Support harness-scoped tool hooks for policy enforcement and auditing

[astrobot-houston]

Originally proposed by @jchristman in #167. Their description is reproduced verbatim below.

All community-submitted pull requests are automatically converted to issues (bugs) & discussions (feature requests, enhancements) where they can be triaged and prioritized. Once prioritized, a PR implementation is created automatically.

Allow init({ hooks }) to run before/after each model tool execution, with before throws surfaced as tool errors to the LLM; task sessions inherit hooks; session.shell synthetic bash runs the same hook path.

const harness = await ctx.init({
  hooks: {
    tool: {
      before: async (call) => {
        if (call.toolName === 'bash' && String(call.args.command).includes('rm -rf')) {
          throw new Error('Blocked: destructive command');
        }
      },
      after: async (call, outcome) => {
        /* audit */
      },
    },
  },
});

---

Original implementation from #167 by @jchristman

[jchristman]

Hello -- I've implemented this feature so that I can build some deterministic checks around agent executions to ensure that it does not execute commands that are outside what is expected. The implementation is initially quite simple, and raising exceptions on the before hook sends the result back to the agent for corrections.

[jchristman]

Are discussions here ever likely to actually get discussion? I am curious why pull requests are automatically closed, but then not discussed?

[DevTK0]

I think that this will be a very useful feature as well. Will enable workflows whereby hooks can be used for automatic steering of a background agent.