cc-pushback

Mine your Claude Code transcripts for the moments you pushed back into a local database.

cc-pushback collects your corrections, interrupts, rejected plans, and code-review comments — with the surrounding conversational context — into a feedback database, then judges each candidate, refines the accepted ones into atomic training pairs, and grounds them in the code they complain about. Your taste is mostly tacit (you notice a rule when it's violated), and that signal sits unused in transcript files; this turns it into a training dataset.

Install

Run with uvx: uvx cc-pushback --help.

Quickstart

Scan your transcripts and accumulate the pushback into a local feedback database:

uvx cc-pushback scan

scanned 412 files, 1473 new rows

scan is incremental and idempotent: each transcript is parsed only when new or changed, every candidate is keyed by a content digest, and recording a file commits in one transaction, so an interrupted scan never leaves the database half-written. The database lives at ~/.cc-pushback/feedback.db by default (override with --db).

Commands

scan, triage, audit, refine, and enrich sync the dataset to <hf-user>/cc-pushback-traces — a private dataset in your HF namespace, created on first push (requires hf auth login) — whenever a pass changes data; --no-sync skips it. Run uvx cc-pushback COMMAND --help for the full flag list.

What gets collected

scan runs four detectors over each transcript, each tagged with a source kind: transcript messages (transcript_message, the pushback you typed mid-session), plan reviews (plan_review, rejected ExitPlanMode plans and plan-mode re-entries), interrupts and rejections (interrupt_rejection, permission denials and user interrupts), and review comments (review_comment, one row per inline code-review comment). Each row carries the conversational window around the feedback, captured at collection time because transcripts are ephemeral.

Exporting a training dataset

Once the corpus is judged, export turns it into a HuggingFace dataset:

uvx cc-pushback export

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

One canonical traces config — one row per judged event, carrying the context, verdicts, refined pairs, and code evidence — plus three TRL-ready projections (sft, dpo, kto) land as per-split parquet under ~/.cc-pushback/dataset (override with --out), next to a generated dataset card. Splits are a deterministic group split on the session hash, so a session never straddles train and test. --push uploads every config to a private dataset in your HF namespace (--repo-id, default <hf-user>/cc-pushback-traces).

You rarely run export by hand. Every mutating pass that changed data rebuilds the dataset and pushes all four configs to that repo. A failed push exits nonzero with the local writes already committed, and the next sync picks them up; export --push is the manual catch-up.

Mining from another machine

Transcripts live under ~/.claude/projects. Mirror a remote machine's history locally, then point scan at it — --transcripts is repeatable, so several mirrors fold into one scan:

rsync -az yasyf@yasyf:.claude/projects/ ~/.cc-pushback/mirrors/yasyf/
uvx cc-pushback scan --transcripts ~/.cc-pushback/mirrors/yasyf/

Docs

Read the docs for the full guide and API reference.