xlfg-engineering

xlfg-engineering

xlfg is an autonomous, proof-first SDLC harness for Claude Code.

Version 2.8.2 fixes the phase-gate Stop hook so it exits cleanly on empty stdin (stops flaking inside an active /xlfg run, stops blocking unrelated invocations that share the cwd) and adds a pointed diagnostic when an xlfg verify contract accidentally uses pytest-style -k "not ..." with a unittest runner. Version 2.8.1 registered /xlfg-debug as a short alias and added the xlfg-ui-designer specialist (conditional plan/verify lanes for UI-related work). Version 2.8.0 hardens the conductor itself: a Stop hook and phase-state file prevent the pipeline from ending before all 8 phases complete, and loopback iterations are now capped to prevent unbounded context growth.

• /xlfg and /xlfg-debug are the public entrypoints, each batching hidden phase skills
• the plugin commands keep short aliases (/xlfg, /xlfg-debug) via name: frontmatter, while the namespaced forms remain /xlfg-engineering:xlfg and /xlfg-engineering:xlfg-debug
• the batch now includes a mandatory intent phase before context gathering and planning
• spec.md is now the only active home for the intent contract and objective groups
• bundled / messy requests are split into stable objective groups (O1, O2, ...)
• the workflow ships an artifact-graded xlfg eval-intent harness for scoring ask recall, objective splitting, blocker handling, and false assumptions
• hidden phase skills still load just in time, matching Claude Code’s skills model while keeping context small
• every plugin specialist now has an explicit tool allowlist, proactive delegation description, foreground-only bias, stronger execution contract, and bounded turn budget
• review specialists now write lane artifacts under reviews/, and the standalone pack now includes .claude/agents/ parity

What is in this repo

1. A Claude Code plugin in plugins/xlfg-engineering/
2. A standalone .claude/skills/ pack in standalone/
3. A dependency-free Python helper CLI (xlfg) that can scaffold, recall, verify, audit, and grade intent artifacts locally
4. Research notes on recent subagent / harness hardening in docs/subagent-hardening-2026.md
5. Benchmarking guidance in docs/benchmarking.md
6. A repo handoff file in NEXT_AGENT_CONTEXT.md

Quick start

Install via the plugin marketplace (recommended)

Inside Claude Code, add this repo as a marketplace and install the plugin:

/plugin marketplace add flrngel/xlfg
/plugin install xlfg-engineering@xlfg

Claude Code fetches the marketplace manifest from .claude-plugin/marketplace.json, resolves the plugin at ./plugins/xlfg-engineering, and caches it under ~/.claude/plugins/. Commands, skills, hooks, specialist agents, and the context7 MCP server all activate together. After install:

• /xlfg "what you want built" — full SDLC run
• /xlfg-debug "what is broken" — diagnosis-only run (no source edits)

Both short forms are aliases of /xlfg-engineering:xlfg and /xlfg-engineering:xlfg-debug, registered via name: frontmatter on the plugin commands.

Update with /plugin marketplace update xlfg.

Manual standalone install

For environments where the plugin loader is unavailable, copy the full standalone/.claude/ directory into your target repo’s .claude/, then run /xlfg or /xlfg-debug.

Entry model

• /xlfg owns the whole SDLC run and loads hidden phase skills just in time: recall, intent, context, plan, implement, verify, review, compound.
• /xlfg-debug is the diagnosis-only sibling: recall → intent → context → debug. It finds the deep root cause and names the likely repair surface without touching source, tests, fixtures, migrations, or configs.
• Neither command asks the user to run phase subcommands or internal skills.
• spec.md is the run card and single source of truth.
• Optional docs exist only when they change a decision or proof.
• The helper CLI is optional, but when installed it makes scaffold, recall, verification, audit, and intent grading more deterministic.

Local helper CLI

python -m pip install -e .
xlfg init
xlfg start "fix login flow"
xlfg audit
xlfg eval-intent --fixture evals/intent/messy-bugfix-bundle.json --run <RUN_ID>
xlfg eval-intent --suite-dir evals/intent
xlfg verify --mode full

License

MIT

2.7.5 note

• xlfg specialists are now documented and audited as leaf workers: no nested subagent delegation inside specialist lanes.
• Plugin and standalone agent packs now share the same bounded maxTurns budgets again, so stalled lanes fail faster instead of looking hung.
• Review fan-out is leaner by default, and conductor guidance now says waiting is valid only when a preseeded PRIMARY_ARTIFACT and explicit RETURN_CONTRACT exist.

2.7.1 note