Your issues become pull requests overnight.

Start passes with thirdshift architect base main, thirdshift pickup base main or thirdshift secure base main. A Custom process started with no issue, and neither an Issue-branch nor a Spec-branch worktree, follows the same Command limit, running beside other Passes up to it, and does not wait on a Ready issue or an Architect idea. A file that declares either branch worktree is refused when no issue is given, before any session. Your operating system schedules them. [pass] command_limit, 1 by default, is how many Commands on one repository on the machine skip a Pass there; raise it to let Passes overlap: Pickup runs, one Architect run and one Security run side by side, never two Architect runs or two Security runs. A Ticket run you start on an Issue URL counts toward it, but is never refused. thirdshift status, thirdshift cronjobs and the Console show when each repository's Pickup, Architect and Security Pass last fired, including repeated skips. Architect runs every few minutes and Security runs on a schedule, with fixing allowed. Both yield to work the Day shift shaped, and pause when a decision needs you.

See the Console with sample Commands in the Focus light Theme. Open either screenshot for the full-size view. The Console header shows the running thirdshift version. Every Console tab title names its host; the Inbox also shows its Needs you count. Ready for review rows show the pull request's latest CI: green, red, pending, no checks or unknown. A separate Merge conflict pill appears beside CI when GitHub reports a conflict; the row stays Ready for review even with green CI. status shows the same indicators, and show --json includes locally cached GitHub facts and observation times without a network request.

The Console opens on this machine's Inbox: Commands grouped by what needs you, with their stages and logs. Overview shows Needs you, Ready for review, Landing and Working together, each with its heading and count. Folders beside the one-line rows filter to a single section, including Done and Resolved; a clicked row's Command details open alongside. It reads ready rows' pull requests and Needs you rows' issues and pull requests from GitHub in the background once a minute; changed answers update open pages automatically. thirdshift status uses the same kept answer. Failed reads keep the last facts; after five minutes, both show when GitHub last answered and why the latest read failed. With fewer than 500 GraphQL points or REST requests left, both wait for the quota to reset, leaving the rest to the factory. An ended Needs you row leaves when GitHub shows its issue closed or its pull request merged; a Triage row also leaves when needs-triage is removed. A Finding leaves with “findings triaged” once all its private records are triaged on GitHub. A draft advisory without severity, or an open security-finding issue labelled needs-triage, still waits; a finding closed by its merged fix does not. Live Silent and Slow rows wait until their Commands end. A signed-in account can Stop or Requeue a live Command, just as from the CLI. Dismiss is always visible on the row in Needs you or Ready for review, beside Copy where available; d dismisses the selected row and selects the next. An inline Dismissal shows a temporary notice with Undo; Command details also keeps Dismiss and Undo. Command details shows each CLI form: thirdshift dismiss <command> or thirdshift dismiss <command> --undo. A dismissed row returns for a new kind of problem or when it ends ready for review; each kind counts once for the whole row.

The Console and status show the serving OS account's Commands by default. thirdshift console other-accounts readable, or a comma-separated list of names, includes readable local accounts with an account tag, including in Needs you and its tab count. Other-account Commands are read-only: Stop, Requeue, Dismiss and Undo apply only to the serving account's own Commands. The machine owner grants read access with filesystem permissions or ACLs. Account visibility.

On a desktop, Inbox issue numbers stay visible when long repository or Process subjects shrink. Hover a subject for its full text.

Command details opens with the Inbox subject linked to the issue and its full title, wrapping long titles above Stop and Dismiss. An unknown title says “(title unavailable)”; a Command without an issue shows its Inbox subject alone. The recorded issue title is also in thirdshift show <command> --json.

An ended, unsuccessful Command shows its full failure reason immediately below that heading, before the controls, with URLs as links. When its own cause is absent, it uses the first cause from failed dispatched work. Successful or live Commands, and those with no cause, show no reason. The cause also stays in the facts list.

Command details shows the last five lines of recent Command output above Full Command log, updating as output arrives. The small preview helps assess a Slow run without leaving its details.

A Security fix’s Command details links each advisory it addresses, with its title and severity, even if the fix fails or is interrupted. The parent Security run marks which advisories it dispatched a fix for. Older Command records show no advisory line.

