Skip to content

Reference

CLI reference

Find commands by task and check arguments in your installed version.

Inspect state

Start with diagnostics and the overview. Thread, phase and claim queries explain recorded state; they do not grant write permission. Use --help on a specific command before performing a mutation.

ork doctor fails when a file or folder in .git has a different owner from the repository owner, with the count, examples and the exact chown in the fix; it runs nothing itself. The graph analyzers check loads typescript and micromark as ork grafo indexar does and warns with the fix when they are missing. The rede line shows Orkastery Network membership, the home, the last heartbeat and the last failure, reading local files only. With the heartbeat line in crontab, a last run older than 30 min also fails: the heartbeat stopped. The forja section checks the forge of the origin remote, and only appears with a forge remote or with the forja: block in the manifest. With verify.security_scan declared, the security scan check is ok with Docker, or with the native binary of every tool in the loop, and warns, never fails, when one would be unavailable.

In version 0.6.10, ork doctor --json returns the same checks in the ork.doctor/v1 contract, with verdict, failures, warnings, project, machine and the factory heartbeat: open threads, crontab line, last run, its age, atrasada and the notice ready for the owner. The forja section only runs with --forja; without the flag, the JSON carries forja: nao-medida.

ork doctor
ork maestro --json
ork thread status <thread>
ork phase list <thread>
ork claims list <thread>

Commands by responsibility

Conduction covers ork thread, ork phase, ork modos and ork gate. Evidence uses ork claims, ork verify, ork fix and ork retry, and scanning uses ork security-scan and ork privado varrer. Isolation uses ork worktree, ork lease and ork conducao, which shows who is conducting each thread. Continuity uses ork handoff, ork recall and ork memory, and written knowledge uses ork conhecimento. Delivery uses ork ship and ork master. Auditing uses ork audit, with the debt board in ork audit divida. Since version 0.6.0, ork verify includes the Agile Check, and security-scan is a command of its own.

ork --help
ork ship --help
ork master --help

Thread and worktree

With worktree.por_thread: true, which ork init writes, ork thread new creates the worktree and the ork/<slug> branch without a flag; with the key false or absent, only with --worktree auto. --worktree DIR reuses an existing directory, --sem-worktree creates no worktree and warns that on the base branch ork ship will be blocked, and --dry-run shows the worktree that would be created. The same rule applies to --from-finding. In version 0.6.10, with worktree.dependencias in the manifest, the command runs inside the new worktree after the thread is created, and a failure goes to stderr with the exact command, without undoing the thread.

ork worktree sync recreates a branch with no commits of its own at the base SHA. With its own commits and a rewritten base, for example after a force-push, it refuses with tree.blocked, cause base-reescrita, and shows the git rebase --onto that replays only the thread’s commits.

In version 0.6.10, with --pedido, ork thread new stores the request in pedido.md, in the thread folder, with the sha256 in thread.json and in thread_created; an empty request or one over 64 KiB is refused before anything is written. Without --pedido, ork prompt render uses the stored request, checked against the sha256 (thread.pedido.divergente when the file was edited outside the core); a thread with no stored request keeps the sample text.

ork thread new "<nome>" --modo auto --dry-run
ork thread new "<nome>" --modo auto --sem-worktree
ork thread new "<nome>" --modo auto --pedido "<text>"
ork prompt render <thread> --fase GO
ork worktree sync <thread> --dry-run

Who conducts a thread

ork conducao status <thread> shows who is conducting the thread right now (the channel, the session or process, the phase and since when) and releases, with proof, a conduction left without an owner. ork conducao assumir <thread> performs the handoff: it ends the current conduction through the runtime, records who took over, from which channel and why, and reserves the turn for that channel; --por and --motivo are required.

In version 0.6.10, ork conducao also has the three commands of the Conductor’s Sentinel, in the block below. The first shows whether the Sentinel is active on this machine, with the session, the last heartbeat and the deadline, and gives the conductor’s next step for each open thread, computed by the core; with --json, in the ork.conducao-sentinela/v1 contract, it is what the external conductor reads to yield dispatch. The second is the heartbeat: it renews the session’s presence for 3 minutes, or takes it over when no other one is alive; with another live session on the machine, it exits with code 3 and does not take its place. The third removes the session’s own presence, and the external conductor takes over again. The ork-sentinela mod runs the heartbeat and the switch-off by itself.

