bootstrap-hooks

Bootstrap awareness hooks

A skill only helps if the agent knows it's there. better-dev leaves a discovery block in the project's CLAUDE.md (or AGENTS.md) for exactly that, but a fresh session - or a subagent spun up mid-task - may start before it reads that file. These hooks close the gap: on session start, and again when a worker is dispatched, they emit a short note that better-dev's practices are installed here and where the project's overrides live. One job: keep the agent aware that better-dev is present, without ever blocking it.

They add awareness only. They never gate a session, never fail loudly, and stay silent in any repo that doesn't have better-dev installed.

What ships

Two extensionless bash scripts and a manifest, all under hooks/:

• bd-session-start - fires on session startup, resume, clear, and compact. If the working directory sits inside a repo with a .better-dev/ scaffold, it emits a one-paragraph pointer to the better-dev discovery block and, when the project has one, to .better-dev/overrides.md so the agent reads project overrides before applying any built-in default. It also appends a one-line update nudge when the installed better-dev clone (its plugin root) is behind its origin - a best-effort git fetch bounded to 2 seconds, well inside the hook's own 5-second host timeout, so a slow remote can cost the nudge but never starve the awareness note the hook exists to deliver. It degrades to a silent no-op when the clone isn't a git checkout, is offline, or can't be time-boxed. The nudge only suggests git -C <clone> pull; it never pulls on its own, honoring never-blocking. A project that wants hands-off updates can opt in with a better-dev auto-update: ff-only line in its overrides, which lets the clone fast-forward itself - and only fast-forward, so it can never clobber local work.
• bd-subagent-start - fires when the host dispatches a subagent. SessionStart context reaches the parent thread only, so a fresh worker would otherwise run better-dev-unaware. This re-injects a short note - the practices and overrides still apply, see /orchestrating-agents for how dispatch works - and deliberately tells the worker not to re-run onboarding. A dispatched worker should do its briefed task, not re-orient.
• hooks.json - the Claude Code manifest registering both, each with a commandWindows fallback and a 5-second timeout.

Always on when installed, silent otherwise

Both scripts walk up from the working directory looking for .better-dev/. Find it, emit the note; don't, exit 0 with no output. That single check is the whole install detection - there is no flag file, no mode, no per-session state to manage. Installing better-dev turns the hooks on; a repo without it never hears from them.

The scripts never read stdin (a hook that blocks on stdin can freeze a session), never write outside their own stdout, and swallow their own failures - a hook error must never surface as a session error. A top-level trap 'exit 0' ERR on the production path enforces that swallow, since set -euo pipefail alone would exit non-zero on an unexpected error before the trailing exit 0 runs. The one non-obvious mechanic worth knowing if you touch these: a subagent hook's raw stdout is dropped by the host, so bd-subagent-start wraps its note in hookSpecificOutput.{hookEventName,additionalContext} where the session hook can emit the platform's plain field. The selftest subcommand on each script checks both the install-aware and no-op paths and that the emitted JSON parses.

Installing and adapting to other hosts

On Claude Code, point the plugin at hooks/hooks.json (the packaging step wires this) and the two events register themselves. For the mechanics of why the payloads are shaped the way they are, and how to port them, read porting.md in this folder.

The SessionStart script already branches its output field for Cursor, Claude Code, and Copilot, so session-level awareness is cross-host today. SubagentStart is a Claude-Code (and Codex) event; hosts that expose no subagent-spawn hook get session awareness but not per-worker re-injection until an equivalent hook is sourced for them. That's a coverage limit to name, not a failure - the session note still lands.

How this fits

onboard writes the discovery block and the .better-dev/ scaffold these hooks look for; without that scaffold the hooks stay silent, which is the correct behavior before a repo is onboarded. The subagent note points dispatched workers at /orchestrating-agents, which owns how better-dev fans work out to fresh-context subagents. The overrides these hooks surface are the ones the /overrides layer persists.

Done when

A session started in an onboarded repo receives the awareness note; a session in a plain repo receives nothing and sees no error; a dispatched worker gets the short subagent note in the shape the host actually reads; and both scripts pass selftest.