Skip to content

v0.12.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 17 Jun 02:41
· 1332 commits to main since this release
v0.12.0
cc93b66

[0.12.0] — 2026-06-17

v0.12.0 — "Durable & Honest." Crash-resumable runs, human-in-the-loop approval gates, typed structured output across every provider, and a set of honesty defaults (memory off, grounding opt-in, debrief off the critical path) that stop the framework from over-promising or silently spending tokens.

Added

  • Durable execution — kill it, resume it. .withDurableRuns({ dir?, checkpointEvery? }) persists a lossless serialized KernelState per iteration to a SQLite run store (opt-in; zero overhead when off). agent.resumeRun(runId) reconstructs a crashed or paused run from its last checkpoint and continues it to completion in a fresh process — without replaying completed tool work; agent.listRuns({ status? }) enumerates persisted runs. A config-hash guard (system prompt + provider identity) rejects resuming under a drifted config with DurableConfigMismatchError (and DurableRunNotFoundError for unknown runs). Cross-process crash-resume is verified by a subprocess hard-kill (SIGKILL/exit-137) e2e proving the on-disk checkpoint reconstructs and finishes the run. withProgressCheckpoint is clarified as the lighter plan-level hint; durable runs are the crash-resume path.

  • Durable human-in-the-loop approval gates. .withApprovalPolicy(...) pauses a run before a flagged tool call instead of blocking a process in memory: the run terminates with terminatedBy: "awaiting-approval", the paused KernelState is checkpointed to the durable run store, and a pending ApprovalDecisionRef is surfaced on the result/stream. agent.approveRun(runId, ...) / agent.denyRun(runId, ...) resume the run from that checkpoint (approve → execute the gated call, deny → observe a denial and continue); agent.listPendingApprovals() enumerates what's waiting. An onApproval convenience callback covers the in-process case. Works on both run() and runStream(), and reuses the crash-resume infrastructure rather than a separate in-memory gate.

  • Typed structured output, every tier. .withOutputSchema(schema) returns a typed result.object (with result.objectError on failure); agent.streamObject() yields { object: DeepPartial<T> } deep-partial snapshots as the model writes. Schemas come from any Standard-Schema library — Zod, Valibot, ArkType, or Effect Schema — bridged to per-vendor JSON Schema (top-level arrays and lenient degradation supported). Live-verified across Anthropic, OpenAI, Gemini, and local Ollama (qwen3.5 / gemma). New exports include extractStructuredOutput, toSchemaContract, groundedExtract, parsePartial, StructuredOutputError, and the @reactive-agents/svelte / @reactive-agents/vue useStructuredObject bindings.

  • Always-on scaffold-leak guard. A standalone check (detectScaffoldLeak) rejects terminal output that echoes framework internals ([STORED:], _tool_result_N, "compressed preview") instead of synthesizing a real answer. Separate from grounding, ~zero false-positive rate, and stays on regardless of the grounding opt-in.

  • Cortex studio — durable & structured surfaces. The Cortex companion app gains: launch durable runs + approval gates from the config panel, an app-wide interactive approval-toast prompt, a structured-output schema editor with a typed-object viewer, budget-cap + evidence-grounding config controls, and the reasoning kernel enabled by default with a clear toggle.

Changed

  • Memory is now OFF by default. new ReactiveAgentBuilder() no longer enables the memory layer implicitly (GH #122) — agents are stateless unless you opt in. The balanced() and intelligent() presets enable it explicitly, and .withMemory(...) turns it on for any agent. This stops surprise SQLite writes and recall costs on simple agents. Migration: if you relied on implicit memory, add .withMemory() (or use a preset that includes it).

  • Post-answer debrief no longer blocks run(). The LLM debrief synthesis is forked off the critical path — measured 4.7s of a 9.8s frontier run (48%) was the agent synthesizing a debrief after the answer was already produced (~6s local; GH #143). run() now returns the instant the answer + a deterministic fallback debrief are ready (result.debrief, always present on memory runs). The rich LLM debrief synthesizes + persists in a background fiber; await result.debriefRich() only when you want it. agent.getLastDebrief() upgrades to the rich version once it resolves, and agent.dispose() joins pending debrief fibers so a short-lived run(); dispose() never drops the persist. Net: ~46% lower perceived run() latency on memory-enabled runs. A tier-aware skip avoids the debrief synthesis entirely on small local models where it isn't worth the round-trip. result.metadata.tokensUsed now reflects the answer's cost (debrief tokens are background; surfaced via debriefRich()/telemetry).

  • Evidence-grounding is now opt-in (off by default). The previous always-on numeric grounding check false-flagged correctly-formatted figures (e.g. $62,578) when the tool observation was compressed or reformatted — firing a failed at evidence-grounded verifier warning on correct answers. Enable per-agent via .withGrounding({ mode: "block" | "warn" }). When enabled it grounds against the full tool data (resolved via storedKey→scratchpad, not the compressed preview) with rounding/format tolerance ($62,578 ≈ 62578.12 ≈ $62.5k). mode: "warn" is advisory (surfaces a verificationWarning, ships the answer); mode: "block" runs one corrective re-synthesis then degrades to warn — it never hard-fails a run. The prose claim-grounding pass (which false-rejected legitimate summaries at 64–73%) was removed.

  • Hooks accept plain functions. .withHook() handlers may now be plain sync/async functions — no Effect wrapper required (the Effect form still works). Additive, no migration needed.

  • Observability config unified under .withObservability(...). The previously scattered observability surfaces collapse to a single configuration entry point.

Fixed

  • Cortex: budget limit of 0 now means "unset" rather than a zero-cap that triggered a false budget_exceeded; saved/gateway agents correctly thread durable HITL + structured output.
  • Reasoning: code-action sandbox now emits observation.tool-result so .on() observers see its tool output (closes the FM-I tool_call sub-gap, #195).
  • Runtime: the durable RunStore read path now mkdirs its parent directory so listing runs from a fresh agent doesn't throw.
  • Build/types: restored a green typecheck across reasoning/core/runtime/cortex (EntropyScoreLike contract, durable-HITL sentinel reason, structured-output schema variance, durable test fixtures, exactOptionalPropertyTypes), and eliminated a concurrent apps/cortex/ui/build double-build race that could fail the release build or ship a partial Cortex UI.

Changelog

All notable changes to Reactive Agents will be documented here.

The format follows Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Pending release notes live in .changeset/*.md and are aggregated into a versioned ## [x.y.z] section by scripts/release.ts when the tag is pushed. Do not hand-edit released sections.

Assets 2
Loading