René's URL Explorer Experiment

Title: perf(skill): slim incident-summary output + serialize same-host monit-agent probes by ysyneu · Pull Request #68 · flashcatcloud/flashduty-cli · GitHub

Open Graph Title: perf(skill): slim incident-summary output + serialize same-host monit-agent probes by ysyneu · Pull Request #68 · flashcatcloud/flashduty-cli

X Title: perf(skill): slim incident-summary output + serialize same-host monit-agent probes by ysyneu · Pull Request #68 · flashcatcloud/flashduty-cli

Description: Two token-efficiency fixes surfaced by /audit-ai-sre-sessions (audit run audit-2026-06-26). Both touch files kept byte-identical with their embedded copies in fc-safari (logic/runtime/bootstrap/skills/flashduty/) — mirrored in linked PR flashcatcloud/fc-safari#337 (the two must land together). P6 — slim incident-summary.sh (scripts/incident-summary.sh) What: the script dumped ~80K chars of toon per incident (81042 / 79075 in the audited session), most of it empty boilerplate, exceeding the inline output cap and forcing 3-4 sequential read calls to page ~600 lines back in (steps 5,6,8,10,12,21,22,24 of sess_TRdrnq5oA345qF66GgbYrY). Root cause: run() appended --output-format toon to every command. For these read verbs that selects the machine-readable branch, which marshals the full raw response object — every empty field on incident detail (account_locale, account_name, ai_summary: "", …) plus heavy blobs like a change's labels.steps deploy-JSON. Each verb's default (table/summary) renderer is instead a curated projection of exactly the summary-relevant fields (printIncidentFullDetail → id/severity/progress/channel/timestamps/ai_summary/root_cause/resolution/impact/description/labels/responders; change list → id/title/status/channel/time; etc.). Fix: drop --output-format toon from run() so each command uses its lean default renderer — that default is the field projection a fault summary needs. Kept all six commands, set -uo pipefail, the no-errexit intent, and the read-only "print real output" purpose. Adjusted the header note (the lean detail shows the channel by name; to scope post-mortems to the incident's channel, read channel_id via incident info … --output-format toon | grep '^channel_id:'). Tradeoff: the lean change list table omits labels (one correlation signal) along with the noisy steps blob; the incident's own labels still print under detail, and a specific change can be drilled into via the change card's documented toon path. Net: a massive token cut for a fast overview. P8 — serialize same-host monit-agent probes (reference/monit-agent.md) What: in sess_QmoruPkRjrwA2tcHvEbNvT (steps 10,11,12) os.top_processes was rejected with code=overloaded — per-target concurrency limit reached for kind="host" locator="10.101.214.50" — because multiple monit-agent probes were fired in parallel against the same host, forcing an identical retry that re-sent the growing context. Fix: added a Gotchas line steering fan-out to serialize per target (batch a host's tools into a single invoke, whose tools array already runs concurrently agent-side) and parallelize only across distinct targets. Prose outside the GENERATED fence. Verification bash -n scripts/incident-summary.sh # OK shellcheck scripts/incident-summary.sh # no findings env -u GOROOT go build ./... # BUILD OK go run ./internal/cmd/skilldoc check # skilldoc: cards OK (fence unchanged)

Open Graph Description: Two token-efficiency fixes surfaced by /audit-ai-sre-sessions (audit run audit-2026-06-26). Both touch files kept byte-identical with their embedded copies in fc-safari (logic/runtime/bootstrap/ski...

X Description: Two token-efficiency fixes surfaced by /audit-ai-sre-sessions (audit run audit-2026-06-26). Both touch files kept byte-identical with their embedded copies in fc-safari (logic/runtime/bootstrap/ski...

Opengraph URL: https://github.com/flashcatcloud/flashduty-cli/pull/68

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width