The same version brings the owner’s durable authorization, which lets the heartbeat act on its own within the deadline. ork conducao autorizar --ate D --pode limpar,registrar,master,merge,despachar writes the ork.autorizacao-duravel/v1 contract to .orkastery/monitor/autorizacao.json, outside git, for up to 31 days; without --por, the user.name configured in git is the value that counts when it names a person. ork conducao autorizacao reads the authorization in force, or the typed reason why it does not hold. ork conducao batida runs the cron’s deterministic step: the merge of an ork/* PR with the required checks green, delivery registration, MASTER by default, ghost cleanup and dispatch of the next phase, only within the authorization, and exits 3 when another heartbeat is acting. ork conducao atualizar <thread> merges the remote base into the thread branch, and --enviar pushes without force and proves it with git ls-remote.

ork conducao status <thread> --json
ork conducao assumir <thread> --por <who> --motivo "<why>"
ork conducao sentinela status --json
ork conducao sentinela bater --sessao <id>
ork conducao sentinela desligar --sessao <id>
ork conducao autorizar --ate 2026-11-02 --pode limpar,registrar,master --por "<who>"
ork conducao autorizacao --json
ork conducao batida --json
ork conducao atualizar <thread>

Integrations and operations

ork adapter and ork mcp connect the host. ork accounts and ork setup describe runtimes, profiles and what this machine has confirmed. ork onboarding keeps the project’s interview and answers, and ork experiencia shows the preferences of the experience pack. ork sessions, ork monitor and ork pulse observe execution and human attention. ork brain and ork portfolio access memory and the catalog, and ork ledger stats gives the ledger telemetry. ork fabrica, ork roadmap and ork network coordinate work across machines, and ork projetos shows which project each command reads. ork grafo queries the local code graph index, and ork conhecimento queries the project’s written knowledge. The canonical repository reference details each family’s options, and a test checks its invocations against the CLI help.

In version 0.6.10, ork mcp install accepts --ship-transport github-ssh, forja-ssh, forja-https or bare-local, and that is the only push an agent fires. Without the option, github-ssh applies. forja-ssh accepts the forge host and port, reads the server key and only pins it once confirmed by your known_hosts or by --ssh-fingerprint, and checks the key on every push with StrictHostKeyChecking=yes. forja-https uses the auth git-credential of gh or glab, with a login on the exact host, clears the other helpers in the config and refuses plain HTTP or credentials in the URL. The pin lives in $XDG_CONFIG_HOME/orkastery/mcp-ship/, is per machine and requires the same install there; changing mode means removing the orkastery entry and reinstalling.

ork mcp install --project "$PWD" --host claude-code --ship-transport forja-https
ork mcp install --project "$PWD" --host claude-code --ship-transport forja-ssh --ssh-fingerprint SHA256:<fingerprint>

Confirmations on this machine

In version 0.6.10, three manifest choices only take effect after the machine confirms them, outside git. A runtime.sandbox that gives the agent full access (danger-full-access) only dispatches through Codex and Grok Build after ork setup sandbox confirmar <postura>; without it, ork phase run and ork retry run refuse with runtime.sandbox-nao-confirmado. A worktree.dir that leads outside the root, by its real path, only creates a worktree after ork setup worktree confirmar; without it, creation refuses with worktree.dir-fora-da-raiz. And fabrica.compartilhada: true only turns the factory on with the machine’s membership, through ork fabrica entrar.

The two ork setup confirmations live in .orkastery/private (postura-local.json and worktree-local.json, mode 0600, outside git) and apply only to this checkout; the worktree one, only to that folder. The factory membership lives in ~/.orkastery/maquina.json and applies to the machine. ork doctor reports each one that is missing.

