CLI

The bayesian-agent command provides trace ingestion, result summarization, and incremental repair utilities.

Install the CLI

python -m pip install -e .
bayesian-agent --help

evolve

Update a Bayesian Skill registry from one or more results JSON files.

bayesian-agent evolve \
  --results artifacts/ga_deepseek_baseline/sop_results.json \
  --results artifacts/ga_deepseek_baseline/lifelong_results.json \
  --registry temp/beliefs.json \
  --algorithm categorical_bayes \
  --context-out temp/skill_context.md

Arguments:

Argument	Required	Description
--results	yes	Path to a result JSON file. Can be repeated.
--registry	yes	Output registry JSON path.
--algorithm	no	Belief backend. Defaults to categorical_bayes; naive_bayes is accepted as a legacy alias, and beta_bernoulli enables the global success-rate baseline.
--context-out	no	Optional Markdown path for rendered Skill context.

repair-plan

List failed task ids for incremental repair.

bayesian-agent repair-plan \
  --baseline artifacts/ga_deepseek_baseline/sop_results.json \
  --out temp/failed_tasks.json

Output shape:

{
  "sop_bench": ["sop_12", "sop_13"]
}

summarize

Summarize accuracy and token usage for a result file.

bayesian-agent summarize \
  --results artifacts/bayesian_full/results.json \
  --out temp/summary.json

incremental-summary

Summarize the lift from a baseline run plus repair traces.

bayesian-agent incremental-summary \
  --baseline artifacts/ga_deepseek_baseline/sop_results.json \
  --repairs artifacts/bayesian_incremental/results.json \
  --out temp/incremental_summary.json

This command is useful for measuring the additional inference cost required to repair failed tasks.

replay-skill-artifacts

Rebuild per-task Skill evolution artifacts from an existing results.json without rerunning the model.

bayesian-agent replay-skill-artifacts \
  --results results/sop_deepseek_v4_flash/bayesian_full/results.json

By default, artifacts are written next to the result file:

<result-dir>/skill_evolution/
  index.json
  <benchmark>/<task_id>/
    skill_context_before.md
    skill_context_after.md
    posterior_context_before.md
    posterior_context_after.md
    belief_before.json
    belief_after.json
    snapshot_before.json
    snapshot_after.json

skill_context_*.md stores the exact model-facing Skill/SOP text. posterior_context_*.md stores posterior summaries for audit/debugging and is not part of the built-in benchmark prompt.

Use --out-root when the run root is different from the result file's parent directory.

Result File Assumptions

The CLI accepts benchmark-style JSON result files that can be normalized into benchmark names and run lists. Each run should contain at least:

• task_id
• success signal such as success
• token fields such as input_tokens, output_tokens, and total_tokens

Extra fields are preserved in evidence metadata.