feat: let an AgentBatch finish the task in one call

[SSHdotCodes]

What and why

Follow-up to #160. Benchmarking the AgentBatch default against the previous build with gpt-5.6-sol (medium, 6 tasks × 2 rounds, arms interleaved) came out 0.90× on wall clock: a strong model already wrote whole tasks as one browser snippet returning {finalAnswer} (one model turn), while the batch protocol cost a spec turn, a steps turn, and a done turn. Model turns decide task time; the delay between browser actions does not.

This PR makes a batch as cheap in turns as the snippet path:

• answer: a final-answer template rendered from step results once every step succeeded ({stepId} = the step's text, value, or URL; {stepId.field} = any scalar field). The built-in agent finishes the task with the rendered finalAnswer in the same turn. A placeholder with no value leaves finalAnswer unset and explains itself in answerError, so a stopped or empty read never becomes a claimed answer.
• Spec call optional: guidance, tool descriptions, and the skill text are now turn-centric — when the task names its targets, one batch (goto first, read last, answer) or one snippet returning {finalAnswer}; {url} is for unknown pages.
• Docs, changelog, regenerated SKILL.md, and the public types updated; budgets for the agent and MCP tool lists raised (6,000 → 6,500 and 9,000 → 9,250) for the new option and its guidance.

Benchmark, honestly

Same rig as before (gpt-5.6-sol, effort medium, Codex OAuth; 6 tasks × 2 rounds, arms interleaved; "old" = pre-#160 build, "new" = this branch). All 24 runs correct on both arms.

	old (snippets)	#160 (batch default)	this branch
model turns	27	33	28
total wall clock	167 s / 156 s	185 s	207 s
input tokens	49.8k / 48.8k	72.3k	65.8k
output tokens	5.7k / 5.5k	5.9k	6.9k

(old was measured in both sessions; both values shown.) The turn penalty of #160 is gone: turn counts are at parity, and login dropped from 4 turns to 2. Wall clock is not better in this sample: the new arm emits about 25% more output tokens per turn, and with only two rounds per cell the per-turn latency noise is large (the same 3-call form task took 25.6 s on one arm and 10.8 s on the other). gpt-5.6-sol did not use answer in any run (it chose browser snippets in 9 of 12 runs, batch in 3), so the one-call finish is exercised by the tests, not yet by this model in the wild. So: this PR removes the structural turn cost; it does not demonstrate a speedup with gpt-5.6-sol. Raw results: ~/bw-bench/results-2 on the author's machine.

Checklist

• bun run release:check passes locally (versions, lint, typecheck, build,
  unit tests, published declarations, tarball).
• Every browser connection still goes through the guard proxy. No new
  launch path, transport, or fetch.
• No secret value reaches the model sandbox. The answer template is
  rendered from step results that are already redacted (password readings
  are [redacted]), and the rendered result goes through the envelope's
  redaction like any other field.
• Dependency pins are still mirrored everywhere. No pin changed.
• Public API changes update types/*.d.ts in this same commit.
  types/common.d.ts gains AgentBatchOptions.answer and
  AgentBatchResult.finalAnswer / answerError.
• The two branch-protected CI job names are unchanged; no workflow changes.
• No unit test imports src/worker.ts or dist/src/worker.js directly.
• User-visible behaviour is reflected in CHANGELOG.md, docs/agent-batch.md,
  README, docs/{agent,javascript,getting-started}.md, and the regenerated SKILL.md.
• Nothing private is being committed.

How it was verified

• bun run release:check and the managed-browser suite (BETTERWRIGHT_REQUIRE_BROWSER=1 bun run test) on Bun 1.4.0 / macOS arm64, including a new end-to-end check that a batch with answer returns the rendered finalAnswer.
• Unit tests for template rendering (shorthand fall-through, explicit fields, unknown step/field refusal, no answer on a stopped batch) and for the agent loop finishing in one turn from a batch finalAnswer.
• The A/B benchmark above, re-run on this branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Td3ZP4tb9UKEQHLaoAo6NN

[greptile-apps]

Greptile Summary

Summary

Answer templates now reserve braces exclusively for valid step-result placeholders and return a recoverable answer error for malformed input.

Two previously reported issues were disproved by an executed focused runtime check:

• Unsupported placeholders bypass validation: {1price} and {step.1} now throw during rendering, and batch execution returns answerError without finalAnswer.
• Nested braces bypass validation: {{flash}} now throws during rendering, and batch execution returns answerError without finalAnswer.

The same check confirmed that stray closing braces and unclosed opening braces are rejected through both rendering and batch execution.

Confidence Score: 5/5

No blocking failure remains.

The executed checks disproved both previously reported answer-template failure paths, and no accepted blocking finding remains.

T-Rex Logs

What T-Rex did

• Ran the brace-template-validation.ts script from the repository root to exercise the focused runtime harness against five invalid brace templates; renderAgentBatchAnswer and executeAgentBatch were invoked with {1price}, {step.1}, {{flash}}, {flash}}, and {flash}, and all five templates threw during direct rendering, returning answerError with no finalAnswer in batch execution.
• Encountered a prerequisite blocker during the initial setup: Bun 1.3.14 reported Unknown lockfile version for Bun v2, halting progress.
• Validated the post-blocker behavior: the after-log shows each of the five invalid templates has finalAnswer: null and an expected answerError.

_{Ran code and verified through T-Rex}

_{Reviews (3): Last reviewed commit: "fix: reserve braces in answer templates ..." | Re-trigger Greptile}

[SSHdotCodes]

@greptileai review — the latest commit rejects malformed answer placeholders (every brace group must match {stepId} or {stepId.field}; otherwise answerError, no finalAnswer), with tests through both the renderer and executeAgentBatch.

[SSHdotCodes]

@greptileai review — the latest commit reserves braces in answer templates: the renderer scans linearly, every { must open a placeholder ending at the next }, and nested ({{flash}}), stray, or malformed braces produce answerError with no finalAnswer; tests cover {{flash}}, {{flash}, {flash}}, and stray braces through the renderer and executeAgentBatch.