[saphid]

Problem

While a Claude Bash command runs, T3 Code tails Claude's task output file for live output. When the command writes more than 256 KiB between polls (for example cat of a large file or a noisy build), the tail skips ahead to the last 256 KiB. That byte offset can land inside a multi-byte UTF-8 character, so the live output after the skip starts with one or two � characters. Any non-ASCII output can hit this: accented text, CJK, emoji, or box-drawing progress bars.

#28643 fixed the same symptom when the file is truncated and rewritten. The skip-ahead path still has it, because a fresh decoder can't make sense of a cut that starts mid-character.

Change

After a skip, drop the leading UTF-8 continuation bytes before decoding. That's the same pattern the terminal history tail reader already uses (readHistoryTail in apps/server/src/terminal/Manager.ts). The cut line was already partial, so losing the rest of one character costs nothing.

Scope and approval

Very small, focused fix for an obvious bug, so no prior issue: the function already skips ahead on purpose and resets its decoder for this case. The reset can't help when the cut lands inside a character. The change is 9 non-test lines in one function. The final tool result isn't affected, only live output.

Overlap check: #28643 (merged) fixed only the rewind-after-truncation path. Open #26501 (id validation) and #31855 (a new tail reader for persisted output) touch the same file but not this path.

Verification

• Extended the existing "skips ahead when the command writes faster than it is read" case so the flood is é\n lines and the cut lands on the second byte of an é.
• Base c6ca6056: cd apps/server && vp test run src/orchestration-v2/Adapters/ClaudeTaskOutputTail.test.ts → 1 failed, AssertionError: expected '\n�\n' to equal '\n\né'.
• With the fix: 5 passed.
• vp run check:changed: format, lint, knip, typecheck (3 packages) and tests (2 files, 7 tests) all pass.
• Runtime check against a real temp file with the real tailClaudeTaskOutput (300 KB of 日本語\n lines written in one go after the first read):
  • base: skip chunk starts "\n��本語\n日本"
  • this branch: "\n本語\n日本語\n"
• Parent verification with the actual base/head reader and the real terminal-output normalizer: a 360,005-byte ANSI-styled UTF-8 file retains one replacement character on base and zero on the candidate; neither normalized result is truncated. This checks that terminal normalization does not mask the fix.
• Not checked: a live Claude session flooding output in the app. This change only affects live output, which the final tool result replaces.

Reviewed with sol-loop: 1 round with GPT-6.1 Sol on high. Sol verdict: approved, no findings.

Created with Claude Opus 5.5 in Claude Code (T3 Code).

🤖 Generated with Claude Code

[macroscopeapp]

Approvability

Verdict: Approved at f614805

Macroscope's review found this PR approvable — This prevents stray replacement characters when Claude live output skips into the middle of a UTF-8 character. The change is limited to live-output presentation; command execution and final results remain unchanged.

^{You can add or adjust custom eligibility rules. Learn more.}

[coderabbitai]

No actionable comments were generated in the recent review.

Walkthrough

The tailer now skips leading UTF-8 continuation bytes when a read starts after skipped output. The fast-writer test uses repeated é characters and checks the chunk boundaries.

Changes: Claude task output tailing

Layer / File(s)	Summary
Skip-ahead UTF-8 tail handling apps/server/src/orchestration-v2/Adapters/ClaudeTaskOutputTail.ts, apps/server/src/orchestration-v2/Adapters/ClaudeTaskOutputTail.test.ts	When a read starts after skipped output, the tailer skips leading UTF-8 continuation bytes before decoding. The fast-writer test checks output containing repeated é characters and retains the chunk-length assertion.

Estimated code review effort: 2 (Simple) | ~8 minutes

Change: Bug fix

Possibly related PRs

• pingdotgg/t3code#28643: It changes decoder reset behavior in tailClaudeTaskOutput when the file shrinks below the read offset.

Suggested labels: bugfix

Suggested reviewers: scratchyone, macroscopeapp

Merge Risk: Minimal · up to f6148

No actionable merge-blocking risk remains from the supplied review evidence.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title names the actual change: Claude live command output no longer starts with a replacement character after the tail skips ahead. It follows the conventional commit format and matches the summar…
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.
Description check		The PR description is thorough and follows the repository template. It has a Problem section that explains the skip-ahead path can start inside a multi-byte UTF-8 character and produce replacement cha…

Full details: Title check

Explanation

The title names the actual change: Claude live command output no longer starts with a replacement character after the tail skips ahead. It follows the conventional commit format and matches the summary, the description, and the objectives. The title does not need to list the test change.

Full details: Description check

Explanation

The PR description is thorough and follows the repository template. It has a Problem section that explains the skip-ahead path can start inside a multi-byte UTF-8 character and produce replacement characters, with examples of affected content. It has a Change section that describes dropping leading continuation bytes after a skip and cites the existing readHistoryTail pattern. It has a Scope and approval section that explains why no prior issue is linked: it is a very small, focused fix of an obvious bug, and it notes the overlap with related PRs. It has a Verification section with the exact test command, base and head results, the full check run, a runtime check with real files, and an explicit statement of what was not checked. The description also ends with the model and harness used, as the template asks. The title is not in the template's conventional commit form in the description body, but the PR objectives show a conventional-style title. The main gap is that the Scope section relies on the author's own judgment that the fix is an obvious bug, without a maintainer-approved issue or discussion. The template allows that exemption when the author explains why the fix qualifies, which the description does. Overall the description is mostly complete and specific.

Finishing Touches

Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: c6ca605
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: c5e7976b-d256-427d-a80f-39e2877a10f5

Commits

Reviewing files that changed from the base of the PR and between c6ca605 and f614805.

Files selected for processing: 2

• apps/server/src/orchestration-v2/Adapters/ClaudeTaskOutputTail.test.ts
• apps/server/src/orchestration-v2/Adapters/ClaudeTaskOutputTail.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

Verified in a real dev environment

verified unchanged head f614805ae28923cc640d20e8357c577c449e8c8e against github/main at e89abadaee70e0d73fb19a5ce40a686b1a92310e on linux. installed both checkouts with vp i and ran one isolated dev server at a time. created the project and threads through the web app using playwright and /usr/bin/google-chrome; no mocks or seeded data.

the changed path is claude-specific, so i exercised real claude haiku 5.5 foreground bash turns through the local proxy. both emitted task_started with task_type: local_bash and is_backgrounded: false. i expanded the running command row before the flood and inspected its live output, before the final result replaced it.

same command on both revisions:

node -e 'process.stdout.write("start\n");setTimeout(()=>{process.stdout.write("\x1b[0m\x1b[0mé\n".repeat(40000)+"end........\n");setTimeout(()=>{},30000)},5000)'

the 440,018-byte output forces the 256 kib tail cut onto byte a9, inside é. the ansi resets keep the decoded text below the client's display limit, so it retains the boundary.

main live: 47,681 characters; 1 replacement character; prefix "start\n\n�\né\n"
head live: 47,680 characters; 0 replacement characters; prefix "start\n\n\né\n"
both final results: 0 replacement characters; both turns completed
head neighbors: ascii output, UTF-8 bytes split across writes, and stderr output passed
real codex gpt-6.1-sol command turns passed on main and head
blacksmith: check:changed c6ca6056309904f89b90b6055d797973be0a0939 passed
format, lint, knip, graphql, typecheck (3 packages), tests (2 files / 7 tests)

before, while the command is running:

after, with the same command:

Not verified

native electron, ios/android, macos/windows, and relay/tunnel or multi-environment connections were not exercised on this linux vm. no source changes were needed.