Also in version 0.6.10, ork eval only runs the catalog of the package that is running: a catalog found from the cwd that is not the package’s own refuses with eval.catalogo-alheio. In an Orkastery checkout, use that checkout’s own, node core/dist/index.js eval, as CI already does.

ork setup sandbox --json
ork setup sandbox confirmar danger-full-access
ork setup worktree confirmar
ork setup worktree revogar
node core/dist/index.js eval

Accounts, profiles and sessions

ork phase run --perfil <id> dispatches through the requested account, or refuses with runtime.profile-invalid, runtime.quota-exhausted or runtime.auth-missing, without switching profiles on its own. runtime_profiles.distribuir chooses among profiles: ordem, the default, is the first in the store; carga is the one with the fewest live sessions on this machine. An unknown value falls back to ordem, with a warning.

ork sessions lists the sessions of every account, with the PERFIL column and the ghosts; --global also inventories the account’s Claude and Codex sessions outside the project root, and --exigir-limpo exits non-zero with an invalid source or a session without a thread. ork sessions limpar-fantasmas releases each ghost from ork with a ledger record, without calling stop or rm in the runtime. When the runtime refuses the worktree directory or waits for new terms to be accepted, the phase waits for the owner to run the indicated command, and ork retry run then re-dispatches the same prompt.

In version 0.6.10, for ork retry run, directory trust counts when the account’s .claude.json records it for the directory itself, for one above it up to the repository root, or for the worktree’s main repository, and a re-dispatch refused again for the same blocker neither spends an attempt nor raises the effort.

ork accounts esgotamentos measures, without writing anything, the dispatches that landed on an account another project had marked as exhausted while the mark was still valid; the account appears as an opaque id, and the total is a floor.

ork phase run <thread> GO --prompt "<texto>" --perfil <id>
ork sessions --all
ork sessions --global --json
ork sessions limpar-fantasmas --dry-run --json
ork accounts esgotamentos --desde 7d --json

Grok Build by explicit dispatch

In version 0.6.10, ork phase run <thread> GO --runtime grok --model <model> dispatches the phase through Grok Build without an interface, requested explicitly or through the block setup, and never through the project’s runtime.adapter. There is no default model: the model is always the owner’s, and the effort goes to Grok Build under the same name. The child process runs without XAI_API_KEY, using the subscription the profile’s login stored, in the workspace sandbox by default: read-only with runtime.sandbox: read-only and in the PLAN that closes the block, and with no sandbox only when danger-full-access is confirmed on this machine through ork setup sandbox confirmar <postura>.

The command ork setup <mode> --bloco N --runtime grok --model <model> makes Grok Build a block’s runtime, and the block fallback accepts Grok Build with its model. ork mcp install --project <root> --host grok writes the MCP entry to the project’s .grok/config.toml, which is not versioned; at dispatch, the core’s MCP server starts outside the Grok Build sandbox, on 127.0.0.1 only, with a per-dispatch token and only that session’s tools, and dies with the session.

Grok Build documents no command that checks its login, so ork accounts check keeps its profile’s state, and a login error at dispatch takes the profile out of rotation with runtime.auth-missing. ork sessions also lists the Grok Build sessions of each profile. The live trial that validates the runtime is still missing.

Citable context and human attention

ork brain context returns entities, parents, citations, freshness and gaps without granting write access, and ork brain dossie builds a read-only decision dossier: link, citable context, alternatives, who decided and evidence. ork roadmap status builds a read-only report showing where the owner’s attention is needed. ork master pedir creates a score request for an authenticated human response; it is not an agent-supplied score. ork master <thread> --score requires --por with the name of the person giving the score. ork master <thread> --aceitar-omissao accepts by default only the given thread, with --dry-run to preview; without a thread, it accepts every delivered thread in the project.

Without --classe, the failure class comes from the reason of each failed gate, and base-avancou is added when the branch pulled in the base before delivery; this also applies to --aceitar-omissao.

