Tycho: never accept the model’s summary — only artifacts it can’t narrate
Jul 30 Swail Tycho: completion-hook ledger that checks files, commands, exit codes, and claims against git/FS — no LLM in the trust path; VERIFIED/FAILED/STALE for overnight review.
Source: Tycho — Verify what your AI coding agent actually did
Swail Labs’s Tycho (CLI/tycho-cli; Claude Code and Codex hooks; Jul 30 public trail) is the boring half of overnight coding: a completion hook that records what the turn did and checks it against artifacts the model cannot fake by narrating. Files touched, commands run and their exit codes, claims in the agent’s own output — then git, the filesystem, and the harness event stream as ground truth. The motto is literal: never accept the model’s summary as evidence. No API key, no second model, nothing leaves the machine; stdlib-only Python (or a standalone binary). On Claude Code the human channel carries the verdict for free; on Codex the same offline checker spends a little context because that harness has one shared field.
The receipt that matters for morning review is the acceptance ladder: attempted → executed → artifact_changed → claim_supported. Gaps render as empty rungs, not invented failures. Verdicts are VERIFIED, FAILED, STALE (evidence older than the code), or backfilled UNVERIFIED. tycho blame ties a file to the turns that touched it; tycho review flags hunks no test or command exercised; exit codes are one-per-adverse-kind so CI can choose what blocks (FAILED, STALE, unexercised, trailer mismatch). Honest limits stay in the open: unkeyed SHA-256 is tamper-evidence against accidents, not authenticity against a motivated agent; some provenance families are advisory; a backfilled transcript never gets a live verdict.
Why it matters for code review
Always-on harnessed momentum compounds only when overnight “done” is checked against process evidence, not session prose. Peak-session Claude mastery still wins on a single change you can watch; continuous factories lose the night when green summaries outrun anything a hook recorded. FOMO that agents aren’t working overnight is cheaper than waking to PRs whose only proof is the builder’s diary.