Skill guide

Autoresearch

Run a goal-directed research mission: interleave web research with the persistent Python REPL, then finish on one structured verdict. Autoresearch never implements.

Where it sits

The public workflow stays four skills and four role agents:

deep-interview -> ralplan -> ultragoal
               └─ optional autoresearch when research must ground the plan

Use Autoresearch before a deep-interview to map a landscape, or after a spec to gather evidence. Ordinary lookup that will be followed by planning belongs in deep-interview / ralplan, not a research mission.

Four, not five: worker coordination is IRC. The retired team skill duplicated that surface. gjc team and gjc rlm are gone — see the RLM removal note.

Start a mission

Availability: Autoresearch ships in Gajae Code v0.15.0 and later. Earlier binaries do not include it.

# Cold intake — clarify goal, constraints, and deliverables before any research tool fires
/skill:autoresearch "which approach benchmarks best on this dataset?"
gjc autoresearch "what changed between these two releases?"

# Spec intake — zero clarification questions; the spec must declare autoresearch-mode
gjc autoresearch intake --spec .gjc/_session-<id>/specs/deep-interview-<slug>.md

# Inspect or retire the current mission
gjc autoresearch read --json
gjc autoresearch clear

gjc autoresearch clear retires the mission artifact only. It does not dispose the session Python kernel — reset that with the python tool's own clear action.

Modes

Every mission states an explicit mode at intake: web, data, or mixed. Mode is never inferred from a file sitting in the workspace. Data-context loading is gated to data/mixed; web never attaches data context.

web — current external information.
data — local files, REPL analysis, and experiments.
mixed — both, interleaved in one ledger and one verdict.

Mission loop

A mission is intake → research runs → verdict — not a linear fourth skill after ultragoal. Cold intake clarifies the goal before any search or REPL work; spec intake reads a deep-interview spec and starts with zero clarification questions. During research, web findings and REPL/data experiments mix freely: a web result can motivate an experiment, an experiment can trigger the next search, and both land in the same ledger.

The mission ends on one best-effort verdict with status, evidence[], caveats[], and the evaluator identity. The verdict is self-issued by default; an optional critic pass records a separate evaluator identity. An inconclusive verdict stays open for follow-up rather than closing as finished.

The python tool is a persistent per-session REPL. It does not require an active Autoresearch mission, keeps variables across calls, and appends every execution to a session JSONL transcript.

Boundary: Autoresearch produces findings and a verdict, never product code. Downstream implementation goes through ralplan / ultragoal with approval-gated execution.

Artifacts

Mission state lives per session under .gjc/_session-<sessionid>/autoresearch/:

mission artifact and an append-only JSONL ledger
session-scoped run records and the TUI run-table dashboard
ledger events such as mission_created, mode_set, run_logged, verdict_issued, and mission_cleared

Resume with gjc autoresearch read --json. Do not hand-edit .gjc/.

Handoff

Autoresearch can chain into /skill:deep-interview, /skill:ralplan, or /skill:ultragoal from any live phase (intake / research / verdict). Finalize-only exit is gjc autoresearch clear.