With a PR/MR URL, Jury resolves an isolated checkout. The default state root is ~/.jury:
checkouts live in ~/.jury/checkouts/, runs in ~/.jury/runs/.
--dir <path> uses <path>/checkouts/ and <path>/runs/ instead.
Your calling repository is not switched to the PR branch.
Without a URL, Jury works directly in the current repository, or the repository selected by --dir.
Other commands use the current Git repository when available, otherwise ~/.jury.
Explicit --dir "" selects ~/.jury. Use the same state root when listing or opening saved runs.
jury runs --dir ""
jury --web-only --dir "" --run <slug>
jury finding list --dir "" --run <slug>Run artifacts include event logs, generated prompts, reviewer output, and folded console data.
Agent stdout and stderr are decoded as UTF-8 across process-output chunks, preserving
Unicode text and filenames in live output, saved reports, and failure diagnostics.
Reviewer replies also preserve literal {{...}} text from findings and verdicts;
command-template expansion does not reinterpret the reply's contents.
Timeline durations use recorded event timestamps; rounding is applied only to display labels,
so short reviews retain their elapsed seconds in the saved console data.
They are retained locally; remove old records manually when you no longer need them.
Related-PR task checkouts retain local commits for resume. Single-PR URL checkouts are removed when the CLI finishes, including local-only commits.
For durable single-PR local-only fixes, review your own source-branch checkout with
jury --dir /path/to/checkout --trunk <base> --push=false (no URL), or export commits
from the temporary checkout while the run is active.
The console and standalone demo show TRAE's own mark beside its participant name in timelines, rounds, conversations, and exchanges. The icon is embedded in the page and follows the participant color in both light and dark themes, without requiring an external image download.
Closing or reloading the console page releases that connection's live-stream watcher and timers, including while its initial event log is loading. The review continues independently of the browser connection.
Use --port 0 with a review or --web-only to let the OS choose an available
console port. Jury prints and opens the URL with that assigned port. A requested
nonzero port still falls back to the next available port when it is busy, trying
up to 20 higher ports without exceeding 65535. If that range is occupied, Jury
reports the address-in-use error; choose another port or use --port 0.
jury review <same-pr-url> --resume <slug> --reviewer claude --push=false
jury review <same-pr-url-1> <same-pr-url-2> --resume <slug> --push=falseRepeat the same --dir root if you used one. Keep related-PR URLs in the same order.
All reviews save their push setting and retain it on resume unless explicitly overridden.
Older runs without a saved setting resume with pushing disabled; pass --push=true
to enable it explicitly. For related PRs, --push=true publishes retained fixes
with fast-forward pushes. If a remote branch changes
independently, start a new task rather than resuming obsolete heads.
--rounds 3 caps review rounds; it is not a monetary budget. Each round may involve multiple
reviewers, judge calls, and replies. Costs and latency depend on provider subscriptions and usage.
A review that reaches its round limit remains incomplete in the saved state. Do not use exit code
alone as evidence of approval; inspect target.state and the reviewed commit in the run record.
For single-checkout reviews, Jury checks that HEAD still matches the commit used
for the review or reviewer replies before reporting completion. If another process
commits during that phase, the run remains incomplete with HEAD changed during review;
inspect the new commit and run the review again. This check does not lock the checkout.
When committing judge fixes, Jury includes new files even if Git's
status.showUntrackedFiles display setting hides them. Ignore rules still apply,
and Jury does not change the saved Git setting.
Reviewer execution failures exit nonzero with --web=false. With the console enabled, it stays
available for inspection until Ctrl-C.
URL routing distinguishes route segments from owner, repository, and subgroup names.
Names such as pull and merge_requests are preserved, including when removing
view suffixes, query strings, and fragments from related-PR URLs. Saved review
labels use the same request number as checkout resolution. GitHub
/owner/repo/pull/<id> and GitLab /<group>/<project>/-/merge_requests/<id>
forms are supported; existing separatorless GitLab links continue to work.
GitHub uses gh for PR metadata and cloning plus Git for commits and pushes. Authenticate both
as needed (gh auth login, your SSH key or Git credential helper). Source-branch push permission
is needed when pushing is enabled.
GitLab and self-hosted compatible services use Git credentials and standard
refs/merge-requests/<id>/head refs, including sharded revision refs where available.
MR descriptions are unavailable and the resolver uses the remote default branch as the base.
Use URL mode only for MRs targeting that branch. If the source branch is ambiguous or on a fork,
check it out yourself and run jury --dir /path/to/checkout --trunk <target-branch>.
