Pre-flight (proactive) mode

Reactive troubleshooting fixes the problem after a command fails. Pre-flight tries to prevent it: right before the agent runs a risky or version-sensitive command, searchoclock injects a one-shot "verify the current approach as of today before you write it" nudge. It is off by default and fires only on a narrow slice — not on every command — because not every dev wants deep research on everything.

Turn it on

SEARCHOCLOCK_PROACTIVE=1

That's it. It registers on PreToolUse(Bash) and is non-blocking by default (permissionDecision: allow + a short additionalContext brief). It never blocks unless you opt in.

Why PreToolUse

It's the only lifecycle event that fires before each individual command and can both inject context (additionalContext) and gate the action (permissionDecision). SessionStart/SubagentStart can inject but can't see the concrete command; UserPromptSubmit fires per-turn before a tool is chosen (noisier). Pre-action interception is what enables prevention rather than only post-failure repair.

Requires a CLI new enough to support additionalContext in PreToolUse (≈ v2.1.9+); on older CLIs the hook fails open (allow, no context). This machine reports a current 2.1.x.

What it fires on (selective — SEARCHOCLOCK_PROACTIVE_TRIGGERS=dep,destructive,version)

• dependency — first touch / install / upgrade of a package (npm/pnpm/yarn add|install, pip/uv add|install, cargo add, go get, brew/apt install, …). Briefed once per package per agent, then silent.
• destructive — rm -rf, drop/truncate table, force-push, *migrate, terraform apply/destroy, kubectl delete, prod deploy, … (on the normalized command, so a wrapper/obfuscation can't hide it).
• version-sensitive — fast-moving SDK/framework surfaces or version-pinning/lockfile ops.

Ordinary reads/edits/builds get nothing (zero noise). Tune the set, or add ignores with SEARCHOCLOCK_IGNORE_CMD_PATTERN.

Optional gating for destructive commands

By default destructive commands get a context nudge only (no block). To add friction:

SEARCHOCLOCK_PROACTIVE_GUARD=ask   # PreToolUse permissionDecision: ask (escalate to you)
SEARCHOCLOCK_PROACTIVE_GUARD=deny  # block with a reason

It never uses allow-to-widen (a hook can't bypass an explicit deny rule anyway), and it gates on the deterministic destructive lattice — never on a model confidence guess.

Cost

The trigger pass is pure regex on the normalized command (no model, no web call) — same latency profile as the reactive path. The injected brief just instructs the agent to verify-before-writing and, only for a flagged version-sensitive/first-touch surface, optionally dispatch the researcher. Actual research stays the agent's choice on the narrow flagged slice.

Posture

Reactive-by-default (cheap, always-on, ignorable) → proactive-by-exception (opt-in, narrow, nudge) → deep research (opt-in or severity-escalated). No dev and no project is ever forced into deep research.