fixer-wave-orchestration
Use when several edit subagents will change one repo at the same time: fanning fixers out over an audit's verified findings list, splitting a feature or refactor by file ownership, or accepting what t
SKILL.mdskills/managed/fixer-wave-orchestration/SKILL.md
GitHubFixer wave: findings → ownership-split fixers → acceptance
Dispatch and acceptance for any wave of concurrent edit agents in one repo; repeatable per audit round. Scouts find (shallow-codebase-audit), per-change acceptance is verify-delegated-edits, Kotlin/Android damage probes are subagent-damage-repair, and the Android readiness slice plan is android-readiness-agent-wave.
Dispatch
- Split the work into agent briefs by exclusive file ownership — no file appears in two rosters. Consolidate same-file findings into one agent even when themes differ; that agent owns the whole file.
- Decide cross-agent contracts before dispatch and write them into the shared context — agents cannot negotiate mid-flight. Contract-block format:
subagent-damage-repair(Prevent it). - Each brief carries, per finding or work item: file:line, the verified mechanism (not just the symptom), and the intended fix approach. Plus hard rules: house comment style (say WHY, name the concrete failure), no commits.
- Build rule, by toolchain: .NET agents may self-check with
dotnet build <sln> --nologo -v minimal→ 0 errors and targeted suites only, under central package management (no inline Version=); Gradle/Android agents run no gradle, lint or adb at all — parallel Gradle runs fight over the lock. Either way an agent's green is not acceptance (Acceptance 3). - Test files are owned by nobody: required test edits (constructor arity, signature callers, fixture shapes) are REPORTED by the agent and applied by the parent or assigned explicitly — otherwise an agent's "all green" hides a broken sibling test project.
- Shared-file hot regions: hub-message the owning agent mid-flight ("keep clear of line N — parent applies that fix there") and apply those yourself after the agent lands.
- List not-worth-acting findings explicitly as out-of-scope so nobody wanders into them.
Acceptance (verify-delegated-edits applies in full)
git status --porcelain=v1 --untracked-files=all+git diff --stat: touched set == union of rosters; zero deletions; zero scratch files.- Read every hunk, then the RAW file at collision regions and concurrency-sensitive changes: compact diff renderers drop words and can hide a deleted argument line. A real regression was caught exactly this way — a build stayed green while an optional record parameter was silently omitted. Optional defaults convert compile-time failures into silent null data that tests on fakes cannot see.
- Gate once, after all agents settle: concurrent edits break sibling builds transiently, so an agent's "0 errors" was measured mid-flight of the others. Kotlin/Android: run the
subagent-damage-repairprobes first. Full gate: build 0 errors + every targeted suite. Flaky/environmental suites: compare failure signatures to the documented baseline and attribute new ones by stashing the agents' files and rerunning (stash-vs-worktree method:verify-delegated-edits). - When an agent changed a shared helper signature (e.g.
Note(msg)→Note(msg, problem)), grep ALL call sites — a real failure rendered as neutral/informational is the failure mode of missing one. - Commit thematically (one commit per fix cluster); straggler test-file updates get a small follow-up commit.
Shell traps (each cost real time)
- bash expands
$inside double-quotedpowershell -Command "...". Pass code via an env var and invoke with single-quoted bash:powershell -NoProfile -Command '$env:CHECK | Invoke-Expression'. - PowerShell:
'-p:x=' + $varinside a comma-list is array + string concat (commas bind tighter) — the value becomes a stray positional arg (MSB1008). Use expandable strings"-p:x=$var".