feat: onAfterResponse webhook for post-response integrations

[cpfiffer]

Summary

Add a config-driven webhook that fires after every agent response, enabling external integrations (analytics, logging, observability) without modifying core code.

Motivation

Community user (bibbs) runs 3 lettabot instances on a Raspberry Pi with a local analytics server that evaluates agent responses after each exchange. Currently requires patching bot.ts, config/types.ts, and main.ts on every upstream rebase (~17 lines). This is the most generally useful hook for the community -- anyone doing logging, analytics, or external integrations would benefit.

Related to #388 (message-level hooks) but much narrower in scope. This proposal is fire-and-forget only (no response transformation), which makes it significantly simpler to implement and maintain.

Proposed Config

features:
  webhooks:
    afterResponse:
      url: http://localhost:3847/api/evaluate
      headers:
        Authorization: "Bearer ${ANALYTICS_TOKEN}"

Implementation Notes

• Fire-and-forget POST after sentAnyMessage is confirmed and response is non-empty
• Should never block or affect the response pipeline
• Body should include: agent ID, user text, agent response, channel, chatId, timestamp
• Environment variable interpolation in URL and headers (already have precedent in config)
• Failure should be logged but never surface to the user

Context

This is one of 5 issues filed from a community user's multi-bot deployment feedback. See also: #388 for the broader message hooks discussion.

[cpfiffer]

Update: Letta Code Stop hook may already cover this

On review, the Letta Code hook system's Stop hook fires when the agent finishes responding and receives user_message, assistant_message, and preceding_reasoning. A shell hook could curl to an analytics endpoint with this data -- no lettabot changes needed.

Example:

#!/bin/bash
input=$(cat)
curl -s -X POST http://localhost:3847/api/evaluate \
  -H "Authorization: Bearer $ANALYTICS_TOKEN" \
  -H "Content-Type: application/json" \
  -d "$input" &
exit 0

What the Stop hook provides: user prompt, agent response, reasoning
What it doesn't provide: lettabot-specific metadata (channel, chatId, group context, delivery status)

If the analytics use case only needs the exchange text, this is already solved. If channel/chat metadata is required, that's a real gap -- but could be addressed by enriching the Stop hook input with session metadata rather than building a separate webhook system in lettabot.

Leaving this open for now pending confirmation of whether the Stop hook approach works for the original requester's analytics pipeline.

[cpfiffer]

Update: Blocked by letta-code headless hook support

After investigation, the Stop hook provides exactly the right data for this use case (user_message, assistant_message, preceding_reasoning). However, lifecycle hooks don't fire in headless mode -- they're only wired in the interactive CLI (App.tsx), not in headless.ts which is what the SDK subprocess uses.

Tool hooks (PreToolUse, PostToolUse) already work in headless because they're in tools/manager.ts. The gap is only lifecycle hooks (Stop, SessionStart, SessionEnd, UserPromptSubmit).

Filed: letta-ai/letta-code#1282

Once headless hook support lands (~50-80 lines in headless.ts), this issue is solved without any lettabot changes. Users would add a Stop hook to ~/.letta/settings.json that curls to their analytics endpoint.

Keeping this open as blocked-by until letta-code#1282 is resolved.