cc-steer 0.10.0

Your best training data is rotting in ~/.claude.

Your best training data is rotting in ~/.claude. cc-steer mines every correction, interrupt, and rejected plan from your transcripts into judge-refined, TRL-ready SFT/DPO/KTO pairs on HuggingFace.

Get started

uvx cc-steer scan

One pass over ~/.claude/projects fills ~/.cc-steer/feedback.db with every correction you typed, every plan you rejected, and every inline review comment — conversational context included. stats shows what landed:

Driving with an agent? Paste this:

Run `uvx cc-steer scan` to mine my Claude Code transcripts into ~/.cc-steer/feedback.db.
Then run `uvx cc-steer stats` and report how much steering was collected per source kind.
Docs: https://yasyf.github.io/cc-steer/

---

Use cases

Build a training set from feedback you already gave

You've spent months telling Claude "no, not like that", and that signal evaporates as transcripts rotate out. Judge the corpus, then distill the accepted events into atomic pairs, the corrective ones grounded in the code they fault:

uvx cc-steer triage
uvx cc-steer refine
uvx cc-steer enrich

uvx cc-steer pairs prints the deliverable: training pairs distilled from your own steering, each carrying the conversational window and code evidence behind it.

See how you actually steer, across every project

Your taste is mostly tacit — you notice a rule when it's violated. The corpus makes it legible:

uvx cc-steer list --source plan_review

On my machine the split is 698 mid-session corrections, 219 rejected plans, 41 review comments, and 22 interrupts. Another machine's history folds in too: mirror it with rsync and point scan --transcripts at it (repeatable, so several mirrors fold into one scan).

Push a private SFT/DPO/KTO dataset to your HuggingFace namespace

A judged corpus in SQLite trains nothing. Export projects it into TRL-ready configs:

uvx cc-steer export --push

traces: train 1156  test 115
sft: train 499  test 67
dpo: train 363  test 44
kto: train 1156  test 115

Four configs land as per-split parquet in a private <hf-user>/cc-steer-traces, next to a generated dataset card. Splits group on the session hash, so a session never straddles train and test.

More in the docs

• Incremental scanning — content digests and one-transaction commits make re-scans cheap and interrupt-safe — scan
• Judge, audit, eval — prompt-versioned triage, a seeded audit sample, and mechanical metrics with no LLM calls — triage
• Pair dashboard — browse refined pairs and their full lineage in a local web UI — view-samples
• Python API — drive the scanner and the feedback store from your own code — reference

Status: alpha — the pipeline runs end to end; the judge prompt still iterates (v6 today).

Read the docs for the full guide. Licensed under PolyForm Noncommercial 1.0.0.