In version 0.6.10, when a thread outside the roadmap is waiting for the owner, the closing of ork roadmap status gains a line that counts those threads and points to ork pulse, and --json gains the optional foraDoRoadmap field; with no such thread, the output does not change.

ork brain context --thread <thread> --ids <ids>
ork brain dossie --thread <thread>
ork roadmap status --json
ork master pedir <thread> --formato telegram
ork master <thread> --aceitar-omissao --dry-run

Choose the project being read

Every command reads one project, chosen in this order: --projeto <name|path>, anywhere in the command; the ORK_PROJETO variable; in a host without a project directory, the only known project or the projeto.escolha refusal; and, in the terminal, the current directory. An ambiguous or unknown name is refused with the candidates, exit code 4; ork never chooses on behalf of whoever asked.

ork projetos lists the projects known to this machine, without secrets. ork init, ork thread new and ork fabrica entrar register automatically; ork projetos registrar and ork projetos esquecer handle older copies without deleting anything from disk. ork maestro, ork board, ork board plan, ork fabrica and ork roadmap status begin by stating the project read and what was not read.

In version 0.6.10, ork roadmap reservas also starts with the project read, and the other commands that read a project the reader cannot guess (in a host, or outside its directory) end with the Projeto consultado line on stderr.

ork projetos --json
ork roadmap status --projeto meu-produto
ork roadmap reservas --json
ork projetos registrar
ork projetos esquecer <name>

Network, reservations and feature numbers

ork network roadmap combines the network roadmap from any directory, with typed gaps for anything it did not read. ork roadmap reservas flags orphaned reservations held by threads already closed on this machine, and --soltar-orfas releases them or passes them to another open thread on the same item, with a record; --remoto picks the remote that is read, origin by default. ork roadmap feat reserves the next FEAT number through an atomic push; two machines never get the same number, and a reserved number is never reused.

ork network status shows the machines in the person’s Orkastery Network, from any directory, with the source, typed gaps and what was not read. ork network entrar checks or creates the private home on the forge and publishes the first snapshot before recording membership; ork network publicar writes the snapshot right away, and ork network sair removes it from the home.

ork docs sincronizar brings ledger and git facts into roadmap items. --so limits it to the requested items; in the worktree of a thread linked to an item, that item is the default, and --todos covers every item again. ork docs verificar checks product and roadmap documentation against the code and git, and exits non-zero on error; on a PR, with --pr, another thread’s merge and an index that diverge from main only warn. ork docs init creates standards, templates, indexes and the Markdown lint in the project.

ork network roadmap --json
ork network status --json
ork network entrar --maquina <nome>
ork roadmap reservas --soltar-orfas
ork roadmap feat --thread <thread>
ork docs sincronizar --escrever --so RM-NNN
ork docs verificar --pr

Company network and collective reservation

ork network empresa criar creates or rereads the network’s private repository and stores the policy, with --nivel observar, avisar or exigir. ork network empresa inscrever records the pending enrollment and the maestro’s expected identity; ork network empresa aceitar, run by the authenticated maestro, checks forge, host, accountId, network and projects before activating.

ork network empresa publicar checks the account again, rebuilds the closed projection, applies dado_privado and the secret scan and writes the v2 and v3 contracts in the same transaction, without force. ork network empresa status shows the state, the last projection, anything pending and the collective preflight; ork network empresa sair and ork network empresa revogar stop new submissions.

With an active enrollment, ork roadmap pegar and ork thread new --roadmap reserve on the collective network first and show coverage that is missing, not checked or pending; exigir without a check refuses. --forcar only handles the local guard and never takes another identity’s collective reservation. Without an enrollment, the output stays as before.

--coordenacao reserva, colaborador or acordo chooses how to join the item. An agreement needs the --atestacao-forja, --atestacao-id, --atestacao-endpoint and --atestacao-sha256 options, and the auditable exception needs an owner, a justification, an expiry and the --excecao-* reference. Over MCP, ork_thread_new takes the same data in the coordenacao object.

ork roadmap reconciliar resumes an interrupted saga and undoes only the effects that still belong to the operation. ork ship, ork ship publicar and ork ship pista show the advisory collective preflight, without changing delivery readiness.

