ContinuousClaudeV4.7: Skill for Claude Code

Orchestrate. Never research directly. Workers research inside Ouros; you plan, decompose, delegate, synthesize. Carry: hypothesis status, confidence scores, iteration number, ouros session names. No web searches yourself. Workers share state through Ouros sessions — not through your context window.

Pipeline: ASSESS → PLAN → PREPARE → EXECUTE → VALIDATE → EVOLVE → (loop to PLAN) Terminates when: all hypotheses confidence >= 0.8, context budget >= 75%, or user stops.

ASSESS

Read research question. Classify scope: focused — single question, 1-3 hypotheses, 2-3 iterations exploratory — broad topic, 5+ hypotheses, open-ended comparative — N alternatives to evaluate, structured comparison Estimate iteration budget: focused ~3, exploratory ~6, comparative ~4. Check available tools: ouros harness, exa, nia, bloks. Create ouros session name: {topic-slug} (e.g., "attention-scale", "auth-patterns").

PLAN

Hypothesis contract first — what would "understanding this" look like?

research_contract.json:

{ "question": "How do attention patterns change with model scale?", "scope": "exploratory", "iteration": 1, "ouros_session": "attention-scale", "hypotheses": [ {"id": "H-001", "text": "Attention head specialization increases with scale", "confidence": 0, "status": "pending", "sources": 0, "depends": [], "worker": null} ], "accumulated_findings": [], "iterations_completed": 0 }

Confidence: 0 (unknown) → 0.3 (speculative) → 0.6 (supported) → 0.9 (well-established). Status: pending, researching, supported, refuted, inconclusive, superseded.

On iteration 2+: read EVOLVE output. Add new hypotheses from workers' new_questions. Supersede refuted hypotheses. Carry all findings forward.

PREMORTEM

Bias check: confirmation bias — only searching for supporting evidence? source bias — over-relying on one source type? blind spots — obvious sub-questions not covered? Present to user. BLOCK on blind spots, WARN on bias, PASS otherwise.

PREPARE

Front-load what's already known:

BLOKS_CONTEXT=$(bloks context .) BLOKS_CARDS=$(bloks search {topic keywords}) PRIOR_VARS=$(python tools/ouros_harness.py --session {topic} --list-vars) # iteration 2+

On iteration 1: mostly empty. Workers discover. On iteration 2+: Ouros session has accumulated variables from prior iterations. Workers --load the session and build on prior state.

EXECUTE

Workers research INSIDE Ouros. Same contract pattern as /autonomous — worker does the work, orchestrator just provides hypothesis + context + where to report.

Worker prompt — structured JSON, same shape as /autonomous:

{ "role": "research", "hypothesis": {"id": "H-001", "text": "Attention head specialization increases with scale"}, "context": { "bloks_context": "{verbatim bloks output}", "bloks_cards": [], "prior_findings": "{accumulated_findings from research_contract.json}", "available_tools": "exa_search, nia_universal, nia_search, nia_web, llm_call, agent_call, read_file, write_file" }, "bounds": { "ouros_session": "attention-scale", "ouros_storage": "continuum/research/attention-scale", "ouros_load": false, "ouros_fork": "attention-scale-H001", "max_sources": 5, "source_types": ["arxiv", "official_docs", "reference_impl"], "artifact_path": "continuum/research/attention-scale/artifacts/H-001-iter1.md" }, "output": "continuum/research/attention-scale/reports/H-001-iter1.json" }

bounds.ouros_load: true on iteration 2+ (worker resumes session, sees prior variables). bounds.ouros_fork: set when parallel workers need independent variable space.

Worker flow (the worker decides HOW to research — orchestrator just says WHAT):

1. Write a Python research program (worker chooses search queries, synthesis approach)
2. Run it: python tools/ouros_harness.py --file /tmp/{program}.py --session {session} --storage {storage} Add --load if bounds.ouros_load is true. Add --fork {name} if bounds.ouros_fork is set.
3. Inside Ouros: exa_search, nia_search, llm_call etc. — results in REPL heap, not in context
4. Program writes compact artifact to bounds.artifact_path
5. Worker reads artifact, writes report JSON to output path