What’s New in v2

Looking for v1.0 docs? They live on the release/v1.0 branch on GitHub. This page describes what changed since.

v2.1 (2026-07)

A minor release. 44 commits on top of v2.0.0 — no breaking API changes, additive behavior across verification, LLM I/O, compute routing, and cost accounting.

Headlines

  • Verification gates walk child sessions and retry once — The DATA/EXEC/LLM gates on TaskOrchestrator now traverse subagent sessions when scoring a task, so evidence produced inside a spawned subagent counts toward the parent’s gate. A subagent that returns success but fails the produces_uris gate gets one continuation turn seeded with the gate’s complaint before the failure bubbles up. See Task Orchestration → Verification.
  • Trajectory-aware verification + session_end event — The verification gate now reads the full session trajectory (tool calls, artifacts, timing) via a new session_end event in the provenance log, rather than only the final result. See Provenance Log Schema.
  • Multimodal attachments ship exactly once per session — Each PDF/image gets a UUID artifact_id at the agent-loop boundary; LLMClient sends the full base64 block on first pass and a text-fallback marker on every subsequent pass. Anthropic prompt-cache layout stays byte-stable. Subagents get their own isolated set. See Tools → PDF & image inputs.
  • Anthropic native block format for images/documents — Instead of a provider-neutral shape, image/document blocks now go over the wire in Anthropic’s native format when the provider is Anthropic. Other providers unaffected (LiteLLM continues to normalize).
  • LLM-agnostic caching and reasoning in the agent — Cache-control emission and reasoning-token accounting are now handled at the agent layer, not per-provider. Gemini implicit cache still hits automatically on 2.5+; Anthropic explicit cache_control still emitted for Anthropic.
  • Compute router looks at the local daemon before going to cloud — When a workload fits the current environment, the router keeps it local instead of provisioning a SkyPilot cluster. New env-composition and compute-threshold heuristics decide the split. See Cloud Compute → When SkyPilot vs local.
  • Aggregate cost rollup across LLM, compute, and storage — Costs are tracked on separate axes and rolled up per run and per CellResult. sky.cost_report powers the compute axis. See Configuration → Cost caps.
  • Layered config surface (--config, --set) — CLI now accepts a layered --config <path> plus --set key=value overrides, with kill-switch caps that hard-limit tokens/cost regardless of config. Precedence: --set > --config > env > yaml > default. See Configuration → Configuration layers.
  • ask_user replaces pause_for_user — What was an out-of-band gate is now an agent tool the model can call itself when it needs input. See Tools → ask_user.
  • Scientific web search tools — web_search_openalex, web_search_crossref, and companions surface as first-class tools alongside the existing web search. Research subagent now ingests artifacts (PDFs, figures) rather than URLs only.
  • service_search token-fallback — When the exact-substring pass returns no matches, a second pass tokenizes and singularizes both query and haystack. Rescues multi-word queries whose tokens span fields (molecular dynamics) and plural queries against singular haystacks (simulations vs simulation). Result rows carry match_mode and matched_tokens.
  • Session-scoped exec/fetch logger — Log lines from exec_shell / fetch_url are attributed to the session that spawned them, so parallel sessions no longer interleave in one flat log.
  • Registry adds — PyTorch dockerfile; RCWA image ships S4.so in site-packages with an import verification step.

Migrating from v2.0

No breaking changes to the Python API. Two surface renames worth knowing:

  • pause_for_user gate → ask_user tool. Existing sessions saved before the switch replay fine; new sessions call the tool directly.
  • BaseTool schema now emits parameters instead of input_schema when serialized (aligns with the Anthropic SDK’s tool-use format). If you consumed the schema JSON externally, update the key.

Everything else is additive. --config and --set are new CLI flags; existing env vars and ~/.sciagent/config.yaml continue to work with the same precedence order.