ork network empresa criar --forja github --dono <org> --repositorio <network> --nivel avisar
ork network empresa aceitar --empresa github:<org>/<network> --inscricao <id> --rede github:<org>/<network> --projetos github:<org>/<project>
ork network empresa status --inscricao <id> --json
ork roadmap pegar RM-NNN --coordenacao colaborador
ork roadmap reconciliar RM-NNN --operacao op-<id>

Check claims in the sandbox

In version 0.6.10, ork claims conferir <thread> --claims C3,C7 re-runs only those claims in the ork_verify sandbox, in the thread’s worktree registered in git, under the thread’s conduction and without writing a receipt: no verify_run and no claim state. ork claims conferir <thread> --tarefa T2.1 checks instead the verify of a task in the thread’s docs/plan.md, when it is a single command in backticks; --prazo goes from 10 to 600 seconds, 180 by default, and --json returns ork.claims-conferencia/v1.

It exits 1 when any claim fails, 2 without a sandbox or with an invalid request, and 3 when another session is conducting. The real check is still ork verify, which does not run confined: in the sandbox, a node child process fails, so a test claim that spawns a process can fail there and pass in ork verify.

ork claims conferir <thread> --claims C3,C7
ork claims conferir <thread> --tarefa T2.1 --json
ork verify <thread>

Withdraw claims and fix from the review

ork claims retirar gains --motivo-tipo, with three values: substituicao, obsoleta or erro. With --substituida-por <claim-id>, the withdrawal only goes through if the replacement exists, is active and passed the latest verify with the same command; otherwise it refuses and asks for a new verify.

Without a replacement, the claim stays active, and the core opens the typed claims.retirada-sem-substituta request to the owner. Without --motivo-tipo, the earlier free-form path still applies, and the receipt marks the withdrawal as untyped. The history stays in claims.jsonl.

ork claims add accepts --dependencia-externa rede,relogio for a claim whose command depends on the network or the clock. A GO without a commit that changes neither HEAD, the active claims nor the plan repeats neither the verify nor the review; a claim with an external dependency still asks for a new verify.

ork fix open reads the verdict and the structured findings from the thread’s docs/check.md: a single verdict line, in Portuguese or English (PASSED, FAILED, NEEDS CHANGES or BLOCKED), and one finding per line, with id, severity, file and line, as at the end of the block below. A negative review without a structured finding, with a green verify, refuses with fix.check.findings.

A finding without a claim joins the same fix round and only closes with a new independent PASSED review and a full green verify. ork fix status shows the round’s findings, with file and line, the commit that closed each one and what remains.

Command texts and reasons come out redacted. verify_run, the receipts and the verify executions redact credentials without changing what runs or the identity of the proof. A withdrawal through the CLI shows the redacted reason, and ork_claim_withdraw, over MCP, performs the same typed withdrawal and returns only the fingerprint of the claim’s commands, never their text.

ork claims add <thread> <file> --claim "<claim>" --verificar "<command>" --dependencia-externa rede
ork claims retirar <thread> C3 --motivo-tipo substituicao --substituida-por C7 --motivo "<reason>"
ork claims retirar <thread> C4 --motivo-tipo obsoleta --motivo "<reason>"
ork fix open <thread>
ork fix status <thread> --json

# docs/check.md
Verdict: NEEDS CHANGES
- A1 [blocker] src/filtro.ts:42 - the filter ignores the time zone

Claims that CI runs as written

ork claims add accepts a repeated --verificar and stores every command, in the order given. ork claims verificar does the same with a repeated --comando.

A single-value option repeated with the same value counts as one. With different values, the command exits with code 2 and names the option; with --json, it returns the cli.option.repeated error.

The lint reads the base branch from the manifest. A proof that cites it as a local ref, main instead of origin/main, does not run in the CI checkout, which only has the remote branch. A remote ref, search text, a file name or a similar-looking option does not trigger the refusal.

