Problem
The spec describes agent lifecycle hooks (prepare, filter, response retry config), but they are not implemented.
Why it matters
- Users need per-agent customization (pre-turn context, token budgeting, retry behavior) without forking core code.
Where this shows up
TECHNICAL_DEBT.md → “Agent Hooks” (Category 3).
Proposed fix
- Implement agent config fields:
prepare: callable run before each turn; output becomes {prepared.*} (when/if that namespace is implemented)
filter: token budget / message filtering policy
response: retry policy (retries, retry_delay, etc.)
- If
{prepared.*} is not implemented yet, define a minimal path (e.g., only allow prepare to mutate session/context).
Acceptance criteria
- Hooks are parsed, validated, and executed according to spec (or spec is updated to match final design).
- Tests cover hook execution order and retry behavior.
Problem
The spec describes agent lifecycle hooks (
prepare,filter,responseretry config), but they are not implemented.Why it matters
Where this shows up
TECHNICAL_DEBT.md→ “Agent Hooks” (Category 3).Proposed fix
prepare: callable run before each turn; output becomes{prepared.*}(when/if that namespace is implemented)filter: token budget / message filtering policyresponse: retry policy (retries,retry_delay, etc.){prepared.*}is not implemented yet, define a minimal path (e.g., only allow prepare to mutate session/context).Acceptance criteria