Component changes at a glance

Area v2.0 v2.1
Verification scope Final result of the parent session Full trajectory + walks child sessions + one corrective retry on gate failure
Multimodal I/O Re-sent every turn Sent once per session via artifact_id; text-fallback marker on replay
Anthropic image/PDF wire format Provider-neutral block Native Anthropic block format
Compute routing Threshold → SkyPilot Local daemon first if it fits; env-composition heuristic decides
Cost tracking Per-tool Aggregate rollup across LLM / compute / storage axes on runs and CellResult
CLI config Env + yaml Env + yaml + --config + --set + kill-switch caps
User-in-the-loop pause_for_user gate ask_user agent tool
Web search Generic web search + openalex, crossref, and companions
service_search Exact substring only + token-fallback with plural handling; match_mode per row

v2.0 (2026-05)

v2.0 is a major release. The codebase grew from a single-process agent with file I/O into a multi-substrate runtime with cloud compute, durable provenance, and registry-backed orchestration. v1.0 stays available as a separate marketing-tag artifact (tag v1.0, branch release/v1.0); semver starts at v2.0.0.

Headlines

  • Cloud compute via SkyPilot — Run scientific simulations on cloud clusters without leaving the agent. New tools: compute_run, compute_exec, compute_cluster, materialize, materialize_workspace. See Cloud Compute.
  • Durable provenance log — Every tool call, compute job, artifact, and verification result lands in an append-only JSONL log per session. Cross-LLM verifiable — a fresh model from any provider can audit a session it didn’t run. See Provenance Log Schema.
  • Task orchestration — A unified registry for in-flight work (task_index) covering cloud jobs and background subagents alike. New tools: task_list, task_get, task_wait. See Task Orchestration.
  • Background subagents with checkpoint & resume — spawn(background=True) returns a task_id; per-iteration checkpoints survive crashes; the parent gets a 3-way resume prompt on the next spawn. See Task Orchestration → Checkpoint & resume.
  • Two new subagent kinds — compute (cloud-job orchestration with token isolation) and analyze (post-job derivation: plots, statistics, light fits). The verifier subagent now reads the durable provenance log, so a different model can audit a session it didn’t run. See Architecture → Sub-agents.
  • Updated default models — Sonnet 4.6 / Opus 4.7 / Haiku 4.5 across the tier system. Provider-agnostic via LiteLLM.
  • Two-layer Python config — AgentConfig (agent-loop concerns: model, tokens, iterations, compaction) is now joined by a separate CloudConfig (cloud-side concerns: cost gate, workspace store, autostop default, subagent warm-resume). AgentLoop accepts both as independent kwargs. Existing env vars and ~/.sciagent/config.yaml keys still work; precedence is env > config dataclass > yaml > built-in default. See Configuration.
  • New scientific service — paraview (multi-arch with EGL) for post-processing OpenFOAM and other simulation outputs. The OpenFOAM image now also ships with a SWAK4Foam-extended variant for field-processing, used internally when a case needs it.
  • First cloud-compute case study — Datacenter CFD with OpenFOAM reproduces published results end-to-end on a SkyPilot cluster.

Migrating from v1.0

There are no breaking API changes — v1.0 code that called AgentLoop, SubAgentOrchestrator, or the core tool registry continues to work. The v2.0 surface is additive:

  • The default model has changed from claude-opus-4-5 to claude-sonnet-4-6 for the scientific tier. To pin to v1.0 behavior, pass --model anthropic/claude-opus-4-5-20251101.
  • The compute_*, materialize*, task_*, bg_*, and monitor* tools are new — main agents pre-v2.0 didn’t have them. They’re registered automatically; nothing to opt into.
  • pip install 'sciagent-cli[cloud]' is required for SkyPilot. The base install does not pull it in.

If you have v1.0 sessions saved in .agent_states/, they remain loadable — the session schema is back-compat.

Component changes at a glance

Area v1.0 v2.0
Compute backends Local Docker only Local Docker + SkyPilot (managed jobs + cluster mode)
Output handling Local files Cloud-agnostic URIs (S3/GCS/Azure/R2/OCI) + per-session workspace bucket
Long-running work Foreground subagents Background subagents + checkpoint/resume + task_index
Audit trail Inline tool logs Durable per-session JSONL log (provenance.jsonl)
Verification Three-tier gates (data/exec/LLM) Same gates + cross-LLM verifier reading the provenance log
Subagent kinds explore, debug, research, plan, general, verifier + compute, + analyze
Scientific services openfoam, … openfoam (with SWAK4Foam-extended variant), paraview, …
Default scientific model claude-opus-4-5 claude-sonnet-4-6