eliasforge/fable-agent-orchestration/behavior-contract-harness

Behavior Contract Harness

Use this when an agent workflow should be enforced and measured, not merely described.

The goal is not to make a model sound like another model. The goal is to turn desired behavior into a contract with enforcement points, eval probes, real-work evidence, and regression gates.

Failure Modes

Avoid both:

• Style cosplay: the agent sounds closer to the target behavior but still

stops too early, buries failures, or claims unverified success.

• Hook theater: the workflow blocks obvious bad endings but has no evidence

that it improves real task completion.

The useful shape is:

behavior rule -> enforcement surface -> eval probe -> negative control
-> real-work telemetry -> regression decision

Procedure

1. Name the behavior contract.

Define the behavior in operational terms:

• finish reversible in-scope work;
• lead with the outcome;
• report failures verbatim;
• distinguish questions from implementation requests;
• use parallelism only when it creates throughput;
• verify before claiming done.

1. Pick enforcement surfaces.

Choose the smallest surfaces that can catch drift:

• prompt or instruction text for posture;
• stop hook for promised-but-undone final messages;
• prompt hook for question-shaped requests;
• post-tool hook for failed command honesty;
• pre-compact hook for continuity preservation;
• critic agent for nontrivial completion claims.

1. Classify authority.

Every enforcement surface must declare:

• read-only nudge;
• blocking gate;
• LLM judge;
• telemetry-only observation.

Prefer fail-open for tooling errors. Prefer fail-closed only when a false pass would ship an unsafe or misleading claim.

1. Build eval probes.

Each probe should have:

• task prompt;
• expected behavior checklist;
• candidate output;
• baseline output;
• judge rubric;
• negative control where possible.

Do not use golden transcript similarity alone. Add real outcome checks: task completed, test fixed, diff correct, failure reported, no false done.

1. Measure noise before tuning.

Run repeated baseline and candidate trials. Treat small deltas as noise until they exceed measured run-to-run spread and at least one-probe-flip magnitude.

1. Run real-work telemetry.

Synthetic probes are not enough. Track local-only counts for:

• stop-too-early caught;
• failed-command honesty nudges;
• tool discipline catches;
• critic refutations;
• false positives;
• time and token overhead.

1. Decide by regression gate.

Accept a contract change only when it improves a named behavior without increasing false positives or blocking correct work beyond the declared budget.

Required Negative Controls

Include probes where the correct behavior is to stop or ask:

• destructive action without clear user request;
• credential, payment, publish, or irreversible external gate;
• question-shaped prompt that asks for assessment, not edits;
• implementation request with no reversible next step available;
• failed tests where the correct final answer must report failure.

Review Checklist

Before adopting a behavior plugin or contract:

• license is compatible;
• telemetry is local-only or explicitly disabled;
• hooks are readable and bounded;
• shell scripts have LF line-ending protection;
• Windows, WSL, macOS, and Linux limits are documented;
• evals isolate unrelated plugins and settings;
• goldens or reference outputs are not treated as absolute truth;
• real-work smoke tests exist;
• false-positive budget is stated;
• no private project data, raw session logs, secrets, or closed methodology is

included.

Output

Return:

• behavior contract being enforced;
• enforcement surfaces and authority class;
• eval probes and negative controls;
• real-work telemetry or smoke test evidence;
• accepted changes, rejected changes, and residual risks.

Attribution

This skill is an adapted, public-clean synthesis inspired by `rennf93/opus-fable-playbook`, MIT licensed, inspected at commit 160acb6a59f83d80e45ef8e1148bac28bd46c143.

No code, hooks, scripts, golden transcripts, or raw project content from that repository are vendored here. The useful public workflow primitives were rewritten as a generic behavior-contract harness for this skill database.