Skip to content

Product

MOD-02 · Evidence and delivery

Behavior and limits for evidence and delivery, based on the reviewed product catalog.

FEAT-004 · Claims and verification

A claim needs a command executed against the actual HEAD. Pending results, timeouts and withdrawn evidence do not count as a pass.

In version 0.6.10, CHECK has two modes with the same coverage and the same verdict. Agile Check, the default, runs the preparation once, reuses its output under a hash guard and runs the test lanes and the scan in parallel, without exceeding the machine’s CPU; Classic Check keeps the same commands in the same order. check.mode: classic in orkastery.yaml returns the whole project to Classic, and ork verify --check classic|agile applies to that run only and overrides the manifest. An unavailable Agile piece falls back to the Classic equivalent, and verify_run records the mode, where it came from and what ran in its place.

With verify.security_scan in the manifest, ork verify runs ork security-scan as the fourth command, after build, test and typecheck. The scan compares HEAD with the base by fingerprint and classifies each finding as new, existing or dismissed; a new secret and a new high or critical vulnerability in dependency, SAST, IaC or workflow fail the run, while quality and license only report. CHECK triages each new finding with the security-auditor, and GO-FIX fixes it with the security-remediation skill or records a dismissal with ork decisao registrar. Without the key, nothing changes.

For pull request delivery, the CI documentation check also requires a CHANGELOG line when the change touches core, adapters or marketplaces, and the roadmap item state after the merge. Verify and commit through MCP no longer stall on a lease held by an already closed thread.

On a virtual machine whose CPU is stolen by the hypervisor, a local ork verify can fail on timeouts; the reliable proof is CI. ork verify measures steal from /proc/stat and records it in verify_run: when only clock-bound tests fail under steal above 40%, the failure becomes verify.timeout, which retry runs again, never verify.regression. The registry of clock-bound tests follows the ork.instabilidade/v1 contract, and JUnit marking protects only the declared cases.

Since version 0.6.2, when the runner names the failures, verify compares test by test within the same command: a new failure in a command that already failed blocks with verify.regression and carries the name in the reason. A new claim that uses rg, calls the CLI by name instead of node core/dist/index.js, reads .orkastery/ or carries literal escaped quotes is refused before it is recorded, with the fix pointing to grep and node core/dist/index.js. In a repository with CHANGELOG.md and a published v* tag, verify also refuses a new line in an already tagged version and requires the Não publicado entry under Adicionado, Alterado, Corrigido or Segurança.

In version 0.6.5, verify understands the JUnit report from any runner declared in orkastery.yaml, per command and per Agile lane; without the declaration, nothing changes. With it:

  • failures and errors are recorded by name, without the cap of ten, with counts and duration in verify_run; the exit code still judges, and the XML never turns a red command green;
  • with the filtro key, the fix round first reruns the names that failed and, if they pass, the full suite before the receipt can turn green;
  • under high steal, a failure made up only of cases marked in the relogio key is left without a verdict and gets a second isolated round through the same filter, which never turns it green;
  • on a warm Agile round, a lane with the selecao key and a valid cache uses the runner’s native selection for the mapped changes, and any doubt runs the full suite, with the reason;
  • a partial receipt, from a filtered pass or from selection, never counts for ork ship.

FEAT-005 · Independent CHECK in CI

The claim bundle is rerun against the candidate’s exact SHA. A run for another thread or commit does not prove this change. Each thread has its own bundle, .ork-ci/<thread>.json, which CI finds by branch name; a thread branch without a bundle fails, and main runs only the manifest commands.

With the opt-in dado_privado policy declared, ork verify scans the thread diff against the base and ork ci prepare scans the same diff and the bundle text before writing the file: a finding fails with policy.violation in block and records policy_warn in warn. The patterns cover home paths, email addresses outside the allowed list and machine-format hostnames, plus the terms of the local list kept outside git; the report states file, line, column and pattern, never the value. ork privado varrer runs the same scan over the whole versioned tree.

Since version 0.6.0, GitLab has parity with GitHub: GitLab CI as the independent CHECK, pulse, the network, delivery by MR and SHIP through MCP, through a forge port in which each platform is an adapter with capabilities.

FEAT-006 · Delivery with verified push

SHIP serializes the merge through a lease and checks the remote. Dry runs and local commits are not publication receipts.

The ship_done receipt records the commit that incorporated the thread branch (mergeSha) separately from the base tip proven on the remote (pontaDaBase and shaRemoto).

With verify.security_scan declared and without ci.required_for_ship, SHIP stops at human.pending before any push when a loop category (secret, dependency, sast, iac or workflow) was left unscanned at the candidate SHA; only --allow-unscanned "<who>" delivers, with ship_unscanned_authorized in the ledger.

In MCP, ork_ship chooses its transport in ork mcp install --ship-transport: github-ssh, forja-ssh, forja-https or bare-local. forja-ssh checks the server key against the person’s known_hosts or --ssh-fingerprint and uses StrictHostKeyChecking=yes, without the user’s SSH config; forja-https uses only the gh or glab auth git-credential with a login on the host, with no prompt and no token in ork. The pinning lives outside the project, in $XDG_CONFIG_HOME/orkastery/mcp-ship/, and is rechecked at every step; an origin or config change during the push leaves the delivery unproven, with mcp.ship.transport.mudou-no-push.

FEAT-007 · Worktrees and leases

Each thread edits its own worktree. Tree, path, card and service leases coordinate resources; auditing rejects disagreement between Git and recorded state.

The worktree is created with the thread when the manifest has worktree.por_thread: true, the ork init default, or with --worktree auto; later, ork worktree ensure creates it. ork worktree sync refuses with tree.blocked, cause base-reescrita, when the base was rewritten, and shows the git rebase --onto that replays only the thread’s commits.

Leases and the queue live in .orkastery/leases at the project root, whether ork is called from the root or from any worktree. A legacy worktree lease is only considered within a 30-minute window, is never deleted and never proves ownership; ork lease list shows leases, families and queues.