Build a self-contained explanation of the repository as it exists at the researched commit. The default product is a technical map: what exists, where it lives, how it behaves, how its parts connect, and how that behavior is tested.

Recommendations, redesigns, implementation proposals, critiques, and causal diagnosis are separate assignments. Include one only when the user explicitly puts it in scope, and label it apart from the current-state findings.

Read the research protocol in full before exploring. It defines the direct-question, artifact, codebase, dependency, web, and follow-up branches; the protected-file boundary; source precedence; delegation; artifact placement; and permalink rules.

Use the user's stated question as controlling input. If the system supplies a task artifact directory, inspect it exactly as the protocol requires. Read every directly named input file in full in the main context before delegating any work.

Create a coverage ledger with one row per question and columns for research area, required evidence, test evidence, assigned investigator, and status.

This step is complete when exactly one inquiry is in scope, all authorized inputs have been read, and every question has a ledger row. If the available inputs are ambiguous, ask the protocol's single selection question and stop.

Split the ledger into cohesive areas that can produce independently checkable findings. Mark dependencies between areas so discovery work precedes analysis that needs its results. Choose the applicable investigator roles and tool branches from the protocol; the skill must remain executable when named specialists or delegation tools are absent.

This step is complete when every ledger row belongs to at least one bounded investigation, each investigation names its expected evidence and testing coverage, and dependent work is ordered.

Run independent investigations concurrently and dependent investigations in waves. Keep every investigator read-only and require paths with line ranges, relevant symbols or entry points, connections to adjacent components, test locations and methods, uncertainties, and external links when web or dependency sources are used.

Wait for every dispatched investigation. Reconcile overlaps against live repository evidence and record conflicts instead of silently choosing or averaging them.

This step is complete when each ledger row contains traceable evidence or states the exact evidence that could not be found and why that gap matters to the question.

Before drafting, read the report standard in full. Synthesize by concept and behavior, not by investigator or file. Connect boundaries and data/control flow, and document the testing pattern beneath every component area—including an explicit finding when no test evidence exists.

This step is complete when every material claim is traceable, every original question is answered or precisely unresolved, cross-component connections are explained, and every finding satisfies the report standard.

Gather repository metadata and choose the destination before drafting, following the artifact rules in the protocol. Then read the research document template and write the complete artifact with no placeholder text.

If unresolved questions remain, perform one additional targeted investigation pass after the first complete draft. Integrate answers into their proper sections and remove resolved entries; retain the remaining entries as explicit open questions. Do not add a second research appendix.

This step is complete when the artifact path and chronological prefix are valid, metadata reflects the researched snapshot, every required section is populated, and the coverage ledger has no unclassified or silently omitted row.

Read the final-answer contract and follow its applicable branch. Include an artifact link when the host provides one; otherwise provide its local path. Include durable source permalinks only when the protocol's reachability check passes.

For later questions about the same inquiry, update this artifact in place. Expand or correct the relevant findings and references rather than appending an unintegrated answer dump.

This step is complete when the user can open the artifact, see whether any question remains open, and identify the next available workflow action.