Where are the leads?
Run the frozen cue expressions over code, tests, and support files. Excludes caches, build output, and vendored trees.
pohe-agent scan <repo>A portable agent package for the 52 POHE records. Scan an arbitrary repository for harness-pattern cues, pull the structural candidates into source review, and turn what survives into a pattern-guided evolution proposal.
What is worth reading? Is the mechanism in this scope? What may safely change?
Run the frozen cue expressions over code, tests, and support files. Excludes caches, build output, and vendored trees.
pohe-agent scan <repo>Anchor each cue to its enclosing declaration, then check context, trigger, structure, and effect. The packet reports missing roles, the negative control, and contract probes. Parallel file analysis is deterministic with --jobs.
pohe-agent analyze <repo> --pattern P01Build a proposal that protects the contract, names authority, adds the test obligation, and records rollback.
pohe-agent plan <repo> --findings findings.jsonThe scanner is Python standard library only. The agent instructions and skills travel with the repository.
Python 3.10+. No API key, no network access, and no repository upload.
git clone https://github.com/xhyccc/pohe-agent.git
cd pohe-agent
python3 -m pip install -e .
pohe-agent scan /path/to/your/harnessEvery layer is readable, testable, and replaceable.
The operating manual: the evidence ladder, the standard workflow, the hard rules, and the handoff points between detection, matching, and evolution.
Detection finds leads. Matching adjudicates a candidate against the pointcut, contract, and negative control. Evolution plans and verifies the change.
pohe-pattern-detectionpohe-pattern-matchingpohe-harness-evolutionscan, match, analyze, plan, list, and show. Deterministic walking, JSON output, and Markdown reports.
All 52 records: pointcut, contract, variation, test obligation, negative control, evolution question, cue, and canonical exemplar where one exists.
A three-file mini harness with a bounded loop, deferred tool discovery, and a permission check. Scan it before pointing the tool at your own code.
Catalog integrity, repository walking, cue retrieval, structural candidates, report boundaries, and plan generation. Runs on every push.
Not only the bundled example: 52 records against 41 frozen repositories, compared cell by cell with the published cue matrix.
All five misses are files the scanner excludes by design: four above the 2 MB limit and one minified third-party asset. Most extra cells are cells where re-running the study's own detector retrieves cues the frozen matrix omits. This is cue-level agreement, not a claim that every match is an implemented pattern. Method and per-cell differences
The current calibration audit retrieves all 51 canonical anchors and rejects none, while keeping 33 missing-structure cases and one medium negative-control risk visible. It does not force exemplars to pass. Read the calibration report
The tool can produce E1 on its own. E2 and E3 require reading the source, the tests, or a runtime check. The package keeps that boundary visible in every report.
Read the full evidence keyscan and match.Read the target repository's own instructions and confirm the task: detection only, matching, or evolution.
Retrieve cues, read the family table, and keep unobserved records visible as unobserved.
Run analyze, read the anchored scope and missing roles, then open the source. Answer all four pointcut dimensions and all four contract fields. Apply the negative control.
Write the evidence level actually reached, the files and tests, and the remaining unknowns.
Protect the contract, stay inside the variation seam, and name authority, rollback, and a retirement trigger.
Add the contract test first. Exercise failure and termination paths. Canary one real task before rollout.