Above its groups, the Inbox counts the runs active on this machine. Below the notification control, every Console reader page shows whether host limits allow another session, refreshing every ten seconds while visible. The indicator links to the Resources page, also available with thirdshift resources, for host memory, CPU load and the User config limits that can delay new sessions. It says when limits are off or measurements are unavailable; capacity is not reserved.

Experiments compare a Candidate with a Baseline on private Eval suites. thirdshift experiment <candidate> shows its Trials’ worst-case session-minutes, with an estimate where this machine has history, and asks on a terminal unless given --yes, then builds both Arms and runs interleaved Trials with two workers, a fixed Bar and resume after interruption or a Paused Command outcome at a quota stop (85% for Claude and Codex, a Claude rejection, or a Claude warning with no readable utilization). Trials and Graders run in bwrap sandboxes. Code Graders score mechanical checks offline from report packets and the prepared repository, then fresh model Graders score written criteria from blind packets with provider-only access. Graders cannot read raw Trial logs, Arm mappings or the Experiment’s other files. Results show the Verdict, wrong findings, borderline marks, paired intervals, each Arm's minutes, tokens by Model, reported cost and Void causes. Trials and Graders join the Command's Trace. Void Trials retry fresh up to twice; an exhausted task drops out of both Arms. More than a tenth of judged tasks dropped in suites that are not advisory makes the overall Verdict inconclusive.

Experiments, including release Experiments, report approved Eval tasks missing Trial inputs before building either Arm. Before each Trial's session, the host prepares its repository's pinned Rust toolchain and listed components, sharing preparation across Trials with the same pin. An installation failure makes the Trial void with session_unavailable. Spec review Trials keep the production prompt and grade only the reports; their sandbox blocks gh calls and provides no PR service.

The Bar fixes separate pass and fail levels, 70% and 95% by default. Each metric shows both intervals from the same draws: fail below zero at the fail level; otherwise pass when the interval at the pass level sits above the margin, marked worse, within the margin if it lies wholly below zero. Today's suites fall short of an 80% A/A pass rate at the 70% default: code-review needs about 20 Eval tasks, Spec review about 80. --pass-level and --fail-level set them before the first Trial. Saved Experiments keep their original Bar and Verdict.

An A/A Experiment uses one commit and matching Harness, Model and Effort flags in both Arms to measure ordinary Trial variation. Select its suites explicitly with --suite; both Arms run through to a Verdict. After completion, thirdshift experiments <id> --record-difficulty records each Eval task's Difficulty from both Arms' non-void Trials. thirdshift evals lists the readings and whether each is discriminating: between 0.15 and 0.95 inclusive, over at least 6 Trials. Fewer Trials or no reading means unmeasured.

The Bar fixes each suite's discriminating Eval tasks from recorded Difficulty before the first Trial. Only those tasks are judged; other approved tasks run and are graded as measurement. experiments <id> and --json show measurement Trials per task and counts of judged, measurement-only, void-dropped and judged tasks measured on another setup. Measurement changes no interval, required recheck or prediction. An empty judged list is inconclusive. A changed list or changed Difficulty for a retained task refuses resume after the first Trial, including readings that stay inside the band. Saved Experiments keep their original method.

The Console's Experiments pages group saved Experiments into Running, Needs you and Done. Each Experiment says what it tests and updates live with its Trial grid, Verdict intervals, Arm comparison and Bar, each folded to one line until opened. Open a Trial's Peek for its Scores, reasons and Session evidence; copy recheck commands to run in the CLI. Session links share thirdshift experiments <id> --session <session-id>'s confined evidence reader, keeping healthy evidence readable even when unrelated Trial metadata or usage spans are corrupt. Approve proposed Eval tasks there or with thirdshift evals approve <suite> <task> --json. Approval starts no work and affects future Experiments; unapproved proposals never run.