For an offline local MR review with --push=false, Jury can use the caller's cached
origin/HEAD when default-branch discovery cannot reach the remote. The caller's
origin must match the MR repository and the cached default must resolve to a commit.
Without that cache, pass --trunk <target-branch> with its cached origin tracking ref.
Jury never infers the default from the branch currently checked out in the caller.
Related PRs may mix GitHub and GitLab-style URLs. Two PRs sharing a source branch are rejected.
jury reply exits nonzero if any selected reviewer fails to run or returns an
invalid result. It saves every reply before reporting the failure. After correcting
the agent issue, retry jury reply with the same --dir and --run, optionally
selecting the failed reviewer with --reviewer. A successful invocation that raises
remaining findings still exits zero; inspect its report for the review verdict.
| Symptom | Next step |
|---|---|
| Missing judge or reviewer | Install its CLI, select installed agents, or disable unused entries in jury.config.json. |
| Login, quota, timeout, or nonzero reviewer exit | Read the saved report, fix the provider issue, and resume. A failed reviewer never counts as approval. |
| Console stays open after completion | Expected. Press Ctrl-C; use --web=false for scripts. |
| Saved run cannot be found | Match the original --dir; URL reviews default to ~/.jury. Run jury runs --dir "". |
| Push rejected | Inspect branch protection and remote changes. Retained related-PR fixes can be resumed after recovery. |
| No reviewer selected | The judge is excluded. Enable or select at least one other reviewer. |
Use jury review (or a bare PR URL) for the loop and jury --web-only for saved-run viewing.
The old review-once and web commands were removed. Prefer --reviewer or --jury;
--agents is a hidden compatibility alias. jury agents still checks executables.
jury, codejury, and cr executable names remain available.
On a fresh setup with exactly one installed supported agent CLI, jury review uses
that agent for both judge and jury, keeping its separate reviewer and judge command
permissions. With exactly two installed CLIs, either can be randomly selected as
judge; the other reviews. Opt-in built-ins are included in this detection.
When one CLI serves both roles, a reviewer's rebuttal still returns the finding to the judge's triage queue, subject to the usual turn limit. A judge's own explanation does not reopen the finding just because the names are identical.
The selected roles appear before the run starts and are saved with it. jury reply
and jury review --resume <slug> reuse those assignments even if installed CLIs
change. An unavailable saved agent causes an error rather than a fresh assignment.
Explicit --judge, --reviewer/--jury, a saved global judge, or repository agent
role/enabled settings take precedence over automatic selection for new runs.
Explicit CLI role overrides on resume use the normal selection rules. You can also
select the same agent directly, even when other CLIs are installed:
jury review <pr-url> --reviewer codex --judge codex --push=falseNo additional flag is needed. Reviewer and judge use separate prompts and session
IDs; a fresh review never resumes the judge's history. Codex reviewer turns and
replies are read-only, while judge turns use workspace-write, including on resume.
The roles are saved for reply and review --resume. Same-agent sessions do not
provide model diversity or guarantee independent judgment. Default reviewer pools
still exclude the judge. Zero or more than two installed CLIs retain the existing
configured/default selection behavior when roles are not explicitly selected.
Pass a bare branch name such as --trunk master, not origin/master or a full
Git ref. Its refs/remotes/origin/<branch> tracking ref must be available and
share history with HEAD; fetch the base branch before reviewing if needed.
Jury validates this before launching agents. Review prompts use Git’s triple-dot
diff to compare HEAD to the merge base, and fail if that base becomes unavailable.
Choose the language of agent-written findings, summaries, judge reasons, replies
and code-fix explanations with a language tag such as en or zh-CN:
jury review https://github.com/owner/repo/pull/123 --language zh-CN
jury review https://github.com/owner/repo/pull/123 --resume <slug> --language en
jury reply --dir <worktree> --run <slug> --language zh-CNThe preference is saved with the run and passed to every reviewer and judge.
Subsequent rounds, replies and resumed runs reuse it when the flag is omitted.
Pass the flag again to switch languages, including when answering comments written
in another language. reply --dry-run --language en previews that language without
saving the change. A new run with no language option keeps the existing behavior.
Tags are validated and canonicalized (zh-cn becomes zh-CN). The preference
controls generated prose; programming syntax, identifiers, existing code conventions,
JSON keys, verdict values and exact protocol/stop markers remain unchanged.
Code comments, string literals and test names follow the repository's conventions
and task requirements rather than the response-language preference.
Replies with explicit DISPUTE: <finding-id> markers reopen only the named
findings. Agreements with other findings remain settled. Each complete rebuttal
ends at the next live dispute or separate new finding; restatements of the disputed
finding and quoted examples stay in its evidence.
