Cap intent LLM request text at 64 KiB so classify cannot unbounded-prompt the provider

[cursor]

Problem

POST /v1/intent already rejects message bodies above 64 KiB, but orchestrator depth-0 hybrid/llm classify (apply_intent_ingress → resolve_intent → build_llm_user_prompt) embeds in.text wholesale. TUI, HTTP chat, and POST /v1/events can carry much larger payloads (HTTP request bodies are capped at 16 MiB), so a single ingress turn can send an unbounded provider prompt — cost and availability, not a style nit.

Heuristic classify is unchanged: it still sees the full text. Only the LLM user prompt is capped.

Fix

• kIntentLlmTextMaxBytes = 64 * 1024 (same number as the HTTP route).
• build_llm_user_prompt UTF-8-safely truncates request text and appends \n[truncated], matching presence/advisor field caps.
• Isolated to src/intent.cpp + include/intent.h + tests/test_intent.cpp.

Tests

unit_intent:

• mode=llm with 64 KiB + 4 KiB of x produces a [REQUEST] body ≤ 64 KiB that includes [truncated].
• A request of exactly 64 KiB is forwarded unchanged.

Suite 27/27 locally (100 assertions) + ASan + UBSan.

Independently mergeable against main (0456350). git merge-tree --write-tree vs #321, #347, #350, #351, #353, #355, #356, #357, and #358 is CLEAN including CHANGELOG this run. Do not re-fix #321–#358.

  

---

Note

Low Risk
Localized intent prompt shaping with tests; heuristic routing and HTTP reject behavior are unchanged aside from preventing oversized LLM prompts.

Overview
Orchestrator hybrid/llm intent classify now caps the [REQUEST] block in build_llm_user_prompt at 64 KiB (kIntentLlmTextMaxBytes), matching POST /v1/intent. Oversized TUI, chat, or event text is UTF-8–safely shortened and tagged with \n[truncated] before the provider call.

Heuristic classify is unchanged and still sees the full in.text. New unit_intent cases cover over-limit truncation and exact-boundary passthrough; CHANGELOG documents the fix.

^{Reviewed by Cursor Bugbot for commit 2ca3ecc. Bugbot is set up for automated code reviews on this repo. Configure here.}