The hard rules move from the goal tree into a proper AGENTS.md at the capsite root (discoverable by agents without reading the tree; the workspace AGENTS.md's commit-trailer policy stays where it is, linked). Checklist keeps a one-line pointer, so a continuation prompt still surfaces them. Error hygiene (no header/param values in messages) groups with secrets discipline there rather than as its own rule. Generated-by: opencode (GLM)
1.2 KiB
capsite working rules
Applies to AI agents working in capsite/ (goal tree + how-to-run:
capsite-checklist.md; background:
the capsite notes; workspace-wide policy:
AGENTS.md — commit trailers live there).
Dev servers are the user's job
The assistant never starts, stops, or restarts long-running dev servers
(wrangler dev, pnpm dev/astro dev). The user runs them and owns the
terminals; the assistant observes via .tmp/capsite.script (the script
capture; use watchme). One-shot codegen (wrangler types), linters,
typechecks, and tests are fine.
Secrets discipline
Worker secrets never in source or wrangler config. .dev.vars (gitignored)
locally; wrangler secret put deployed. As of the DO split, the cookie
lives only in the webapp DO's storage; DIIGO_WEB_ID is the one
diigo-dump secret. Error messages must not contain header or param values
(they cross RPC as exceptions and land in logs); URLs appear query-stripped.
Keep config types generated
Run wrangler types after any wrangler.jsonc change
(worker-configuration.d.ts is generated, gitignored; hand-rolled
cloudflare:workers shims drift and gave red squiggles).