The CLI and saved-run console show input, output and total tokens for every review, judge/fix invocation and reply, with subtotals per round, per model and for the run. A judge/fix turn is counted even when it rejects a finding without editing code. The expanded invocation table includes failures and retries. Resuming adds new invocations to the saved run rather than replacing earlier usage. Codex and Qwen report cumulative session usage; each resumed invocation counts only the increase over that session's last saved totals, including across CLI restarts. Qwen model counters are compared with the saved totals for each model. If a resumed thread has no saved baseline, that invocation's usage is unavailable; its reported totals establish a baseline for subsequent invocations. An interrupted turn without usage invalidates its session's baseline, so its unmeasured work is not charged to a later invocation. Commands that resume the last session follow the same rules. If an interrupted call never identifies the session it used, saved baselines for that agent remain uncertain until refreshed. Empty Qwen error results cannot restore an invalidated baseline, including across repeated failures. When resuming an older run, known reviews, judge turns and replies without usage snapshots remain unavailable and keep combined totals partial. Saved judge session records establish known old fix invocations. Bare verdicts and heartbeats cannot establish additional calls: manual resolutions use the judge name too, and heartbeats lack invocation IDs and may arrive late. Historical invocation counts are therefore limited to what those old logs can establish.
Counts come from agent output, never estimates. Codex JSON, Droid/Qwen result JSON
and OpenCode JSON expose supported usage fields. Other output formats, including
the current plain-text Claude integration, show unavailable. A configured model
is labeled configured; an identity reported by the provider is labeled reported.
If neither is available, the model is unknown. These are agent CLI identities; a
provider or billing account is not inferred from the model name.
Input includes cached input once. Cache reads and writes are also shown separately: Codex and Qwen already include them in input, while Droid and OpenCode report them separately. OpenCode reasoning is included in output. Qwen's reported total is preserved; thinking counted beyond input and candidate output is included in output. When Qwen reports model counters, they supply the model breakdown and replace its aggregate usage rather than being added to it. Cache counts that are absent remain unavailable. Partial totals add only available counts; missing invocations, missing input/output fields and unfinished or failed invocations mark them partial. Provider output can omit work from failed API requests or internal retries, so these counts are not a billing record. Plain text that mentions tokens is not usage data.
Usage snapshots are saved to events.ndjson as output arrives and folded into
run.json. The saved-run reader also recovers newer usage snapshots after an
interruption. If a provider emits usage only at the end, termination before that
event leaves it unavailable. A forced kill can preserve only bytes already received.
Protocol references: Codex events, Droid result JSON, OpenCode token normalization.
Each reviewer, judge and reply invocation gets an owned directory under
~/.jury/tmp/agents. Jury points TMPDIR, TMP and TEMP there and supplies
GOCACHE and GOMODCACHE defaults inside it. Tools using those variables or their
normal temporary-directory API stay within that invocation's storage. File-based
prompts use the same directory. Jury removes it after success, failure or timeout,
once the agent's process group has stopped. On Windows, a reviewer's exit does not
confirm that its descendants stopped, so launched invocations retain their caches.
Read-only Go module directories are also removed when cleanup is safe.
An explicit GOCACHE or GOMODCACHE in the parent environment or agent env
configuration takes precedence and remains your responsibility. Set these when
persistent Go caches are preferred over a fresh cache for each invocation.
JURY_AGENT_TEMP_ROOT selects another managed root; an agent-specific value is
supported, and the same value must be used when inspecting it from the CLI.
Jury overrides inherited temporary-directory variables for each invocation.
If temporary-cache initialization fails (for example, the root is unwritable),
Jury records an error for that agent and lets other reviewers finish. Their reports
and the failed round status are saved before the review stops.
Cleanup keeps ownership metadata until cache content is deleted, so interrupted deletions can be retried. Interrupted invocations are reclaimed when the next agent starts, or explicitly:
jury cache # list directories and ownership status
jury cache clean # reclaim known stopped invocationsCleanup preserves live owners and process groups, unknown directories, symlinks,
and launches whose process identity could not be recorded. Windows invocations
remain uncertain after their owner exits, including after timeout termination;
startup cleanup and jury cache clean preserve them. An uncertain entry requires
manual inspection before removal. Processes that deliberately detach from their
agent's process group are outside this lifecycle. Explicit external caches and arbitrary
/tmp or /private/tmp directories are never claimed or removed. This does not
retroactively identify old, unmarked caches created before managed storage.