A suite with fewer than 20 discriminating Eval tasks is an Advisory suite: its Verdict and intervals are shown, but cannot pass or block a release. --min-discriminating-tasks sets the minimum for one Experiment; the Bar fixes it before the first Trial and refuses a changed minimum on resume. Only suites that are not advisory decide the overall Verdict, Void dropout guard and required rechecks. When every suite is advisory, the Verdict is inconclusive. CLI text, JSON and the Console mark Advisory suites and show judged and measurement counts. The release summary lists each Advisory suite and its Verdict and asks whether to release when all suites are advisory. Saved Experiments without the minimum keep their original Bar and Verdict.

Each Trial and each Grader has its own memory cap. A Trial its cap kills is void; repeated cap kills are reported for its Arm. The User config sets the caps, with defaults of 1 GiB for throttling, 2 GiB for killing and 1 GiB of swap.

The Verdict waits for your rechecks of borderline and decisive Scores. thirdshift experiments <id> --json lists those waiting and which flip changes the provisional Verdict, with its resulting Verdict. thirdshift recheck <id> <trial> pass --reason '…' records your decision beside the Grader's Score and keeps its history. To review a batch, use thirdshift recheck <id> --pending, then record one decision and reason with thirdshift recheck <id> --batch <fingerprint> pass|fail --reason '…'. A changed pending list refuses the batch.

Each Trial removes its Harness caches when its session ends. After grading and all rechecks finish, the Experiment removes every attempt's repository copy. Paused Experiments keep repositories for resume; Session evidence, reports, packets and Scores remain readable.

scripts/release.sh runs on Linux. It refuses other hosts before Git or GitHub work. On macOS, use an Ubuntu 24.04 Linux VM such as Lima; on Windows, use WSL2.

thirdshift experiment --release <commit> compares touched suites with the last release tag. The release script skips Experiments by default. Add --with-experiment to scripts/release.sh on each invocation to run one: it waits for rechecks before pushing and asks whether to release on a fail or inconclusive Verdict. Reruns with the flag reuse the Experiment, including before opening a pushed branch’s PR. It selects the newest main commit whose latest push CI passed, tags that content plus the version bump, waits for publication, then merges the bump PR. After a failed unpublished release, rerunning re-cuts from newer green main content or automatically reruns failed jobs; published releases are never moved. If the publication check cannot be completed, the script refuses and moves nothing. Recovery retains the prior release boundary for the summary and Experiment. Main can move during the release; a failed merge leaves the release standing and the PR for the Day shift. No touched suite or a last release with no Trial flag carries on with a note. Release checks.

Completed Ticket runs keep their Run metrics in their Command records: outcome, Repairs, CI rounds, Unaddressed findings counted separately as Standards, Spec and Security, stage times against Usual ranges, and minutes by stage and in total. The Unaddressed findings counts come from Delivery's final PR body and include the Security review. Each Session stage keeps the review's Standards and Spec finding counts from its summary, with missing counts left missing. thirdshift show <Issue URL> --json includes each Ticket run's metrics, including Ticket runs and Base fixes. The Console's Command details show these same metrics and mark unavailable signals as Missing.

thirdshift metrics <command> [--json] shows one Ticket run and its children; thirdshift metrics [--json] shows weekly trends by repository, Harness, Harness version, Model, Effort and kind of session, with week and repository filters. Session traces supply tokens by observed Model with the cache split, reported cost, tool errors, permission denials, killed work and repeated tool calls. Completed Ticket runs sum these by session kind; past Ticket runs get them from converted traces when metrics are read. Missing values stay missing. Every signal stays separate.

Confirmed Bug origins count against the Ticket run that introduced each bug when metrics are read, even if confirmed after it ended. Metrics match this machine's Ticket runs against the Scoring data for both hosts, and weekly Ticket run trends count them in the Ticket run's start week. Rejected and unsettled origins count for nothing. Command details in show --json and the Console include these read-time counts and the available metrics of past Ticket runs, without rewriting their records.

Run metrics and Experiments share Session usage accounting: partial totals refine only reported fields, preserving known input, cache counts and cost. Input includes cache reads and writes; missing measurements stay missing. Reported Session cost wins, otherwise each Model's reported cost replaces its call costs. Saved completion metrics stay fixed.

Every Ticket run also checks its sessions' shell commands for force pushes, shared-branch history rewrites and recursive deletes outside the Worktree. Run metrics retain each matching command as an advisory flag; flags never block or fail a Ticket run.