investigate

Investigate

The core research loop. Run N investigation cycles, each testing a hypothesis and applying keep-or-discard discipline.

Usage

Invoke this skill by name with an optional cycle count. On slash-command hosts, prefix with / (e.g. /investigate 10).

# Run 10 investigation cycles
investigate 10

# Default: 5 cycles
investigate

Default: 5 cycles if no argument given.

Inputs

Always read before starting:

• reaper-workspace/notes/problem-statement.md — model assumptions and property definitions
• reaper-workspace/notes/ideas.md — the ideas to investigate
• reaper-workspace/notes/current-understanding.md — the "branch tip" of accumulated knowledge
• reaper-workspace/notes/results.md — what's been tried and what happened

Lazy-load only when needed (do not read upfront — load only if a cycle's hypothesis requires it or the "when stuck" protocol calls for it):

• reaper-workspace/notes/paper-summary.md — the source paper (if provided)
• reaper-workspace/notes/literature.md — known prior work (organized by same-goal and same-approach)
• reaper-workspace/papers/ — downloaded PDFs and per-paper notes (<id>-notes.md) from the literature review
• ../reaper/references/methodology.md — proof/analysis patterns

The Batch Loop

Investigation runs in batches, not sequential cycles. Independent hypotheses are investigated in parallel by default; sequential execution is the fallback when dependencies exist.

Overview: Plan → Dispatch → Merge → Repeat

while cycles_remaining > 0 and not converged:
    Plan:     read ideas.md + notes/results.md → dependency graph → batch of independent hypotheses
    Dispatch: spawn one subagent per hypothesis (parallel, in one message)
    Merge:    collect results → update notes/results.md → integrate "keep" insights into current-understanding.md

Plan Batch

1. Read ideas.md and notes/results.md
2. Identify all unresolved hypotheses. Check notes/results.md to avoid repeating failed approaches.
3. Build a dependency graph: does resolving H2 require knowing the outcome of H1? If so, H2 depends on H1.
4. Select the largest set of independent hypotheses that can run concurrently. Cap the batch size at cycles_remaining.
5. Allocate cycle numbers: pre-assign a contiguous range per subagent (e.g., batch 1 gets 001-003, batch 2 gets 004-006). Each subagent gets one or more consecutive numbers from its range.
6. If all remaining hypotheses form a dependency chain, the batch size is 1 — this is the sequential fallback (see below).
7. If all hypotheses are resolved, stop and return control to the orchestrator. The orchestrator will call /brainstorm to generate new ideas if needed.