Smarter truncation for oversized script results

Status summary: implemented; unit tests green (typecheck/lint/knip/format clean). CI green on PR #2400. Followup task filed for typed Results<...> handles.

When an agent script returns a big result, renderScriptSettlement slices the pretty-printed JSON at scriptResultHistoryLimit (30k chars) — bisecting JSON mid-key and showing only the start of the payload (often one giant array's first entries). The model learns nothing about overall shape; in this chat the agent had to guess a prior result's structure and got it wrong.

Fix: on overflow, render (1) an inferred TypeScript type of the whole value, (2) a small inspect-style preview (a few items per array, capped strings/depth, valid JSON with truncation markers), (3) the existing spill notice. The full result already spills to script-results/<id>.json; the type tells the agent how to write the follow-up readFile script.

Decisions (grilled with Misha 2026-08-04):

• Drill-in stays the spill-file + itx.workspace.readFile recipe. Typed Results<"agent-output-400"> cast is a followup (see tasks/typed-script-result-handles.md).
• Type inference is hand-rolled (sync, zero deps, Workers-safe) — not quicktype (Workers compat scar tissue: the quicktype-core>readable-stream pnpm override).
• Preview builds on the existing truncateJsonToBytes (was PostHog-only) with an aggressive arrays/strings/depth policy pre-pass.
• Budget shrinks on overflow: type ~3k chars + preview ~10k bytes + notice. 30k stays the passthrough threshold, per-agent patchable as before.

Checklist#

• apps/os/src/lib/infer-json-type.ts — inferJsonType(value, {maxChars}) → TS type text: structural merge of array elements (union cap ~3), cardinality comments (// 1234 items, // ~45k chars), small literal unions for repeated short strings, budget enforced by collapsing deepest subtrees to unknown shape lattice + merge + depth-shrinking renderer; also renders wide keyed maps as Record<string, T> and drops undefined-valued keys to match the spilled JSON
• unit tests apps/os/src/lib/infer-json-type.test.ts 11 tests incl. optionality-from-undefined and budget enforcement
• move apps/os/src/domains/integrations/truncate-json.ts (+ test) → apps/os/src/lib/, update posthog import; no behavior change git mv; only importer was posthog.ts
• add previewJson(value, {maxArrayItems, maxStringChars, maxDepth, maxBytes}) to truncate-json — policy pre-pass reusing existing marker idiom, then byte-budget guarantee policy pass over the MeasuredJson tree (tiny ≤100-byte subtrees survive untouched), then truncateJsonToBytes for the ceiling
• rewire renderScriptSettlement JSON overflow branch: type block + preview block + spill notice (recipe now const data: Result = JSON.parse(...)); raw-text results keep slice path with ~10k preview; failure branch unchanged; inferJsonType failure degrades to no type block renderOversizedJsonResult in agent-processor-implementation.ts; budgets 3k type / 8k preview / 10k raw-text, each clamped to historyLimit
• update system-prompt one-liner in agent-defaults.ts (~line 207)
• update render expectations in agent-processor.test.ts (~1040–1140); e2e spill test should pass as-is JSON spill test asserts type block + typed recipe; new test covers array-eliding preview with 400 rows
• file followup tasks/typed-script-result-handles.md

Implementation log#

• 2026-08-04: implemented as specced. Full os unit lane green (2545 passed). Demo render on a Slack-history-shaped 270KB payload: type block ~20 lines with thread_ts?: string optionality and literal unions, preview 3 messages + [truncated 797 items; from 271814 JSON bytes] marker — total rendered item ~2.5KB vs 30KB before.