The CHECK review lives only in the check artifact, the docs/check.md in the thread’s state, written through ork_artifact_write. An untracked copy of docs/check.md, docs/goal.md or docs/plan.md in the worktree stays there, but becomes a typed warning in the ledger and shows up in ork thread status.

ork claims add <thread> <file> --claim "<claim>" --verificar "<command 1>" --verificar "<command 2>"
ork claims verificar <thread> C3 --comando "<command 1>" --comando "<command 2>"
ork claims add <thread> <file> --claim "<claim>" --verificar "git merge-base --is-ancestor origin/main HEAD"
ork thread status <thread> --json

Delivery, CI and concurrency

ork ci prepare writes .ork-ci/<thread>.json in the thread’s worktree, with its branch; ork ci run --branch finds that bundle by branch name, fails an ork/* branch without a bundle and, outside a thread, runs only the manifest commands. ork ship registrar-pr --repo --pr records a PR merged into an external repository declared in ci.external_repositories, with the merge confirmed through the GitHub API and the declared check green on the PR head.

ork phase run refuses with concurrency.limite, exit code 3, when the project already has max_parallel_threads live sessions in other threads; --esperar waits for a slot. In a session without ledger access, the ork_decision_record MCP tool records the decision through the same contract as ork decisao registrar.

ork ship registrar-pr --dry-run, also with --repo --pr, runs the same checks and answers registraria, without writing ship_done, changing the phase or publishing the factory.

--remoto in ork ship --para, ork ship registrar-pr and ork ci status is the name of a remote, validated before any git call: outside the format, it refuses with ship.remoto-invalido or ci.remoto-invalido, without querying GitHub.

In version 0.6.10, a base push refused by branch protection becomes entrega.base-protegida, with guidance to deliver through a PR or MR. With verify.security_scan declared and without ci.required_for_ship, a loop category without a scan on the candidate SHA stops with human.pending before any push, and --allow-unscanned "<who>" delivers anyway, recording ship_unscanned_authorized in the ledger. ork ship registrar-pr and ork ci status recognize GitLab through the resolved forge, and on GitLab --repo accepts group, subgroup and project, with --pr being the MR iid.

ork ci prepare <thread>
ork ci run --branch <branch>
ork ship registrar-pr <thread> --repo <owner/name> --pr <n>
ork ship <thread> --para main --allow-unscanned "<who>"
ork phase run <thread> GO --prompt "<text>" --esperar

Verify receipt in CI

With ci.politica_recibo and the last verify green and complete, ork ci prepare writes the ork.ci-projecao-replay/v1 projection into the bundle’s recibo field, with no present HEAD. Without the key, the historical receipt stays.

ork ci run --branch applies the policy when it is configured. The runner validates the projection, reruns preparation, claims and canonical commands in an allowlisted environment and emits ork.ci-recibo-runner/v1 with the executed HEAD.

observar reports the state; avisar also uses the job’s channels; exigir fails a PR or MR targeting worktree.base_branch when the receipt is missing, invalid, divergent or the replay is degraded.

In a PR or MR with an enrollment on the company network, ORK_REDE_EMPRESA_READ_TOKEN gives read access to the network to attach the preflight. Without the credential, or on a fork, the output says the collective check was not performed, and the CI result does not change.

ork ci prepare <thread>
ork ci run --branch <branch> --json

OpenTelemetry telemetry

ork telemetria status shows the private opt-in, where it comes from, the destination without credentials and what is pending per thread; when off, it does not enumerate threads or read the endpoint. ork telemetria ligar and ork telemetria desligar store the choice in the private ork.telemetria-config/v1 file, with 0600 permissions; OTEL variables alone do not turn the feature on.

ork telemetria enviar sends OTLP/HTTP JSON batches with a cursor per ledger. --dry-run shows the same LogRecords from the start, or from the --offset byte, without headers, network or cursor.

ork telemetria maturidade computes the six agentic indicators over the local ledgers, in the window between --desde and --ate, with no network or opt-in. DORA stays separate and comes out as not confirmed without external facts.

