Harness Discipline

Badger inherits a harness-rewrite posture. The thesis (paraphrased from the upstream’s own framing): a clean rewrite of an agent harness is more valuable than a museum-quality archive of the leaked snapshot. What survives the rewrite is the technique — what gets dropped is incidental complexity.

Badger applies that thesis to Chimera in three ways.

1. Tighter max-step budget

--max-steps defaults to 25 in badger, vs 50 elsewhere.

Why: long agent trajectories are expensive both in tokens and in context drift. A 25-step ceiling forces the agent to plan more carefully up front. When the trajectory legitimately needs more steps, the operator opts in explicitly:

chimera badger -p "..." --max-steps 80

2. Rerun-on-failure as a first-class flag

Other Chimera CLIs treat the agent’s first attempt as definitive. Badger ships --rerun-on-failure: when the first attempt produces a test failure or an obvious syntax error in tool output, the harness resets and retries with a refined prompt up to --max-reruns extra attempts.

chimera badger -p "Fix the broken pytest cases under tests/api/" \
  --rerun-on-failure --max-reruns 2

The detection is intentionally conservative — see rerun-on-failure.md for the marker list. False negatives (real failures that slip past) are preferred over false positives (rerunning a healthy trajectory).

3. Parity-tracker

The parity subcommand diffs the live agent against a declared schema (tools, max-steps, slash commands, default model, rerun_on_failure flag). When the schema and the live snapshot match, parity returns rc=0; when they diverge, rc=1 with a diff report.

The schema lives at PARITY.md, PARITY.json, or PARITY.yaml. Any of those formats works — see parity.md for details.

What badger is not

A rewrite of any specific upstream codebase. Comparative language uses “badger”, “the upstream”, or “the harness-rewrite tradition”.
A drop-in replacement for ferret/otter/mink. Pick badger when the harness-rewrite posture is what you want.

When to reach for badger

You want a focused trajectory, not a sprawling one.
You expect transient failures (flaky tests, syntax-error from a speculative edit) and want automatic recovery.
You operate against a parity contract — a schema you want to enforce in CI.

When not to reach for badger

You want sandbox-first / IDE-first ergonomics: use ferret.
You want a server-first posture: use otter.
You want a Kimi-on-Ollama experience: use mink.
You want shell-mode toggles: use stoat.