Continue and steer an investigation

Continuation and steering apply to minimal studies.

Continue after a deadline or change direction

Open the study in Autonomous research, or open its experiment monitor. Choose Continue investigation, describe the follow-up, choose a new wall-time budget and select working files to inherit. Choose Prepare directly to review the normal study brief and agent/model selection, then Start research. Preparation alone starts neither an agent nor a simulation.

The new study links to its parent. Selected programs/readers are copied to research/inherited/. A compact continuation_input/brief.json summarizes counts, recent unreviewed notes and pending reviews. A hash-verified snapshot outside the writable research folder preserves the original files and a record summary. Large raw results stay at the parent’s receipt paths; move/copy them explicitly if using a different machine. References do not grant new scientific acceptance. Prior exploration stays exploration, and old observations cannot satisfy a new prospective commitment. No prior method approval is automatically imported. Reuse its scoped validation as context when explaining the next method, instead of repeating irrelevant tests.

The original hypothesis, instructions, reviews and deadline are unchanged. Each phase has a separate budget and may select a different agent/model. Required instrument rules are inherited unless explicitly supplied for a new CLI phase. The parent should remain available for receipt and raw-data inspection. If it is still active, the handoff captures recorded context at the new phase’s launch; later parent results are not silently imported.

CLI equivalent:

simjecture study --campaign runs/phase-2 --continue-from runs/phase-1 \
  --instructions-file next-phase.md --inherit-file calculation.py \
  --inherit-file analysis.py --backend builtin --model deepseek-flash \
  --provider-config /private/provider.json --wall-seconds 7200

Omit --hypothesis-file to retain the original question. Specify a different one for an explicitly revised question. Configure the execution backend appropriate to the host. For a new phase the backend is checked afresh. Selected files are limited to 128 files, 1 MiB each, with a 16 MiB handoff including metadata. Hidden configuration, known credential filenames, symlinks and lab.py are excluded.

Prepare the continuation with an agent

The continuation dialog also offers Prepare with agent. You can leave the follow-up description blank if you want to discuss the direction first. Select initial files and a proposed budget, then choose that button to open the chat.

The interactive agent receives a parent-study context file: original question, previous instructions, limitations and review state, selected files, your guidance and the proposed budget. It can ask focused questions and call draft_study to prepare or revise the continuation brief. It may refine the inherited file list; the host checks every selected path against the parent’s research workspace. Parent lineage and the inherited instrument registry survive later chat turns and agent-written briefs. Prior observations remain context, not fresh evidence.

Prepare with agent also appears beside the parent-study link in an existing continuation conversation, so you can discuss a directly prepared draft later. If no model is configured, the parent selection is retained: choose the agent in chat and use that button again. The resulting brief appears in Autonomous research. Preparing in chat uses the selected interactive agent, but starts no autonomous research allocation; Start research remains a separate action.

Give advice during a study

Choose Send guidance beside the study controls. Add an observation, a paper URL, an alternative interpretation or a suggested experiment. It is queued durably for the next agent checkpoint. Operator guidance shows its timestamp and whether it has been included in a worker prompt. Delivery does not mean the agent has implemented the suggestion. Recent guidance is repeated after session recovery, and independent reviewers also receive the advice with its scope and timestamp.

simjecture steer --campaign runs/phase-2 \
  --message "Measure boundary Poynting flux independently; do not infer it from energy closure."

Guidance is advisory. It does not rewrite commitments, change acceptance rules, extend time or interrupt active simulations. Use the existing Pause/Stop controls for execution control. For new requirements, a new hypothesis or more time, use Continue investigation. This keeps the original contracts auditable while leaving the research agent free to plan its next experiments.

Review and status recovery

The built-in reviewer now records whether output was empty or truncated and its provider finish reason. It makes at most two transport attempts per invocation; a DeepSeek retry disables thinking after an empty/truncated response. Usage includes both attempts. Persistent failure remains an error, never approval. Invalid oversight retries back off up to 30 minutes while other research can continue.

Live status reconciles expired budgets and recorded supervisor identity. Final reports are written on supervisor exit, including an external-error exit. A continuation does not erase or silently resolve a parent’s pending review.

Conversation continuity

From Study details · Evidence & review, Continue investigation opens the workspace with the study selected. For a workspace-owned study, the next proposal stays in that conversation. If its agent is busy or it already has an unlaunched proposal, resolve that work before preparing another proposal. Earlier versions that inferred a launched brief are supported without rewriting their records.

Standalone CLI studies still create a conversation on submission. If the network fails after submitting a standalone continuation, check the conversation list for Continue: before submitting again: this preparation API does not yet provide an idempotent replay receipt. Repeated clicks while a request is pending are suppressed. An owned-study retry never silently overwrites its pending proposal.