René's URL Explorer Experiment

Title: feat(skilldoc): generated+validated flashduty skill — lean SKILL.md router + 19 domain cards by ysyneu · Pull Request #56 · flashcatcloud/flashduty-cli · GitHub

Open Graph Title: feat(skilldoc): generated+validated flashduty skill — lean SKILL.md router + 19 domain cards by ysyneu · Pull Request #56 · flashcatcloud/flashduty-cli

X Title: feat(skilldoc): generated+validated flashduty skill — lean SKILL.md router + 19 domain cards by ysyneu · Pull Request #56 · flashcatcloud/flashduty-cli

Description: What Redesign the built-in flashduty skill so SKILL.md is a lean index/router backed by 19 per-domain reference cards, plus the internal/skilldoc tooling that keeps each card's factual content generated from the compiled cobra tree and validated against it in CI. Bar (met): for a core domain like status pages, the agent reaches full competence after exactly 2 loads (SKILL.md + one reference/.md) — no --help trial-and-error, no guessed command names. Why The old hand-written skill drifted from the binary (e.g. it taught statuspage create-incident when the real command is status-page change-create), so agents fell back to a --help dance. Root cause: the card was decoupled from the CLI surface. Fix: generated facts + curated judgment in one card, validator-first. Architecture Oracle = the compiled cobra tree, not the OpenAPI spec. A new hidden fduty __dump-commands walks the live tree → JSON (internal/cli/dump_commands.go). cligen transforms spec→binary (positional folding, flag names), so only the binary is authoritative. Each card has a fence (verbs, flags, enums, positionals) rendered from the dump; everything outside the fence (routing, hot-flow, gotchas) is hand-curated. skilldoc tooling: Dump → Harvest (fduty … examples from the cards) → Validate (unknown command / unknown flag / folded-positional-used-as-flag) → Generate/CheckFences (freshness). Dev CLI skilldoc gen [group] | check; make gen-cards / check-cards; CI gate runs skilldoc check on ubuntu/windows/macos and fails on drift. cligen positional folding is captured: a single required *_id folds into a positional (change-create ), the same-named flag stays registered but fails the Args check — the generator renders it as a positional and the validator now rejects passing it as --page-id (the exact error class that previously only a live run caught). Files internal/cli/dump_commands.go — __dump-commands oracle. internal/skilldoc/{model,build,harvest,validate,generate}.go — dump/harvest/validate/generate. internal/cmd/skilldoc/main.go — dev CLI (gen/check). skills/flashduty/SKILL.md + reference/*.md — lean router + 19 cards. Makefile (gen-cards/check-cards), .github/workflows/ci.yml (gate). Validation (local, against a live AI-SRE session connected to a local runner) Status-page A/B (old vs new): --help 13→0, fduty failures 9→0, every task lands at exactly 2 loads, 6/6 success. All 16 operational domains rolled out + eval'd: 16/16 success; after iterating 3 cards, zero real fduty-command failures (residuals are client-side jq scripting, orthogonal to teaching fduty). Chinese product-term routing (12 real queries): 12/12 routed to the correct card, 0 --help, 0 failures — incl. the trap "协作空间" → channel (not the naive "频道"), 富化→enrichment, 分派→route. Quality gates (simplify + two adversarial bug reviews) run; all findings fixed. Full -race suite + check-cards green. Notes Source of truth lives here; fc-safari embeds these cards by copy-and-commit (separate PR into its feat/ai-sre). An in-sync check between the two repos is a follow-up. The safari command group (platform self-config: mcp-server-*, a2a-agent-*) is intentionally not carded — connector setup has dedicated inline-card/Settings handling; the SRE-operational 19 domains are the scope.

Open Graph Description: What Redesign the built-in flashduty skill so SKILL.md is a lean index/router backed by 19 per-domain reference cards, plus the internal/skilldoc tooling that keeps each card's factual content ...

X Description: What Redesign the built-in flashduty skill so SKILL.md is a lean index/router backed by 19 per-domain reference cards, plus the internal/skilldoc tooling that keeps each card's factual cont...

Opengraph URL: https://github.com/flashcatcloud/flashduty-cli/pull/56

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width