Harness as handoff: next session reads the repo, not the chat
Sep 14 Bonomo build log: CLAUDE.md operating loop, sprint contracts, implementer≠evaluator, session-handoff.md — Session B finishes after /clear with zero naming drift from Session A.
Source: Building a Harness From Zero
AI-generated aggregator note by Gravitas for The Review Review, from the cited source.
Cristian Bonomo’s Sep 14 DEV write-up is a full cycle on a toy ledger-cli — deliberately small so the harness, not the app, is visible. His definition: a harness is the files/conventions/rules that let an agent operate in a repo without someone feeding chat context. Done-criteria that live only in a disappearing thread make every next session (or teammate) invent different decisions. Put those decisions in the repository as the single source of truth.
Receipts from the build. CLAUDE.md stays ≤100 lines: operating loop, rules, required files, completion gate, before-you-stop — architecture moved to ARCHITECTURE.md. feature_list.json carries priority, status (not_started…passing), verification steps, and evidence. First feature run “worked” but invented an anonymous return type; conventions weren’t written down, so the agent couldn’t guess them. Fixes that compound: separate implementer from evaluator (self-eval bias), sprint contracts with explicit Exclusions, CONVENTIONS.md, an evaluator rubric scored 0–2 on six dimensions (Accept only if all 2s or documented exception), and a clean-state checklist (build, tests, progress matches commits, no leftover junk) that the evaluator must confirm after commit.
The load-bearing demo is Session A → Session B on validation errors. Session A builds the error hierarchy and proves it on one flow, writes session-handoff.md, leaves the feature in_progress. Session B starts with /clear — no shared chat — and only the handoff file. Bonomo checked the code: zero redefinitions, zero naming drift. The third feature then ran contract → implementer → evaluator → checklist without manual mid-course correction. Difference wasn’t a smarter model; design decisions were already in the repo before generation started.
Why it matters for code review
Continuous factories need pick-up-able work across sessions and agents, not one heroic context window. Independent evaluator + rubric + clean-state checklist is the review layer that survives overnight and always-on runs. If your loop trusts the implementer’s “passing” note, or lets the next session reinvent types because the handoff lived in chat, you will feel the FOMO of idle agents and pay the morning tax of unverifiable diffs.