Field origin: a workspace whose history runs 117 refactor commits to 86 feat — rebuilt in place more than extended, driven by one structured audit, and ending with every recovered rule mechanically gated. A refactor-heavy phase is healthy in agent-built code; an unstructured one is churn.
The audit is a committed report with a fixed shape:
Finding IDs (F001…) are a second coordinate system alongside work-item
IDs: commit subjects carry both (refactor(x): ... (br-ejs1.14 slice 2, F014)), so git log is greppable by work unit and by why.
One epic per audit. Children cluster findings (one bead per root cause, not per symptom), ordered by dependency, not severity. The epic body carries the goal, non-goals, the target layout, and the ordered child list with their finding IDs.
Step 0, always: if the canonical gate fails at HEAD (silently
accumulated debt — see quality-gates), green it in a dedicated chore:
commit titled "Step 0 for " before any child starts. The field
incident: clippy green "for the first time in the epic" only after a 20+
error sweep that pre-existing slices had quietly walked past.
Three field-proven sequencing patterns for migrations that cross crates; each slice is independently green with receipts (per-crate test run named in the commit body):
Course corrections mid-epic get fractional slice numbers (slice 1.5) rather than a renumbered plan — and the correcting commit says plainly what the plan got wrong ("the bead author didn't account for the dep matrix; this slice walks the placement back BEFORE either adapter consumes it, so no downstream churn").
Mechanize every rule the campaign establishes, or it's decoration. Each recovered invariant ends the campaign as a gate (dep-matrix lint, complexity gate, generated-architecture build) — otherwise the next growth phase re-accumulates the same debt. The one convention in the field repo with no gate (a glossary file) is the one already rotting.
next_step(&Event) -> Step; the full decision table is then
tested synchronously — no runtime, no fake channels.Arc<dyn Port>, exactly one public method, no
framework imports.Result<_, String> and
Debug-formatted text in user-facing output — Debug is not a stable
contract; the domain owns the noun, the use case owns the sentence.starts_with becomes an exact pin).Test deletion inside a campaign always states its coverage justification in the commit ("the 735 LOC of old tests are deleted — the verb structs' per-file unit tests cover the same surface"). Hunting test theater is toolable: classify tests by tier and flag the ones that pass whether the system under test works or not.
Small warts found while reading are fixed inline if they sit at the edges and touch no orchestration — and the commit says which warts and why inline was safe. Cross-cutting discoveries are filed, not fixed ("the fat-port wart is cross-cutting; tracked as "). Same discovery-mandatory / opportunistic-fixing-forbidden line the rest of the kit draws, applied at campaign speed.
igouss/gitweb-rs
9ef2429
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.