简体中文 · Try a task in your assistant
For one computer, start with local SQLite. A new Goal using the unconfigured defaults shown in the SQLite developer exercise gets one SQLite authority database in the selected runtime. It holds the durable work state, including Todos, leases and operation receipts. SQLite may create WAL and shared-memory sidecars while open. Git files, the registry and host session artifacts live in their own locations.
Use loopx --format json todo list --goal-id <goal-id> from the project that
owns the Goal. Check authority_read.source_authority for sqlite_v0 on the
SQLite path. Existing Goals keep their recorded provider. Use the
storage selection and migration guide
before moving them. Your agent host runs the model and tools. LoopX retains
the work state used to decide what can continue after that host stops.
Use a stronger model when the route is open-ended, the evidence conflicts, or the failure calls the plan into question. Have it leave a reviewable contract with a versioned input, the action and its limits, an acceptance check, and a condition to return control. A cheaper model can then apply that contract to another eligible item. Run a deterministic check directly when it can do the whole job.
For example, first work out which test distinguishes a regression from a stale fixture. Then hand off the test, the exact fixture version and the expected result. A worker can repeat the check and attach evidence. A changed fixture, an unexplained failure, or newly required authority sends the task back for a decision. A write with an uncertain result needs the original receipt readback before another attempt.
Compare accepted results on the same tasks and budgets. Count cached and uncached input, output, extra calls and recovery, retries, elapsed time, and human review. Report any unknowns. The measurement guide shows how to compare model and context choices under declared conditions.
The current decision gives the agent an opening read and an index into an immutable capture, a saved copy of the applicable context. Read the required facts for the action at hand. When more detail is needed, use that capture's returned reader and follow its pages. For a compatible continuation, verified saved-file differences can reuse unchanged data. A lost context baseline can require a full read again. Validate scope and freshness before using a source to make a decision. See the operator guide.
| If a status looks current but… | Look for |
|---|---|
| the underlying source is old | When the source was last read, against its freshness window. |
| a required source was never read | Its missing or unbound read status. |
| a capture producer is failing | Last success time, separately from the last check time. |
| a read returned partly unreadable records | The unreadable count and affected source. |
| the summary covers the wrong registry or only part of the scope | Expected scope, actual coverage and typed omissions. |
The five-incident postmortem walks through these checks. A current timestamp can coexist with incomplete coverage, so inspect both.
The public eight-task, four-arm Context ON audit reports 3.0% for the first read body and 5.0% including CLI JSON. It sums 374,449 and 630,550 bytes against approximately 12.53 MB of full packets for 112 in-cycle continuations. The 213 total wakes also include 101 new-session or post-compaction baseline reads. End-to-end model tokens, cost and task outcomes have their own measurements. Aggregate and source revision.
The November 12, 2026, 22:40 CST EdgeBench snapshot lists Context ON ahead or marginally ahead on 7 of 10 rows. Rust is a small native-score edge (27.1801 vs 27.1757) and ties at two decimal places. OFF leads on Portfolio, Codec and OpenTTD. Only four reported pairs are within 15 minutes of one another near 18 hours. The other six involve historical runs, early stops or incomplete verification. Treat the tally as a description of this snapshot. Isolating the effect requires the exact versions, treatment, evaluator and budgets for both arms. The public aggregate also leaves ANN OFF's 3,746 QPS and both actual durations unverified. Revised score table and provenance.
Check the 2.3.0 release
and your installed host for the supported native path. Later changes on
main and research measurements carry their own revisions.
The remote authority task service documents an opt-in source-checkout candidate. It accepts authenticated claim, progress, lease and read operations through a service. PostgreSQL holds the authoritative transactions and receipts. A compare-and-set (CAS) check prevents an outdated writer from committing against a newer state. If a response is lost, read back the original operation under the same identity before retrying. Evidence for the candidate includes real-backend tests with independent clients, but enabling cross-host execution requires its own deployed host and service qualification. The current scoped-read candidate adds explicit Todo and audience checks. Its scoped listener rejects writes.
Open PR #8747 proposes broader resource access control for cross-person collaboration. The first candidate's coarse scope is tenant, Goal, actor and allowed operations, pinned to a store incarnation. The proposal intersects principal grants, credential scope, resource ACL, and applicable delegation/source permissions. Missing or unverifiable grants are denied. Resource access, execution leases and budgets continue to use their respective authority checks.
A human Gate can name one decision, such as whether to publish an already reviewed note. Keep the publication Todo waiting with the exact decision and target visible. An independent verification Todo can continue under its own admission. Approval resolves the named Gate. The publish step still needs applicable execution authority. Work graph and Gate scope.
The community assistant evaluation uses these sources for onboarding and boundaries (CQ01–03, CQ05, CQ08), versioned mechanisms and experiments (CQ11, CQ14), and collaboration and durable FAQ updates (CQ15, CQ20). Check the version in the question before quoting an answer.