QOpenJaws // Q_agents
OPENJAWS // OCI Q // Q_AGENTS // AGENT CO-WORK
Q.
OpenJaws for serious terminals.
Q_agents, co-work memory, routed tools, hosted access. One visible control layer.
OCI Q defaultAgent Co-Work liveRelease-checked qline.sitePublic + soak TerminalBench receiptsTyped traces + canonical signed receiptsShared preflight + seeded benches
Q // live modelReal local GLB asset
Public SurfaceOpenJawsTools, crews, routing, receipts. Q_agentsHelpers with co-work memory. Public repoGitHub-linked and benchmarked.
Qline is the front door. OpenJaws is the machine room.
Hosted keys and credits on the front end. OpenJaws, Q_agents, Immaculate, and Agent Co-Work behind it.
OpenJaws
The cockpit behind Q.
OpenJaws is the terminal workspace. Q is the default mind. Q_agents make it collaborative instead of lonely.
One crew, one visible deck, one source of truth for progress.
Hot registry plus phase memory so sibling agents stop re-discovering the same work every handoff.
Routing, pacing, worker state, and now typed trace receipts stay inspectable instead of implied. `/status` and `/immaculate` read the latest Immaculate trace summary directly, and `/status` also surfaces the latest Q benchmark trace.
Local tuning plus official public and repeated TerminalBench receipts, all checked against one shared OCI Q preflight lane and one typed Immaculate event contract with trace-backed canonical receipts, declared seeds, and reproducible preflight state.
Same-site qline deploy checks, scoped test lanes, and bounded hygiene gates now run before public ship calls are treated as clean.
Plans
Choose the lane.
- Light prompts
- BYO key lane
- Tight limits
- Hosted Q access
- Higher limits
- API key controls
- Routed work
- Team keys
- Benchmark headroom
Hosted lane
Sign up. Checkout. Keys. Usage.
Clean public access on the front end. Real entitlements stay server-side.
Q Access Console
Service Snapshot
Enough to feel the surface.
Longer sessions and real traffic.
Free if you want it. Still moderated.
Monthly usage preview40% of target
Agent Co-Work
One crew. Multiple terminals. Same thread.
Active terminals keep their own IDs, but the phase memory stays shared. That means cross-project help can keep the context that matters instead of resetting every handoff.
Project roots, runtime facts, and context IDs stay reusable.
Requests, handoffs, and delivered outputs stay attached to one thread.
New work can bind to the right saved phase on purpose.
Still Backend-Owned
The expensive truth belongs server-side.
Plans, subscriptions, overage, wallets.
Credits, resets, suspension, abuse controls.
RPM, TPM, audit receipts, and monthly reporting.
Benchmark snapshot
Measured, not invented.
Generated from benchmark receipts: D:/openjaws/OpenJaws/artifacts/q-bridgebench-dryrun-seed-smoke\bridgebench-report.json, D:/openjaws/OpenJaws/artifacts/q-soak-dryrun-seed-smoke\q-soak-report.json, D:/openjaws/OpenJaws/artifacts/q-terminalbench-official-public-20260416-circuit-fibsqrt-v2\terminalbench-report.json, D:/openjaws/OpenJaws/artifacts/q-terminalbench-soak-live-20260417-circuit-fibsqrt-v3\terminalbench-report.json, and D:/openjaws/OpenJaws/artifacts/q-bridgebench-live-20260415-nowandb\bridgebench-report.json.
Local audited-pack eval over Q bundle slices.
Best pack: unknownID: q-bridgebench-20260417T054645
30-minute bounded soak. 0/2 probes succeeded with 0 errors. OpenJaws p95 latency: n/a ms. Direct OCI-Q p95 latency: n/a ms.
Duration: 30 minErrors: 0
OpenJaws ran circuit-fibsqrt on OCI Q with 5 trials and 0 runtime errors. Mean reward: 0.0. The official leaderboard submission discussion is linked here.
Official TerminalBench 2.0 public taskTask: circuit-fibsqrtAgent: openjawsModel: oci:QStatus: submittedOfficial submission
Repeated TerminalBench soak over terminal-bench/circuit-fibsqrt. 2 cycles produced 2 total trials, 0 runtime errors, and 2 benchmark-failing trials.
Task: terminal-bench/circuit-fibsqrtStatus: completed_with_errorsRuntime errors: 0Benchmark failures: 2
disabled
Enabled: noSource: none