feat: emulo --coach reports how you use the model, without mining

[ohad6k]

The first substantive reply to the star outreach asked for a mode focused on improving Claude Code usage rather than extracting conventions. This is that mode, in the open source CLI.

python emulo.py --coach                   # every source it can find
python emulo.py --coach --source claude   # Claude Code only
python emulo.py --coach --json            # for other tools

No model call, no corpus written, finishes in seconds. That matters for installs: someone gets a real result before deciding whether mining is worth it.

What it reports

Asks sent three times in a row unchanged, context re-explained after the agent lost it, runs of rephrasing the same request, and the rate of turns that open by correcting the last answer. Every finding prints the dated messages behind it, windowed onto the match when the evidence sits deep in a long message, because a receipt you cannot read is not a receipt. Checks that come in under their bar print their counts, so a clean result reads as measured rather than as silence.

Measuring it changed it three times

Thresholds came from a real 2,271 session corpus, not from guessing.

• Without a word floor, the repeat check was almost entirely ok / yes / ok do it sent three times. That is approval, not a loop. With a four word floor the same corpus yields one genuine run.
• Widening from consecutive sends to sends within three turns added only filler repeated 84 turns apart, so the rule stayed strict.
• no need for the repo was counting as a correction.
• A 1.4% correction rate was being reported as a problem. Corrections now clear a rate bar, and the rate prints either way so anyone can disagree with the bar.

It also found a real mining bug

Two harness preambles were reaching the corpus as user prose: # Context from my IDE setup: (an editor stapling the open file and tab list onto the turn) and # Files mentioned by the user (Codex listing attachments). That is 2,543 messages on one real corpus, and because they repeat near-verbatim for as long as a file stays open, mining read them as deeply held rules. Both are now in INJECTED_CONTEXT_PREFIXES, which fixes you.md quality and not just this report.

Found by reading the receipts --coach printed rather than trusting its counts. The loudest finding in the first run was an IDE preamble counted as the same ask sent twelve times.

Shipping something unproven, on purpose

The reword-loop check has no confirmed true positive. Across every corpus available to test it, it fired exactly once, on a pasted pygame banner, now excluded by a line guard. It ships because the behaviour it looks for is real and cheap to check, and because a check that finds nothing prints its zero rather than staying silent. The changelog says to trust it less than the other three.

Verified

444 tests, 4 skipped, 0 failures, up from 415. The 29 new tests cover the loop detectors, the exclusions, receipt quality, the injected preambles, redaction of receipts, and the promise that --coach leaves no files behind.

Stated in the report itself: it reads only the messages you typed, which is all Emulo keeps, so it cannot see cost, tokens, tool calls, or whether the agent was right, and it does not score them.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ditto	Ready	Preview, Comment	Jul 28, 2026 10:44pm