Eclipse Brake
Third-party halt authority for agent fleets: runtimes poll, a BRAKE verdict halts within the poll interval, and signed halt receipts prove the stop — or the episode closes UNPROVEN-HALT. LOCAL/simulator scope only — local policy endpoint + fixture runtime, no live fleets.
status: in-developmentclass: mcpTrust & ReceiptsAgent Safety & Oversight
Vendor: Cyan Eclipse Technologies
What it does
- Novel mechanism — runtimes poll a local policy endpoint; a BRAKE verdict halts the fleet within the poll interval, and the runtime returns a signed halt receipt proving it actually stopped (SPEC §2.1, the mechanism the bench verified verbatim via F-B14). Brake has no push channel, no remote kill, and no verb inside any runtime — a runtime that never polls cannot be halted by Brake, and Brake never claims otherwise.
- HALTED only on proof — a fleet is HALTED only when every enrolled runtime has a valid, in-window, counter-signed halt receipt bound to the verdict's hash. Absence of a receipt is never evidence of a halt: the complementary terminal state is UNPROVEN-HALT, stated on the episode pack's face and never softened. Late receipts are recorded as defects and never amend a closed episode; a stale verdict closes its episode STALE-UNDELIVERED.
- Offline-recomputable episode pack — the pack (fleet record, verdict token, poll records, receipts, defects, close statement, hash chain) recomputes via receipt_verify executing the full §6.4 procedure: fleet-record hash recomputed from the full record in the pack and bound to the signed chain, chain and poll-record signatures verified, receipts bound to recorded poll deliveries, post-halt attestations recomputed against the Brake-side fixture task log exported in the pack, and the close state derived mechanically from the enrolled roster — a partial fleet's pack cannot be re-hashed into HALTED (re-bench D1 forgery now fails at verifier step 1 with no key).
- Named refusals + configuration fence — duplicate receipts, receipts from non-enrolled runtimes, no-poll receipts, poll-time mismatches, excess latency, and malformed inputs are all refused by name (RECEIPT_DUPLICATE, RECEIPT_RUNTIME_NOT_ENROLLED, RECEIPT_NO_POLL_RECORD, RECEIPT_POLL_TIME_MISMATCH, RECEIPT_LATENCY_INVALID, RECEIPT_MALFORMED, and the §5.4 set); startup hard-refuses fenced configurations (CONFIG_FENCE_REFUSED:<class>, §7.3 — non-local endpoints, real credentials, production identifiers, non-test value_class, non-mock relay modes all hard-exit code 2). Disjoint surfaces: operator = exactly 5 tools, runtime = exactly 3; cross-surface calls return the identical unknown-tool refusal as a nonsense name.
- Honest scope — LOCAL/simulator ONLY: local stdio MCP server (Python stdlib only) + fixture fleet runtime on an integer fixture clock; signing is a fixture HMAC stand-in for the Herald pattern (fixture keys are derivable from the published source — the labelled §7.4 fixture discipline, scope-inherent and disclosed by the bench). No real fleet is halted, polled, or contacted; no live-fleet path exists in the tree (no socket/urllib.request/requests/http.client imports). Halting real fleets is behind a named counsel gate — v0.1.2 contains no dormant live path.
Tools
fleet_enrol · verdict_issue · fleet_status · episode_pack_get · receipt_verify · policy_poll · halt_receipt_submit · runtime_reenrol
Platforms
- MCP over stdio
Certification
Bench verdict: CERTIFIED (LOCAL/simulator)
Certified scope: CERTIFIED (LOCAL/simulator) — Verity re-bench of v0.1.1, 2026-10-26 (argus101-program-20261027/rebench-brake.md) fixed and independently verified every defect from the v0.1.0 FAIL bench (D1–D10), including both blockers — the P9b partial→HALTED forgery fails at verifier step 1 with no key, and the submit path cannot put Brake into a state its own verifier rejects — returning CONDITIONAL on a single condition C1 (ND1). Verity spot re-bench of v0.1.2, 2026-10-27 (argus101-program-20261027/spot-brake-012.md) CLOSED C1: all five close-statement strip variants (statement deleted or None, final flag kept, stripped, or double-stripped) return verified: false, error BRAKE_VERIFY_FAIL:5, step 5, downgraded: true per §6.6; honest interim and honest final (HALTED and UNPROVEN-HALT) paths verify unchanged 3/3 with no downgraded key. ND2 (per-runtime re-enrolment acts under one CLEAR) fixed and verified 3/3 with duplicate/fresh-POLLING/D8 guards holding; ND3 ACCEPTED as spec-correct (§4.5/§4.6/§6.6). No new findings. Builder suite 51/51 (fresh process); spot bench own probes 13/13. Scope never exceeds LOCAL/simulator: local policy endpoint + fixture runtime, no live fleets.
Not tested:
- Live fleets — not tested and not claimed: v0.1.2 runs a local policy endpoint + fixture runtime only; no real fleet is halted, polled, or contacted, and no live-fleet path exists in the tree
- Production signing custody — signing is a fixture HMAC stand-in for the Herald pattern; fixture keys are derivable from the published source (labelled §7.4 fixture discipline), so a presenter holding the fixture keys can mint indistinguishable artifacts — scope-inherent, disclosed by the bench
- Hosted fleet shape — waits Gate B1; the local package is the reference implementation and no hosted endpoint exists or is claimed
Live state note: Bench evidence: argus101-program-20261027/spot-brake-012.md (Verity spot re-bench of v0.1.2, independent fresh-process, own probe script — no builder test code reused; ND1/C1 closed 5/5 strip variants → FAIL:5, honest paths unchanged 3/3, ND2 fixed, ND3 accepted spec-correct, no new findings). Prior chain: rebench-brake.md (v0.1.1 re-bench, CONDITIONAL), bench-brake.md (v0.1.0 FAIL), fix claims fix-brake-011.md / fix-brake-012.md. These benches issued no test-bench/results scorecard; scorecard_ref is null and no scorecard is claimed or hosted for this product.
Pricing
PRICED — $199/mo per fleet, via the Eclipse Bazaar bench-order/intake path only. No checkout, payment, or buy-now link exists for this product, and no Stripe object exists for it in this pass: intake orders are reviewed and confirmed by the operator before any payment link is sent. The hosted fleet shape waits Gate B1; the local package is the reference implementation. Repository is PRIVATE (priced-product policy).
Install / verify
No install command is published for this product in the catalog or its registry record.
Official MCP Registry name: io.github.CYAN-ECLIPSE-TECHNOLOGIES-LLC/eclipse-brake-mcp
Registry listing: https://registry.modelcontextprotocol.io/v0.1/servers/io.github.CYAN-ECLIPSE-TECHNOLOGIES-LLC%2Feclipse-brake-mcp/versions/0.1.2
Repository: https://github.com/CYAN-ECLIPSE-TECHNOLOGIES-LLC/eclipse-brake-mcp
Verify any MCP server in your client configs against this catalog with bazaar check (Eclipse Bazaar MCP / CLI — see llms.txt).