fix(coding-agent): a review round is never refused for its length (TASK-901)

[yalexx]

What broke (TASK-901)

On 2026-09-18 run-q3tjkwm3 (TASK-899) opened PR #926, and the box could not start its review round:

Could not start a review round: The task is too long: at most 4000 characters.

The pull request was handed back needs_owner with its pr phase blocked, and no round was spent on it.

Why

startRun ran every task through normalizeTask, which refuses anything over MAX_TASK_CHARS (4 000). That limit was meant for what a person types, but it also applied to the follow-ups the box writes itself.

A review-loop round's task comes from buildReviewFeedback (src/lib/coding-review-state.ts). It quotes GitHub's failing-check logs (up to 5 × 2 000 chars) and unresolved review threads (up to 20 × 1 000 chars), and is capped at MAX_FEEDBACK_CHARS = 24 000. So any PR with a few CodeRabbit comments produced a round that startRun refused. PR #926 had 4 review threads totalling 17 385 characters (measured with gh api .../pulls/926/comments).

The round does not embed the original task text. It resumes the session that holds that task and names the pull request, branch and base. So there was nothing to truncate or summarise there. The run id is already how the original task is referenced.

The fix

• MAX_TASK_CHARS still applies, unchanged, to typed tasks (owner or assistant): startRun, createDraftRun, the /run and /draft routes and the MCP tool schema.
• Tasks the box writes itself are exempt: the automatic review pass (reviewOf), a review-loop round (reviewLoopOf) and a Vercel deploy-fix turn (vercelFixOf). No route passes those fields through from a request body, so a typed task can't claim the exemption.
• Those tasks have their own ceiling instead, MAX_BOX_TASK_CHARS = 32 000. That is above anything the builders produce today (review feedback ≤ 24 014). A task over it is clipped, ending in \n…(truncated), and never refused. The length of a task the box wrote is never the reason a follow-up fails to start.

The task text goes to the harness on stdin, and the MCP layer only shows its first line, so a longer task needed nothing else downstream. Team planner, worker and reviewer tasks already trim themselves to MAX_TASK_CHARS and are left as they were.

Tests

• coding-review-loop.test.ts › a round longer than the owner's task limit:
  • A completed run with a 6 000-character primary task. Its PR has a failing check (2 000-char log), CHANGES_REQUESTED, and four review threads at the parser's full 1 000 characters each.
  • The round starts: round 1, fixMode: "resumed", the fix run completes and the loop reaches clean.
  • The round's task is longer than MAX_TASK_CHARS and contains every thread and the check log, with nothing truncated. It does not re-embed the primary task.
  • With the fix reverted, the same test ends in needs_owner, which is the incident.
• The same block pins the worst-case buildReviewFeedback output (every cap overrun) and a 200-file reviewPassTask under MAX_BOX_TASK_CHARS.
• coding-agent.test.ts: a typed continuation over 4 000 characters is still refused invalid. A box-written round over MAX_BOX_TASK_CHARS starts, clipped to exactly that length with the marker.

Verification (this pass)

Run on this branch (origin/beta 67a8e6d + this commit). tsc ran first, on its own, with the larger heap this device needs:

• NODE_OPTIONS=--max-old-space-size=8192 bunx tsc --noEmit -p tsconfig.json: exit 0, no errors.
• NODE_OPTIONS=--max-old-space-size=8192 bun run test: exit 0.

  Test Files  1231 passed (1231)
       Tests  19198 passed | 1 skipped (19199)
    Duration  214.54s
  

• bunx eslint src/lib/coding-agent.ts src/tests/unit/coding-agent.test.ts src/tests/unit/coding-review-loop.test.ts: exit 0, no problems.
• With the coding-agent.ts change reverted and the new tests kept, all 3 new tests fail. The 6 000-character round test fails with expected 'needs_owner' to be 'clean'.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes

  • Box-generated follow-up tasks can now support up to 32,000 characters and are truncated when necessary.
  • User-entered tasks over 4,000 characters continue to be rejected.
  • Resumed review loops preserve all findings without duplicating the original task.
  • Review feedback and generated review tasks remain within supported task limits.

• Tests

  • Added coverage for task-length handling across resumed runs and review-loop scenarios.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository: ID-Robots/clawbox/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 063a6233-990c-4ef1-8e6d-85693c2270bb

📥 Commits

Reviewing files that changed from the base of the PR and between 67a8e6d and 0ef9101.

📒 Files selected for processing (3)

• src/lib/coding-agent.ts
• src/tests/unit/coding-agent.test.ts
• src/tests/unit/coding-review-loop.test.ts

Included review availability: 0 reviews are currently available. Your included PR review attempts over the past 7 days set your current allowance at 1 review per hour.

---

📝 Walkthrough

Walkthrough

The change adds separate size limits for typed and internally generated tasks. Oversized typed tasks remain rejected. Oversized review, review-loop, and deployment-fix tasks are truncated and marked before execution.

Changes

Task Length Handling

Layer / File(s)	Summary
Task limits and normalization src/lib/coding-agent.ts	Adds MAX_BOX_TASK_CHARS, detects internally generated tasks, truncates oversized generated tasks, and preserves rejection for oversized typed tasks.
Boundary and regression tests src/tests/unit/coding-agent.test.ts, src/tests/unit/coding-review-loop.test.ts	Tests typed-task rejection, generated-task truncation, review-round resumption, finding preservation, and generated task size bounds.

Priority: ⬇️ Low

Estimated code review effort: 3 (Moderate) | ~20 minutes

Change: Bug fix

Suggested reviewers: clawbox

🚥 Pre-merge checks | ✅ 6

✅ Passed checks (6 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the main change: preventing generated review rounds from being refused because of task length. It is concise and specific.
Description check	✅ Passed	The description provides a detailed summary, rationale, implementation details, testing scope, verification commands, and results. It does not use the template headings or complete the type-of-change …
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 1 functions across 2 files. (1 skipped: 1 …
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Persistent Tunnel Security	✅ Passed	PASS. The authoritative PR range changes only src/lib/coding-agent.ts and two coding-agent test files. The implementation changes task-length normalization for box-generated review and deployment fo…

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

CI Summary

✅ Tests

• Result: passed
• View run
• Coverage: statements 87.81%, branches 81.16%, functions 87.59%, lines 90.45%

✅ E2E

• Result: passed
• View run

✅ E2E Install

• Result: success
• View run