Targeted test
Best when a meaningful before/after signal can isolate the defect or requested behavior.
Repository proprioception for agents
Know where you are.
Make one move.
Read what changed.
Agentic Coding keeps an AI oriented in the repository that exists now: current state, local instructions, live hypotheses, bounded changes, available oracles, and the exact point where confidence ends.
Contest edition 1.0.0 · MIT License · Standalone source, not an independent-install claim
The operating loop
Let each move earn the next one.
Code volume is not progress. A useful episode advances through new observables, keeps the causal story intact, and widens only when the evidence demands it.
- 01MapFind the repository boundary, current state, local instructions, and behavior surface.
- 02LocalizeRead the target beside its nearest tests, callers, contracts, and error path.
- 03HypothesizeName one live causal mechanism and the observation that would falsify it.
- 04ProbeMake the smallest reversible change or experiment that can discriminate the hypothesis.
- 05InterrogateRun the narrowest available oracle that actually exercises the claimed behavior.
- 06InterpretState what the outcome supports—and what it does not.
- 07CheckpointClose, recover, or hand off with an exact re-entry condition.
A resumption surface, not a transcript
Carry only the state that keeps the next context competent.
- Objective
- The intended behavior and the evidence that would advance acceptance.
- Authority
- What is permitted, what is in scope, and what remains excluded.
- Repository state
- Root, branch or working state, local instructions, and known mutations.
- Decisive facts
- Tests, callers, interfaces, configuration, and error-path observations that constrain the explanation.
- Live hypothesis
- One causal story, paired with a falsifier rather than protected by optimism.
- Next probe
- The smallest reversible information-producing move and its exact re-entry condition.
Verification with a declared scope
Ask the narrowest oracle that can disagree.
A passing command is evidence only for the behavior it exercised. Widen the oracle when a change crosses a shared utility, public contract, schema, build configuration, or permission boundary.
Build
Exercises integration, generated output, bundling, and configuration surfaces that a unit test may never touch.
Lint
Checks policy and static conventions; useful, but never a substitute for runtime behavior.
Typecheck
Discriminates contract compatibility and data-shape assumptions without claiming the program worked.
Target readback
Confirms the actual state written by a tool or patch, not merely the tool's success message.
Failure is evidence
Recover by changing the premise.
Repeating the same move after the same failure is not persistence; it is expensive superstition. Preserve the signature and last-known-good state, classify the episode, then alter localization, hypothesis, probe, or oracle.
Environmental faultThe code path may be sound; the runtime, dependency, service, or local setup is not.
Pre-existing failureThe baseline was already broken. Keep that missing before/after evidence visible.
False hypothesisThe observation contradicts the causal story. Retire or revise it.
Patch regressionThe bounded change worsened or displaced behavior. Return to the last-known-good state.
Insufficient oracleThe check passed or failed without exercising the decisive surface.
Unknown mutation stateThe repository may not be in the state the agent thinks it is. Re-anchor before another edit.
Permission and acceptance
Keep three truths separate.
What may be changed
User and host authority define the action boundary. A convenient tool does not expand it.
What was attempted
Record the actual command, patch, returned state, and failure signature without upgrading a tool message into verification.
What an oracle established
Independent acceptance stays distinct from authored implementation and carries the scope of the oracle used.
A useful invocation
Start with the evidence contract.
Use $agentic-coding to diagnose this failing repository behavior. First map the live state and local instructions, then identify one causal hypothesis with a falsifier. Make only the smallest authorized probe or change, run the narrowest discriminating oracle, and report what the result supports, the unresolved risk, and the exact re-entry condition.
The skill is designed for live repository work inside a host that already provides the relevant files and tools. This standalone repository preserves the public contest source; independent plugin installation is not claimed by the contest evidence.