From xlfg-engineering
Internal xlfg phase skill. Use only during /xlfg-debug runs to reproduce the failure, separate symptom from mechanism, and write an evidence-backed diagnosis without changing source code.
npx claudepluginhub flrngel/xlfg --plugin xlfg-engineeringThis skill is limited to using the following tools:
Use only during `/xlfg-debug` orchestration.
Mandates invoking relevant skills via tools before any response in coding sessions. Covers access, priorities, and adaptations for Claude Code, Copilot CLI, Gemini CLI.
Share bugs, ideas, or general feedback.
Use only during /xlfg-debug orchestration.
Input: $ARGUMENTS (RUN_ID or latest)
Produce an evidence-backed root-cause report and likely repair surface without modifying source code.
RUN_ID, DOCS_RUN_DIR, and DX_RUN_DIR.context.mdmemory-recall.mdspec.mdtest-contract.mdtest-readiness.mdworkboard.mdrepo-map.md, why.md, diagnosis.md, research.md, harness-profile.md, debug-report.mddocs/xlfg/knowledge/current-state.mdspec.md as canonical. For prompt or agent debugging, treat the prompt text, tool permissions or contracts, context windows, evaluation criteria, and false-success traps as first-class diagnosis surfaces. Do not collapse everything into one suspicious line too early.xlfg-repo-mapper unless a fresh same-run repo-map.md already exists and clearly covers the failing surfacexlfg-why-analystxlfg-root-cause-analystxlfg-harness-profiler when any runnable harness, failing command, or log stream existsxlfg-env-doctor when environment, local server health, race timing, infra, or dependency drift may be causalxlfg-test-strategist when the reproducer or disproof probes are still vague after first-pass diagnosisxlfg-researcher only when freshness or external docs materially change the diagnosisStatus:, and contains real findings instead of preparation notes. If it does not, use SendMessage with the returned agent ID to resume the same specialist once before treating the lane as failed. If no agent ID is available, re-dispatch the exact same packet once.git bisect, trace replay, targeted instrumentation, or log slicing when they materially shorten origin searchtest-contract.md so it becomes a diagnosis proof contract:
fast_check or manual stepsanti_monkey_probe that would still fail under a shallow patchGUESS when the repo evidence is not strong enough to claim certaintyspec.md as the single source of truth:
Outcome / whyResearch and contextExecution shape for a diagnosis-only runSolution summary, where chosen solution becomes the likely repair surface only — no code changesTask map with diagnosis tasks onlyProof summary with reproduction, disproof, confidence, and remaining unknownsdiagnosis.md with the causal chain and rejected shortcuts.debug-report.md as the final diagnosis artifact.workboard.md so stage status, blockers, and the next safest repair or proof step are visible. Add a debug stage line if the ledger does not already have one, and mark implementation-oriented stages as SKIPPED (/xlfg-debug) when that improves clarity.repro-notes.md, probe-log.md, history-findings.md, env-plan.md, research.md. Do not create implementation tasks.debug-report.md must begin with Status: DONE or Status: BLOCKED or Status: FAILED and contain:
# Debug report
## Problem summary
- ...
## Expected vs observed
- ...
## Smallest honest reproduction
- ...
## Causal chain
1. ...
2. ...
3. ...
## Deep root problem
- ...
## Strongest evidence
- ...
## Fake fixes to reject
- ...
## Likely repair surface (no edits made)
- ...
## Unknowns / confidence
- ...
## Next safest proof step
- ...
PRIMARY_ARTIFACT with Status: IN_PROGRESS, the scoped mission, and a short checklist so the specialist is resuming a concrete work item instead of starting from an empty chat turn.PRIMARY_ARTIFACT: <exact path>
FILE_SCOPE: <bounded files or paths>
DONE_CHECK: <single honest check or NONE>
RETURN_CONTRACT: DONE|BLOCKED|FAILED <artifact-path> only
LS or Glob instead of Read on directories; use Grep plus chunked Read windows instead of loading an oversized file in one shot.