces rate-limited sepolia.base.org. Shared-RPC panel warning with CTAs when on the bundled key.

🛤 Two-train release cadence (qlol)

  • Promote-main workflow + canary trigger (#252, #253) — every push to next ships a @canary to npm; named-stable cuts trigger npm @latest + auto-FF of main. Handbook + CLAUDE + hotfix runbook updated.

📜 Canonical docs

  • PRINCIPLES.md (#230) — privileged canon. Agents read it at session start; all decision-making runs through it.
  • BRAND + GLOSSARY (#196) — builder-pitch learnings captured into canon.

🪲 Ghost-task floor (band-aid for #300)

Caught during the cut itself.

  • Backlog floor across all ingestion paths (#301 + #303 + #305) — fresh operators on v0.1.6+ skip the 9 known pre-pool-rebuild ghost tasks on Base Sepolia. Three pieces because the band-aid surface had three call sites (adapter default, main.ts shadow, DiscoveryAPI consumption). #300 is the proper-fix investigation.

Live verification

A3 closed end-to-end on Base Sepolia during the cut. Hermes-on-DeepSeek-V4-Flash produced a patch for sympy__sympy-27510 ("Printing multiplication by negative number with custom (infix) function not correctly parenthesized"); the patch applied, the FAIL_TO_PASS tests passed, and Op A's evaluator scored it verdictCode: 1. Full transaction hashes and envelope CIDs in log/decisions/2026-05-19-v0.1.6-stewardship.md.

Full changelog

feat

  • (#181) feat(ebu7): network explorer — design, indexer schema/routes/enrichment, SPA
  • (#209) docs(52x3): plug-in builder entry point — spec + plan
  • (#210) feat(et6s): jinn create plugin scaffold (two patterns)
  • (#213) feat(1pbc): jinn solver-plugins publish + revoke — on-chain plug-in registry via setMetadata
  • (#214) feat(attd): Ponder indexer — PluginPublication entity + plugin:* MetadataSet handler + PublishedArtifact base
  • (#215) feat(ttz8): five Discovery API REST routes per spec §6.5
  • (#216) feat(hfmf): /build SPA route + canonical /docs/build/ tree
  • (#251) feat(explorer): lead Network view with solve-rate hero

fix

  • (#227) fix(0nih): auth-gate /v1/discovery/* daemon API routes
  • (#234) fix(fufn): SWE-rebench v2 eval admission + verdict-time substrate recheck
  • (#237) fix(h74p): retry safe-binding to absorb fresh-Safe race window
  • (#238) fix(hjex.4): surface setAgentWallet revert reason in attention.hint
  • (#255) fix(3nc5): narrow Stage 1 bindResult before reading txHash
  • (#257) fix(3nc5): Dockerfile build stage copies client/docs/ for /build SPA
  • (#262) fix(u34i): bootstrap reliability stack
  • (#279) fix(k1ng): setAgentWallet retries on returned ok:false, not just thrown errors
  • (#292) fix(u34i): post-bootstrap operator-app polish — Tenderly default RPC, one-shot funding, discovery holder, /build design, Hermes catalog, harness-aware PluginPicker
  • (#298) fix(hermes): infer --provider openrouter from <org>/<model> model ids (closes #293)
  • (#299) fix(hermes): inject JINN_NETWORK_TOOLS_CLIENT_ROOT into MCP server env (closes #294)
  • (#301) fix(mech): bump Base Sepolia TaskCreated backlog floor past v3 pool rebuild (#300 band-aid pt.1)
  • (#303) fix(mech): remove main.ts shadow of TaskCreated backlog floor (#300 band-aid pt.2)
  • (#305) fix(mech): apply backlog floor to DiscoveryAPI candidates too (#300 band-aid pt.3)

refactor

  • (#140) refactor: rename claude-code-learner → learner + Hermes design docs
  • (#174) refactor: solution envelope role schema
  • (#212) refactor(nghf): staged bootstrap — fleet-level Stage 1 + ensureStage1 / ensureStage1And2 entry points
  • (#228) refactor(gxuf): extract PanelCard to DRY /build SPA panel sections
  • (#248) refactor(vh74.2): per-harness readiness registry + daemon-level Claude gate removal

chore / release plumbing

  • (#164) chore(2cl.21): bd-mirror writes Sprint iteration + human Epic options
  • (#221) chore: sync main into next after current release
  • (#226) chore(52x3): simplify epic surface (post-merge cleanup)
  • (#252) chore(qlol): two-train plumbing — promote-main workflow + canary trigger
  • (#253) docs(qlol): two-train cutover — handbook + CLAUDE + hotfix runbook
  • (#256) chore(3nc5): bump client to v0.1.6 for the Monday cut

canon

  • (#196) canon: builder-pitch learnings in BRAND and GLOSSARY
  • (#230) canonical docs: introduce PRINCIPLES.md as privileged canon

test

  • (#173) test: real two-operator corpus-read e2e on Anvil fork (incl. x402 USDC payment)
  • (#224) test(r83r): reference plug-in + cold-start E2E acceptance gate
  • (#275) test(u34i): tier-2 regression coverage — Playwright sequential-state E2E, late-mount lint, boundary tests

docs

  • (#306) docs(stewardship): READY-FOR-CUT decision-log entry for this cut

other

Known issues / v0.1.7 follow-ups

v0.1.6 ships three band-aid patches in the floor stack (#301 + #303 + #305) for the ghost-task class. Each has a paired investigation issue framed as "understand the code first, then propose":

Issue Class What
#295 catalog schema Provider as a first-class field on LearnerModelOption. Replaces the regex inference in #298.
#296 plugin layout Per-task plugin-mount audit. Replaces the env injection in #299.
#297 test surface Real-Hermes E2E test shape. Stub-based E2E missed the bug classes caught live.
#300 admission Ghost-task class — symmetric solver-side admission filter, contract-side cancelTask, self-bumping floor anchored to pool updatedAt. Replaces the three-piece floor band-aid.
#302 harness layer codex-code-learner session-start hook missing on dev-layout install.
#304 CI hygiene transcript-watcher.test.ts > shutdown stops further dispatches — 8s waitFor times out under CI load.
#307 release process qlol cutover removed the holistic-review-at-main gate. Investigate restoring it.

Stats

  • Window: v0.1.5 → next HEAD (92ba5361)
  • 296 commits · 48 PRs · 466 files changed, +66,489 / −1,446
  • 3 contributors

🤖 Generated with Claude Code

Contributors

ritsuKai2000
Assets 2
Loading