feat(agent): add pre_verify hook and coding guidance config

[OutThisLife]

Summary

Adds two related coding-agent controls while keeping verification closure owned by the existing verify-on-stop evidence gate.

1. pre_verify hook

Adds a pre_verify event fired when the agent edited files and is about to accept a final answer, immediately after the existing verify-on-stop guard. A plugin or shell hook can keep the agent going one more turn — run a check, defer it, tidy the diff — by returning:

{"action": "continue", "message": "<follow-up for the model>"}

The Claude-Code Stop shape ({"decision": "block", "reason": "..."}) is accepted too. Hooks receive coding, attempt, final_response, and sorted changed_paths so they can self-scope and self-throttle. The hook path is bounded by agent.max_verify_nudges (default 3), preserves message-role alternation, and is silent to the user while it nudges.

Important: Hermes does not add a second built-in verification decision through pre_verify. The built-in decision remains owned by verify-on-stop and the passive verification evidence ledger. pre_verify is for user/plugin/shell policy.

2. Default verify-on-stop guidance

Hermes still ships the default coding guidance, enabled by agent.verify_guidance: true. The default text is preserved:

[Coding] Before you run tests/linters or call this done: if this is creative UI/visual work, hold off on tests and linters until the user says they like the result or you're about to commit. And before every commit, clean your work: keep it KISS/DRY, match the surrounding code style, and be elitist, shorthand, clever, concise, efficient, and elegant.

The behavior change is where it lives: this guidance is appended to the existing verify-on-stop missing-evidence nudge, rather than emitted as an independent default pre_verify continuation. That puts the workflow rule at the exact verification decision point without creating a second default extra model turn.

3. agent.coding_instructions

Adds agent.coding_instructions as a string or list appended to the coding brief as its own stable system block. It is coding-posture only, cache-safe, and takes effect next session.

Scoping

• verify-on-stop: remains the only built-in verification closure path; it decides when missing evidence requires another turn.
• agent.verify_guidance: default-on shipped guidance appended to that missing-evidence nudge; set false to keep the nudge terse.
• pre_verify: user/plugin/shell hook mechanism with no separate shipped default continuation.
• agent.coding_instructions: stable cached standing guidance for coding sessions.

Footprint

One new hook event, one small helper module (agent/verify_hooks.py), three config keys under agent (coding_instructions, verify_guidance, max_verify_nudges), no env vars, and no new model tools.

Test plan

• scripts/run_tests.sh tests/agent/test_verify_hooks.py tests/agent/test_verification_stop.py tests/hermes_cli/test_plugins.py tests/agent/test_shell_hooks.py -q
• scripts/run_tests.sh tests/agent/test_verify_hooks.py tests/agent/test_verification_stop.py tests/agent/test_coding_context.py -q
• Linter check on edited Python files
• Manual diff review for stale default-pre_verify wording and local-only /clean references

[tonydwb]

COMMENT: Well-structured feature adding pre_stop round-end hook and configurable coding instructions. 14 files (just under the 15-file threshold) with 413 additions. The coding_instructions config is cache-safe and properly isolated as a separate system block. The pre_stop hook integrates cleanly with the existing hook system. However, the surface area spans agent core (coding_context, conversation_loop, stop_hooks), config, plugins, and 4 test files. Recommend verifying: (1) the coding_instructions block doesn't break prompt caching when empty, (2) the pre_stop hook fires reliably across all exit paths (cancel, error, normal stop), (3) the hook registration in plugins.py doesn't conflict with existing hooks. Test coverage looks solid.