feat(agent): add LoopDetectHook and ReflectRetryHook for agent self-correction

[MuataSr]

Summary

Agents frequently get stuck in two failure modes:

1. Tool-call loops — repeating identical tool calls without progress until hitting iteration limits
2. Blind retries — failing with the same tool call and arguments, wasting iterations on the same error

This PR adds two lightweight hooks that plug into the existing AgentHook system to help agents self-correct before exhausting their iteration budget.

What Changed

AgentHook — Two new hook points

• before_execute_tools(tools) — runs before tool batch execution
• after_tool_call(name, args, result) — runs after each individual tool call

LoopDetectHook (new)

• Tracks consecutive identical tool-call signatures (SHA256 fingerprint of name + args)
• At warn_at (default 3): injects in-context warning nudging the model to change approach
• At halt_at (default 5): injects stronger message explicitly telling the model to stop calling tools
• Zero runner changes beyond the hook plumbing — works entirely through the hook system
• Configurable thresholds for different tolerance levels

ReflectRetryHook (new)

• Intercepts tool error results and enriches them with attempt history
• Same call with same args = same dedup key (tool_name + sha256(args))
• First error: adds structured hint with "Things to try differently"
• Repeated same error: accumulates history showing the model what it already tried
• After max_retries (default 2): hint pivots to "try a fundamentally different approach"
• Pure transform — no re-execution, the model decides whether to retry

Testing

• 32 tests, all passing — composite hook fan-out, error isolation, loop detection thresholds, retry accumulation, dedup, exhaustion
• Stress tested with 70 parallel subagent tasks (JSON data analysis, math computations, file writes) — 100% success rate, zero loops

32 passed in 0.68s

Files Changed (7 files, +560 −5)

File	Change
nanobot/agent/hook.py	Added before_execute_tools and after_tool_call to AgentHook base class
nanobot/agent/loop_detect.py	New — LoopDetectHook (142 lines)
nanobot/agent/reflect_retry.py	New — ReflectRetryHook (127 lines)
nanobot/agent/runner.py	Hook plumbing for the two new hook points
nanobot/agent/loop.py	Pass hooks through to runner
tests/agent/test_hook_composite.py	22 tests for composite hook system
tests/agent/test_reflect_retry.py	10 tests for ReflectRetryHook

Design Decisions

• Hooks over core changes: These plug into the existing hook system rather than modifying the core loop. Users opt in by configuring hooks, not by changing framework behavior.
• In-context nudges, not hard stops: Loop detection injects messages into the conversation. The model can still choose to continue — it just gets better information to decide.
• Pure transforms: ReflectRetryHook doesn't re-execute anything. It enriches error messages so the model self-corrects on the next iteration.
• Backward compatible: All new hook methods have no-op defaults. Existing hooks continue working unchanged.

Relates to

• Partially addresses Add subagent profiles with configurable tools and skills #1012 (subagent reliability improvements)