This worked example applies the field-guide workflow to public Plannotator OSS history.
The example contains two active, evidence-backed lessons and one lesson that graduated into a helper and regression test. Current Plannotator code, tests, contracts, and accepted decisions remain authoritative.
| Entry | Public evidence | Status |
|---|---|---|
| Keep agent feedback mode-specific | Issue #1107, PR #1109, commit 3f3b3514 |
Active |
| Bind long-running results to launch context | Issue #1112, PR #1115, maintainer-review response, commit 99d11dca |
Active |
| Reject absolute results from relative containment | PR #1117, commit 3455d828 |
Superseded by the existing helper and test |
The source manifest also records three candidates that were not selected for v0:
- bounded memory for large untracked files in PR #1118;
- per-file version history in folder sessions in PR #1105;
- viewed-file collapse behavior in guided review in PR #1068.
They remain possible future examples. Version 0.1 does not need six entries.
The evaluations directory contains:
- a constructed adequate and inadequate plan for the real issue #1107;
- the actual before and after containment predicates from PR #1117;
- four selective-retrieval cases and their fresh-context transcripts;
- an independently completed timed distillation case based on issue #940 and PR #1118;
- independent plan-review and code-review rubric results.
The plan fixtures are explicitly constructed. They do not claim to reproduce a historical plan. The code fixtures use small Apache-2.0 excerpts from the public Plannotator diff.
The entries use only public issues, pull requests, commits, tests, and one public maintainer comment. They do not use private Plannotator annotations, agent transcripts, credentials, or customer information.
The example passed:
- public source audit;
- owner publication approval for the three entries;
- four grounded rubric fixtures;
- four fresh-context retrieval tests;
- independent timed distillation in 4 minutes 5 seconds;
- independent assessment against both rubrics;
- deterministic link, source, privacy, license, and structure checks.