[rodrigobarona]

Fixes #158.

What

• Output budget: resolveAnthropicMaxOutputTokens now uses the 64k budget at every reasoning effort, still capped by the model's advertised maximum (and 16k when the maximum is unknown). Adaptive thinking shares max_tokens with the reply and the tool input, so the previous 16k cap at low/medium/high left too little room for multi-record write scripts. Anthropic bills only generated tokens.
• One-shot recovery: when a response stops with max_tokens, the runtime discards it instead of ending the turn. The truncated assistant message is not kept, and none of its tool calls were dispatched. It then retries once, with a user-turn instruction to keep reasoning brief and split large writes into smaller scripts. Any partial visible text stays in the transcript, followed by a paragraph break. A second truncation in the same turn still ends with the existing retryable incomplete error. model_context_window_exceeded is unchanged.

Tests

• Effort/budget table updated.
• New: truncated response containing a tool_use is discarded, MCP is never called, and no approval is raised. The retry request has no assistant turn and carries the recovery instruction.
• New: a second max_tokens stops with the retryable incomplete error.
• npm run typecheck, npm run lint, npm test (952 passed).

[roger-datocms]

(Claude here [Opus 5.5])

Triage

Status: still relevant. Master (6f57d960) still caps low/medium/high at 16k (dato-agent/src/lib/agentRuntime.ts:5065). It also ends the turn on any max_tokens stop (agentRuntime.ts:6122). The PR's base is current master.

Mergeability: MERGEABLE, merge state CLEAN. The repo runs no CI checks on this PR. I ran them locally on 4534285e:

• npm run typecheck: pass
• npm run lint: pass
• npm test: 952/952 pass

Review

The diagnosis matches the code. Adaptive thinking shares max_tokens with the reply and the tool_use input. 16k leaves too little room for a multi-locale write script.

Correct:

• The runtime discards the truncated assistant message before any tool dispatch. streamSingleResponse emits no tool activity, so the UI shows no orphan tool call or approval.
• Recovery runs once per loop state. cloneAnthropicLoopState spreads the flag, so approval resumes keep it.
• Mid-loop, the instruction lands as a text block after the tool_result blocks. The Messages API accepts that order.
• The cap still respects the model's advertised maximum. It falls back to 16k when the maximum is unknown.
• model_context_window_exceeded behaviour stays unchanged.

Minor points (non-blocking):

• The discarded response's visible text stays in the transcript, e.g. "Creating all sections now." The user can read it as a claim about work that never ran.
• systemPrompt.ts:209 still asks for "one focused script … so a correct request normally needs one approval". That pushes the first attempt toward the oversized script. A hint to split writes beyond a few records would avoid the wasted first attempt.
• A runaway low-effort turn can now generate up to 64k tokens before it stops. That means 4x the worst-case latency and spend. This is acceptable, but it belongs in the release notes.

Recommendation

Merge after a maintainer review. Address the systemPrompt.ts:209 wording here or in a follow-up. Fixes #158.

[roger-datocms]

I made a few additional changes:

• Raised max output length to 128k
• Added prompt caching to the Anthropic models
• Surfaced max-output errors to the editor
• Added some new tests