cc-steer

Collect developer steering signals from existing Claude Code transcripts.

cc-steer [OPTIONS] COMMAND [ARGS]...

Full --help output

Usage: cc-steer [OPTIONS] COMMAND [ARGS]...

  Collect developer steering signals from existing Claude Code transcripts.

Options:
  --version  Show the version and exit.
  --help     Show this message and exit.

Commands:
  audit               Audit a seeded stratified sample of the current...
  compare-arms        Verdict two sweep score reports under the...
  decisions           Mine and summarize AskUserQuestion decision rounds...
  enrich              Ground every refined pair in the code evidence...
  eval                Compute the mechanical metrics for the current...
  evidence            Search the refined-pair evidence index by BM25,...
  exemplars           Retrieve the exemplars the frontier refiner would...
  export              Export the steering lineage as a HuggingFace dataset.
  freeze-eval         Freeze the gate, watcher, steer-type, and pick eval...
  golden              Author, verify, and audit the blind watcher-fires...
  hooks               Manage the global SessionEnd hook that feeds...
  hosted              Calibrate and promote a watcher adapter for a...
  import              Import cc-factory decision outcomes as judged-pair...
  inbox               List recent and undelivered steers — the fallback...
  index               Embed the accepted steering exemplars for the...
  list                List recent feedback events, newest first.
  live                Control live steering: the delivery mode, the kill...
  models              Inspect and flip the registry of lab-trained model...
  pairs               Print the refined training pairs — the pipeline's...
  pipeline            Run the collection stages as one budgeted,...
  predict             Predict a steering decision from recorded evidence.
  profile             Print the per-category steering profile built from...
  prune-gate-samples  Delete gate samples whose rendered text has no...
  rebuild-context     Rebuild stored contexts from their source session...
  refine              Refine every accepted steering event into atomic...
  retrain             Retrain and (conditionally) promote one component...
  sample-negatives    Sample gate training windows: rewound positives,...
  scan                Scan transcripts for feedback, incrementally.
  score-watcher       Score one watcher recipe under the uniform...
  shadow              Analyze the live watcher's shadow-mode proposals.
  stats               Print ingestion counts by source kind and triage...
  thresholds          Re-fit served cascade thresholds from the live...
  triage              Judge every stored candidate lacking a verdict at...
  view-samples        Serve the training-pairs dashboard: refined pairs...
  watch               Tail live transcripts and run the steering cascade,...

Options

--version: Show the version and exit.

Commands

scan: Scan transcripts for feedback, incrementally.
rebuild-context: Rebuild stored contexts from their source session transcripts.
prune-gate-samples: Delete gate samples whose rendered text has no substantive content.
stats: Print ingestion counts by source kind and triage coverage.
list: List recent feedback events, newest first.
triage: Judge every stored candidate lacking a verdict at the current prompt version.
audit: Audit a seeded stratified sample of the current prompt version’s verdicts.
eval: Compute the mechanical metrics for the current prompt version.
refine: Refine every accepted steering event into atomic training pairs.
enrich: Ground every refined pair in the code evidence behind it.
export: Export the steering lineage as a HuggingFace dataset.
pairs: Print the refined training pairs — the pipeline’s deliverable.
sample-negatives: Sample gate training windows: rewound positives, hard and random negatives.
index: Embed the accepted steering exemplars for the frontier refiner’s retrieval.
exemplars: Retrieve the exemplars the frontier refiner would see for a steering context.
hooks: Manage the global SessionEnd hook that feeds continual collection.
live: Control live steering: the delivery mode, the kill switch, and the prompt-submit hook.
inbox: List recent and undelivered steers — the fallback surface when the hook stays quiet.
models: Inspect and flip the registry of lab-trained model versions.
decisions: Mine and summarize AskUserQuestion decision rounds from transcript trees.
thresholds: Re-fit served cascade thresholds from the live scored-moment distribution.
hosted: Calibrate and promote a watcher adapter for a hosted serving substrate, distinct from the local lane.
retrain: Retrain and (conditionally) promote one component through the model registry.
score-watcher: Score one watcher recipe under the uniform Tinker-frame instrument and write .athome-metric.json.
compare-arms: Verdict two sweep score reports under the instrument card’s paired-DeLong rule.
freeze-eval: Freeze the gate, watcher, steer-type, and pick eval views into ~/.cc-steer/eval/.
golden: Author, verify, and audit the blind watcher-fires golden labeling packet.
pipeline: Run the collection stages as one budgeted, schedulable pass.
watch: Tail live transcripts and run the steering cascade, delivering per ~/.cc-steer/live.toml.
shadow: Analyze the live watcher’s shadow-mode proposals.
view-samples: Serve the training-pairs dashboard: refined pairs and their full lineage.
evidence: Search the refined-pair evidence index by BM25, optionally MMR-reranked.
import: Import cc-factory decision outcomes as judged-pair candidates.
profile: Print the per-category steering profile built from the refined pairs.
predict: Predict a steering decision from recorded evidence.