Thursday’s AI news is about making agents operationally dependable, with OpenCode 2.0 leading: its shared service and composable tools change how integrations work, but existing plugins need migration. Around it, Liquid’s downloadable agent model offers a local tool-dispatch candidate with weaker vendor-reported coding results, TurnCost questions compression savings once prefix caching enters the accounting, and revocation research shows why acknowledgments cannot substitute for checking committed effects. CarryOn and DeRef tackle context reuse and tool-call judging, JointServe explores serving several models under memory pressure, and Jev and Figma add new inputs and shared design context. Eight states’ OpenAI lawsuits supply the policy counterpoint: safety claims remain claims that products must substantiate.

Lead — OpenCode 2.0 changes the integration surface

  • OpenCode 2.0 ships a shared service and code-mode tools — orchestration features arrive alongside migration work for existing integrations.

    • Terminal, web, and desktop clients share an automatically started background service; @opencode/client discovers it locally.
    • Code mode exposes MCP/OpenAPI tools as JavaScript functions, composing calls without one model round-trip per operation.
    • Background subagents and shell commands keep foreground work moving; the SDK exposes the CLI runtime.
    • Migration: V1 plugins and server integrations break; terminal settings move to global cli.json, not a project-local file.
    • Migration guide: preserve V1 and verify credentials, permissions, models, and MCP before relying on V2.
    • Flag: SDK development continues; the interpreter alone establishes neither measured speedups nor security guarantees.

    (HN 635 · 253c)

Policy & provenance

  • Eight states sue OpenAI over user-safety claims — coordinated enforcement scrutinizes product safeguards and safety marketing.

    • Separate suits allege harms to minors, elderly people, and other vulnerable users.
    • Attorneys general seek disgorgement and an end to allegedly deceptive safety statements; no damages figure was specified.
    • OpenAI says age-specific protections exist and additional safeguards are being developed.
    • Flag: allegations, not adjudicated findings or enacted rules; review safeguards and claims rather than infer liability.

    (Techmeme · Bloomberg Law)

Agent frameworks & tooling

  • Jev 1.14 adds experimental images, boxes, and longer context — typed-decision users get new inputs and a version-pinning decision.

    • jev-preview targets jev-1.14.0; jev-latest remains 1.13, changing in one week.
    • Image assets work across state/instructions/criteria; box output requires 1.14.
    • Docs: opt-in long context raises request and question limits to 262k tokens.
    • Flag: experimental features lack production endorsement; long context is slower and less accurate. Rankings are internal.

    (HN 301 · 93c)

  • Figma Specs gives design agents a shared reference — Markdown specifications connect design rules to the canvas instead of repeating them in prompts.

    • Specs stand alone or live on-canvas, referencing designs, diagrams, and code layers.
    • Libraries and guidelines provide centralized context; code layers move from private to public beta.
    • Test whether UI agents follow your design system, not merely produce attractive screenshots.
    • Flag: reference material steers a nondeterministic agent; it does not verify constraint enforcement.

    (Techmeme)

  • CarryOn organizes context around work products — reusable task objects preserve obligations and recent interactions instead of repeatedly summarizing everything.

    • Opus 4.7 OSWorld: partial score 43.9%→49.7%; binary completion 15.7%→19.4%.
    • Total tokens fall 51.9% and cost 35.9%, including context-management calls.
    • Attach evidence and reuse conditions to objects; a file’s existence does not prove completion.
    • Flag: one host model; the repository promises code after review, not an installable release.

    (arXiv 2610.18784)

  • DeRef supplies independent references for tool-call judges — judges can reject correct arguments when deriving defaults or recovering hidden evidence.

    • Seven judges show the gap; scaling to 27B does not remove it.
    • A separate model reads full history without the candidate, supplying expected values and sources.
    • Code: three benchmarks; controlled hidden-source AUROC reaches 98.40–99.59, not end-to-end success.
    • Flag: wrong references can reverse verdicts; retain deterministic checks and human review for consequential actions.

    (arXiv 2610.18497)

  • Revocation receipts do not prove tool effects stopped — bounded outcome contracts check enforcement and committed effects beyond successful event delivery.

    • Controlled receiver-to-gateway withholding commits 112/225 calls despite acknowledged, persisted revocation; normal delivery commits none.
    • Tested: eight gateways; two separate 16-host deployments; narrower MCP/OPA checks.
    • Test deny-state installation and independent filesystem/database outcomes, not just successful revocation requests.
    • Flag: guarantees cover declared gateways, calls, and observation windows only; this is not a claimed Keycloak vulnerability.

    (arXiv 2610.15991)

  • TurnCost challenges cache-blind agent savings — re-score actual trajectories before paying a second model to compress tool output.

    • Across 600 trajectories, median observations contain 118–286 tokens; turn count explains more cost variation than observation size.
    • Simulated near-ideal prefix caching yields 98.6% cached input; delegation increases cost in half the tested settings.
    • MIT replication package: CPU-only re-scoring without paid inference.
    • Flag: simulation omits TTL/eviction/backend movement; fewer turns may reduce success. Package predates October’s paper.

    (arXiv 2610.15093)

Models & research

  • LFM2.5-24B-A2B releases downloadable agent weights — test this reasoning-only local model on tool dispatch rather than assume superior coding.

    • Weights/card: 24B total / 2.45B active · 131,072-token context · GGUF/MLX/GPU formats.
    • Vendor LiveClaw: 48.09 versus Qwen3.6’s 38.19; SWE-bench Verified: 51.0 versus 66.2.
    • Vendor 184 tok/s: M5 Max, Q4_K_M, 4K input—not your hardware.
    • Flag: custom license adds a $10M revenue commercial-use threshold; not Apache-2.0.

    (HN 220 · 51c)

  • JointServe schedules weights and KV state together — multi-model agents can prefetch named experts instead of serially swapping whole checkpoints.

    • SGLang prototype combines partial weight residency, host-backed KV restoration, and transfer scheduling.
    • Reports 21.7% lower session completion time versus its strongest whole-model swap baseline; replay gains reach 49%.
    • Testbed: A800 80GB with constrained memory budgets and 1TB host RAM, not consumer hardware.
    • Flag: needs visible target-model call headers; gains depend on hardware/overlap. No deployable repository was verified.

    (arXiv 2610.19822)

All gathered items - what was cut and why (6)