## Context
AI chat spills tool outputs larger than `MAX_INLINE_TOOL_OUTPUT_BYTES`
(16 kB) to a file and lets
the model page them back with `search_output` / `extract_json_paths`.
Three paths bypass this and
let unbounded payloads into conversation history:
1. **`learn_tools` is never spilled.** Tool schemas go inline whatever
their size.
2. **Navigation tools have no inline cap.** They are exempt from
spilling by design (they page
spilled files), but nothing bounds their own output.
3. **Spill failure falls back to full inline.** On any spill error the
service returns the complete
payload with only a warning appended.
## What changed
- **`learn_tools` now spills.** `createLearnToolsTool` takes `{
excludeTools?, spillLargeOutput? }`
(same shape as `createExecuteToolTool`); chat execution enables it. Only
the bulky `tools`
schemas are spilled; `message` / `notFound` / `suggestions` stay inline
and the response carries
a `spilledTools` envelope (fileId, preview, hint) pageable via
`extract_json_paths`. MCP is
unchanged.
- **Navigation tools get a hard inline cap.** Still never spilled, but
output above 16 kB is
head+tail truncated with a marker telling the model to narrow the query
or page with `offset`.
- **Spill failure truncates instead of inlining.** The fallback returns
head+tail within the 16 kB
budget with the original byte size in the marker, keeping the warning.
- New `truncateHeadTail` util: byte-budgeted, marker-aware, UTF-8
codepoint-safe.
## Test plan
- `tool-output-spill.service.spec.ts`: spill envelope unchanged,
under-budget passthrough,
navigation cap for both tools (budget respected, marker mentions
`offset`, no file written),
truncated fallback on spill failure with warnings preserved.
- `learn-tools.tool.spec.ts`: no spill without the option, inline under
budget,
`message`/`notFound`/`suggestions` intact when spilled, spill-failure
warnings surfaced.
- `truncate-head-tail.util.spec.ts`: budget, head+tail+marker,
multibyte-safe cuts.
- 63 tests across 6 suites; `lint:diff-with-main` and `typecheck` green.
## Post-deploy
Watch the `AiChatToolOutputTokens` histogram (p95 should collapse to ~4k
tokens) and the
`AiChatInputTokens` / `AiChatCacheReadTokens` ratio on GPT-5-class
models.
<!-- This is an auto-generated description by cubic. -->
<a
href="https://cubic.dev/pr/twentyhq/twenty/pull/23286?utm_source=github"
target="_blank" rel="noopener noreferrer"
data-no-image-dialog="true"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-light.svg"><img
alt="Review in cubic"
src="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"></picture></a>
<!-- End of auto-generated description by cubic. -->
Etienne
4851489ebc
