refactor(hooks): drop run_effect — deciders are pure inline logic

[trashhalo]

Summary

run_effect (added in 0.30.0) is the wrong seam, and unused.

Wrong seam: it baked in an in-process execution model. DirectExecutor and DBOS both run the effect in the loop's own process, so a bare closure is always reachable. A dispatch-to-pool durable runtime — Temporal — can't honor it: an activity's input must be serializable and a Go/Python closure can't cross that boundary. The only "implementation" drags execution back in-process (local activity + a co-located thunk registry), which defeats Temporal's distribution model. When the sole way a durable executor can satisfy a seam is to subvert its own runtime, the seam encodes the wrong assumption.

Unused: zero production callers — every run_effect() call site was a test. The real decider hooks (cycle-id stamping, search-budget veto, finish-cycle nudge) are pure, deterministic logic — read run_context/messages, return a decision — and run inline, replay-safe, exactly like the existing on_pre_tool/on_before_finish callbacks. The one production hook that does durable I/O (per-chat token-usage recording in hey_lefty) already dispatches a named activity directly — the one durability primitive every executor implements natively.

Changes

• agent/hooks.py: remove run_effect from HookContext (both dual-class branches); delete passthrough_effect; docstring now frames a decider as pure inline logic, with durable I/O delegated to an activity/step the consumer's hook dispatches.
• agent/loop.py: _make_hook_context no longer wires run_effect; drop the passthrough_effect import.
• executor.py: remove DirectExecutor.run_effect.
• tests: drop test_passthrough_effect_runs_inline + test_loop_before_tool_decider_can_run_effect; the MP …stamps_and_runs_effect test becomes …stamps_from_run_context (keeps the stamp assertion, drops the effect).

HookContext stays read-only (run_context + messages) plus the decision return.

Test plan

• ruff check + ruff format --check clean
• ty check exoclaw clean (only the pre-existing loop.py unused-ignore)
• CPython suite: 318 passed, 1 skipped
• MicroPython gate: passed

Follow-ups (separate)

• exoclaw-go: drop RunEffect/RunEffecter from the unpushed 0.30.x sync (deciders run inline).
• exoclaw-plugins #99: remove the now-moot run_effect DBOS durability test (keep the on_before_finish one).

[Copilot]

Pull request overview

This PR removes the run_effect seam from the agent lifecycle hook system so hook deciders remain pure, inline, and replay-safe by construction (with durable I/O expected to be dispatched via executor-native activities/steps instead of closures).

Changes:

• Remove run_effect (and the inline passthrough_effect fallback) from HookContext and stop wiring it in AgentLoop.
• Drop DirectExecutor.run_effect from the core executor implementation.
• Update CPython and MicroPython tests to stop asserting side-effect dispatch via hooks and keep only deterministic decider behavior checks.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
exoclaw/agent/hooks.py	Removes run_effect from HookContext (both CPython/MP branches) and updates hook contract documentation to emphasize pure deciders.
exoclaw/agent/loop.py	Stops importing/wiring passthrough_effect and builds a simpler, read-only HookContext.
exoclaw/executor.py	Removes DirectExecutor.run_effect implementation.
tests/test_hooks.py	Removes passthrough_effect/run_effect tests and updates hook contract test descriptions accordingly.
tests/micro/test_agent_loop_paths.py	Updates MicroPython hook-path tests to only validate run_context stamping (no run_effect).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.