megaprobe

Scout first, then spend. A staged model pipeline for coding agents:

1. Detect the stack from manifests. No model involved.
2. Profile the repo once with a big-context model, cached.
3. Scout each task with a cheap read-only model that returns a structured handoff: relevant files, symbols, repro command, constraints, plan.
4. Verify the handoff's claims in code and remove the false ones.
5. Decide which worker tier fits the task (rules first, pluggable later).
6. Work with the cheapest fitting worker: claude -p or codex exec, using your existing logins.
7. Check with typecheck, lint and focused tests, and escalate one tier on failure.

How it works

flowchart TD
    repo([Repo]) --> detect["Detect stack<br/><i>code, no model</i>"]
    detect --> profile[("Project profile<br/><i>big-context model, cached</i>")]
    task([Task]) --> scout["Scout<br/><i>cheap model, read-only tools</i>"]
    profile --> scout
    scout --> handoff[/"Handoff JSON<br/>files, symbols, repro, plan"/]
    handoff --> verify{"Verify claims<br/><i>code</i>"}
    verify -- "false claims removed" --> decide["Decide tier<br/><i>rules</i>"]
    decide --> worker["Worker<br/>fast / standard / strong"]
    worker --> check{"Check<br/>typecheck, lint, tests"}
    check -- pass --> done([Done])
    check -- fail --> escalate["Escalate one tier<br/>+ failure output"]
    escalate --> worker

Why this order: recent results show a verified repository handoff lets cheap models match the best single model at about ⅕ of the cost, and the handoff mattered more than choosing the model. See the paper.

Status

Design stage. docs/PAPER.md covers the design, prior work, engineering notes, the evaluation plan and the roadmap. No code yet.

Planned usage (v0.1)

megaprobe profile                        # detect stack + build/refresh the cached profile
megaprobe scout  "fix token expiry bug"  # print the verified handoff
megaprobe run    "fix token expiry bug"  # full pipeline: scout → decide → work → check → escalate
megaprobe log                            # per-stage cost, tier, and check results of past runs

Configuration will live in .megaprobe/config.json (worker tiers, decision rules, check commands), and run history in .megaprobe/runs/.