With telemetry on, prompt content only comes from the stamped file, with secrets removed and dado_privado applied before the LogRecord, in gen_ai.input.messages. An unrecoverable source becomes a typed gap, and a transient error stops before the event.

ork telemetria ligar
ork telemetria status --json
ork telemetria enviar --dry-run --json
ork telemetria maturidade --desde <ISO> --ate <ISO> --json

Search memory by meaning

ork memory search --texto searches the tenant by meaning, combining vector and FTS, and is not deterministic; it cannot be combined with --tags or --thread. ork memory index maintains the local, idempotent vector index, and --dry-run estimates tokens and cost without calling the provider. ork memory status --sondar makes one real call and measures its latency. All three use the same tenant search universe; ork memory status and ork memory index display it per collection, count what stays out of search and warn when the index covers less than search can see.

When the universe was not read in full, or memory is off or unavailable, ork memory search --texto exits 1 with a typed reason and no results, also with --json; an empty universe read successfully exits 0. Reading the universe has its own deadline, memory.universo_timeout_ms, 90,000 ms by default.

ork memory search --tags is the deterministic tag search; with --thread, reading is restricted to the thread, and --restrito and --janela require --thread (memory.query.thread-required).

In version 0.6.10, the memory DSN can come from a file, outside the agents’ environment: memory.database_url_file accepts only the ~/.orkastery/segredos/<nome> form, and the two sources are exclusive, so ork never falls back from one to the other. The file must be a regular file of the user running ork, with exact 0600 or 0400 permissions, up to 4 KiB and a single line with the DSN. Missing, unsafe or invalid, memory degrades to files with dsn.arquivo-ausente, dsn.arquivo-inseguro or dsn.arquivo-invalido. ork memory status shows the source, never the value, and ork doctor gains the DSN source and DSN in the environment checks; the second lists, by name, the variables carrying the project DSN or a URL with a password that the agent reads with env.

In the same version, the core/scripts/prova-busca-semantica.sh proof shows, for each pair, the position and similarity of each target in the vector ranking and how many search entries have a coherent vector in the index, without changing the pairs or the exit codes. Without arguments, it uses the default pairs the owner decided on; pairs passed as arguments replace the defaults.

ork memory status --sondar
ork memory index --dry-run --json
ork memory search --texto "<phrase>" --json

Query the code graph

ork grafo indexar builds the index for a clean HEAD, in project state and outside git; --verificar extracts again and checks contract, bytes and determinism. ork grafo vizinhos, ork grafo chamadores, ork grafo importadores and ork grafo caminho respond from the edges, as text or --json, with each edge’s extractor and evidence. ork grafo status shows the HEAD index, ork grafo amostra supports manual edge audits and ork grafo limpar deletes indexes that are not the HEAD of any tree.

Answers are partial by construction: they contain only what the extractor proves, and they say so. With a modified tree, the answer reflects HEAD and warns about it. All of ork grafo needs typescript and micromark in the node_modules of the ork package itself: they are package dependencies with exact versions, and npm install -g brings them. Installed inside a project or through npx, npm hoists them out of the package, and the refusal is grafo.parser.indisponivel, with the fix in ork grafo status and ork doctor. The graph needs Node 20.19, 22.12 or later.

In version 0.6.10, answers check provenance against the tree: the text marks with [fonte modificada na arvore] the edge whose source, target or evidence changed after HEAD and lists up to five sources that changed; in JSON, each edge carries the arvore field, and the answer carries the fontes field.

With the index of an ancestor revision, ork grafo indexar re-extracts only what the change reaches, with the same bytes as a full extraction, and says when it ran in full and why. With --json, --teto-bytes N caps the answer at N bytes, dropping the edges farthest from the target first; a path is never cut and refuses with grafo.consulta.teto-excedido. Without a HEAD index, the refusal names the case (grafo.indice.ausente, grafo.indice.outra-revisao or grafo.indice.outro-extrator) and the fix. The same queries are in the project MCP server behind the grafo.mcp flag, off by default.

ork grafo contexto <thread> builds the thread’s compact package from the worktree diff, GOAL, PLAN and claims, with the links closest to the diff first and evidence as tuples; the ork.thread-graph-context/v2 JSON is capped at 32,768 bytes by default, from 4,096 to 65,536. Without a worktree, the diff is ignored and the answer says so; with the flag on, ork_grafo_contexto returns the same JSON over MCP.

ork grafo indexar --verificar
ork grafo status
ork grafo chamadores <symbol>
ork grafo importadores <file> --json
ork grafo caminho <from> <to>
ork grafo contexto <thread> --json

Agile Check and Classic Check

In version 0.6.10, ork verify runs the Agile Check by default, and the Classic Check remains the CHECK as it always was, with the same commands, the same order and the same tools. --check classic or --check agile overrides the manifest’s check.mode, and verify_run records the mode and where it came from: opcao, manifesto or padrao. The run text shows the check line.

