preset

preset

Composition, not a capability. This package has no plugin of its own — it assembles the others.

preset.Plugins(cfg) returns the plugin list that reproduces agentcore.New(cfg), and preset_test.go proves the two agree field by field via Describe(). If the plugin surface ever drifts from Config, that test fails.

preset.Full(cfg, opts) is that list plus the capabilities Config has no field for: spill, jobs, repeat_guard, session_query, plus opt-in todo, subagents, advisor, ask and finish guards. The split is load-bearing — Plugins is pinned to agentcore.New parity and must stay a pure mirror of Config, so the capabilities that answer to no Config field live one layer up. Full is where a deployment starts; Plugins is where the parity proof lives.

Every Options field degrades to off, never to wrong. A nil Spill leaves the loop's own head+tail truncation in place rather than minting locators into a store that dies with the process — the locator is written to the durable session log, so an in-memory store would make a resumed run read its own spill back as not-found.

Model Experience

None. Composition happens before any model call.

Impact on the agent

It is the starting point for a custom agent. Take the default list, drop or replace what you want to change, append your own, and build:

ps := preset.Plugins(cfg)
ps = preset.Replace(ps, agentcore.PolicyDenyAll()) // replace the permission policy
ps = append(ps, myCapability{})                  // add a capability
agent, err := agentcore.Build(ps...)

The engine owns turn scheduling. Plugins configure capabilities through registry setters, hooks and per-run extensions; they cannot replace the loop.

Two tests here carry the architectural claims:

• TestPluginCompositionMatchesNew — the two entry points cannot drift.
• TestEjectRemovesEveryTrace — building the same agent with and without a capability leaves no sign of it in the one without: no tool, no name in the extension list. This is the test that fails if someone re-couples the loop to a plugin by name, because the loop would keep doing the work with the plugin gone.

Known limitations and deferred work

• Parity is checked through Describe(), so a Config field that Describe does not report could drift undetected. The test guards against a vacuous pass by asserting the dump is not thin, but that is a proxy, not a proof.
• Replace appends. A replacement lands at the end of the list rather than in the original position. Seams are keyed and hooks are prioritized, but extension stop guards run in registration order. Preserve that order when replacing guards. Full replaces its goal/memory options in place.

Tracing is supplied through NativeRun.Telemetry or telemetry.WithContext; there is no observer plugin or provider-decorator registration. Native recovery validates checkpoints and effect receipts directly.

Full additionally accepts GoalLifecycle (an opt-in goal plugin configuration), AdvisorOptions (cadence/caps for NativeAdvisor), and ConsolidateMemory (native rollout distillation when the configured store implements the optional durable consolidation contract). Zero values preserve the earlier composition behavior.