The Agile Check runs the preparation once and the claims against its product, with a hash guard, and runs the test lanes and the scan in parallel without exceeding the machine’s CPU; the lane that starts containers waits for the scan, through check.agile.apos_a_varredura. The Trivy database is cached, the lint image is prebuilt and the native binary comes first; an unavailable piece falls back to the Classic equivalent and is recorded in verify_run. In both modes, verify_run records the duration of each command, claim and run, and the CHECK report and pr-body.md state the run’s mode.

The run and the baseline wait their turn in the heavy resource queue: concurrency.max_parallel_heavy, 1 by default, and 0 disables it. At most that many run at once in the project’s worktrees, and never two in the same worktree. Waiting in the foreground goes up to --esperar or 60 min and then refuses with fila-pesada.ocupada, exit code 3, naming who holds it.

Over MCP, ork_verify starts the same round in a server worker and returns a rodadaId at once, and ork_verify_status follows queue, lane and verdict, under the ork.verify-mcp/v1 contract. When the runner names the failures, the round compares test by test within the same command, and a new failure in a command that already failed blocks with verify.regression. With ORK_CHECK_TMPFS=auto, the Agile lanes use tmpfs when memory, space and writes have headroom.

ork verify <thread>
ork verify <thread> --check classic
ork verify <thread> --check agile --esperar 20

Security and quality scan

ork security-scan scans HEAD, the versioned content in a detached worktree, with the catalog tools: secrets, dependencies, SAST, IaC, workflows, quality and license. Each finding is compared with the base by fingerprint and comes out as new, existing or dismissed. The base is --base, or the merge-base with <remote>/<worktree.base_branch>.

A new secret and a new high or critical vulnerability make the command exit 1; quality and license only report. The report.json, in the ork.security-scan/v1 contract, and report.md live in .orkastery/security-scan/<sha>/ of the main tree, and never carry a secret’s value. Without --json, the summary ends with the path to report.md.

With verify.security_scan in the manifest, ork verify runs the scan as a fourth command, and ork doctor gains the security scan check. ORK_SECURITY_SCAN_REQUIRED=1 makes a loop category tool that is unavailable or failed exit 1. The unsupported-ci and no-requirements-file reasons are declared in the report and do not count as an unscanned category. --check classic or agile chooses the parallelism: in Agile the tools run at the same time, with the Trivy database in persistent cache, and in Classic one at a time.

ork security-scan --json
ork security-scan --base origin/main
ork security-scan --check classic

Private data barrier

In version 0.6.10, ork privado varrer looks for private data in the repository of the current directory: home paths, email outside the allowed list, hostnames in machine format and the terms of the local list, which stays outside git. The list is termos-privados.txt in ~/.orkastery/ and in the project’s .orkastery/, or the file named by ORK_TERMOS_PRIVADOS, and is only read as a regular file of the current user, without a link and with 0600 permissions. Exceptions live in orkastery.privado.json, with an exact pattern and value, no wildcard, and a required reason.

Without --base, the scan reads the names and all the content of versioned files, binaries included. With --base REF, it reads the diff against the merge-base with REF and the untracked files. The report states file, line, column and pattern, never the value, and a file name with a finding comes out as [omitido]. It exits 1 with a finding under the dado_privado policy in block or without the policy declared, 0 in warn and off, and 2 when the scan does not run.

With the dado_privado policy in the manifest, ork verify scans the thread diff and fails with policy.violation in block, and ork ci prepare refuses to write the bundle when there is a finding. In the core CI, the scan is a step of the documentacao job.

ork privado varrer
ork privado varrer --base origin/main
ork privado varrer --json