Standards
Product direction
Monthly planning summary and a link to the tracking source.
October 2026 summary
Planning covers conduction, host integration, governed memory and documentation quality. The top priority is still conduction HITL by alternatives: every request ork opens to the owner becomes a choice among 3 to 5 options, one of them recommended, with no pasted text. Also still in development: a blocker only the owner can clear in the terminal becomes a request to them with the exact command, and a profile per dispatch, with load-based distribution across accounts. This page neither replicates tickets nor turns intent into a promise of availability.
Since the previous review, new items entered the roadmap and others moved forward, some of them into a published version:
- #Noctum mode, autonomous arcs of 6 to 12 hours across many threads and machines: in development, not announced and not yet validated. The six slices and the single-machine integration are in the code, and the link between two machines has only been exercised in a simulated trial. No real pilot has run, and the version that announces the mode requires the first real pilot approved and ratified by the owner.
- Conduction inside Claude Code, with ten mods in four waves: in development, with all ten mods in version 0.6.10. The foundation and waves 1 and 2 arrived in earlier versions, and wave 3 (the Delivery Track, the Task Gate, the MASTER Ritual and the Session Messenger) completed the set; exposure is partial and enablement is in progress.
- Grok Build alongside Codex and Claude Code: in development. Since 0.6.1, explicit dispatch (
ork phase run <thread> GO --runtime grok --model <model>), account profiles, block setup and the core’s MCP server outside the sandbox are in the code. In 0.6.5, F5a makes headless dispatch use a closed list of phase commands (git, node, npm, ork, grep, find, ls and cat): a command outside the list is denied without a prompt, and the sandbox does not widen. Live validation, the F5b trial, is still open. - Agile Check faster than Classic, measured on the same machine (RM-074): closed on November 8 with the measured result. The final bench measured the Agile Check 1.21 times faster than Classic on the cold round and 1.28 times on the warm one, with confidence, below the target of about 1.7 times, and kept the whole battery. The three gaps it found continue in RM-085. The bench compares both modes on the same commit and publishes a ratio only with three valid pairs in alternating order. Version 0.6.2 moved heavy tests in-process and reuses proofs already closed; 0.6.3 cuts the system CPU of the expensive tests and gives the reserved CPUs back to the hermetic lane when the scan and the integration finish; 0.6.4 reserves CPU by each piece’s measurement, reuses the Semgrep scan when the rules do not change and separates the cold round from the warm one on the bench, discarding a pair with external load. Version 0.6.5 brings M8 and M9: the scheduler decides by the median of several samples and, under high steal, reduces capacity in steps instead of falling back entirely to the Classic order, and the integration reserves only the CPU it measures. No test was removed.
- The same item (RM-074) in 0.6.6 and 0.6.7. Version 0.6.6 fixes the session watcher under CPU contention: dispatch waits for the watcher to be ready while it is alive, instead of giving up after a short fixed deadline, and the phase closes again. Version 0.6.7 brings M11, M12 and M13: independent claims start together with the lanes, and the CPU the scan gives back goes to the chained integration or, when the integration already runs in parallel, to the hermetic lane; a claim that repeats a test already green in a lane reuses that lane’s proof when the environment is the same and, if it is not, really runs; and the bench only discounts the CPU delay measured in the round itself, with no discount when the measurement is missing. Selecting
node --testtests by the import graph (M10) was archived without a merge:node --testkeeps the whole lane, and RM-081’s native selection still applies to the runners that have it. The final result is in the previous item, and Classic does not change. - CHECK that understands any test runner and reruns only what is needed: delivered in 0.6.5, with all five slices. Verify reads the JUnit report from pytest, Vitest, Jest, Playwright or gotestsum declared in the manifest (F1), protects marked clock-bound tests under high steal (F2), reruns the failed tests first on a fix round (F3), gives each lane the workers of its reservation and serializes lanes that share a resource (F4), and uses the runner’s native selection on the warm round (F5). A partial receipt never counts for SHIP, and the ratio against a full round still comes from the bench. In 0.6.7, the product’s own process tests close on the observed event (queue, receipt, state, result or process exit), not on a fixed clock deadline; time rules keep an injected clock or the verify’s protection.
- Conduction findings from a network of three VPSs: six items, RM-075 to RM-080. Phase end read from the result (RM-075) arrived in 0.6.5: a session that ends without asking the owner anything no longer becomes a human wait, a clean GO without a commit comes out as
go.sem-mudanca, Codex closes throughtask_completeeven when the watcher loses the state, and retry does not dispatch an already concluded phase again. A receipt valid while nothing changes (RM-078) and claims that the fix passes (RM-079) arrived in 0.6.7, in the next two items. The other three remain planned: the heavy queue with priority and slot reservation, independent review measured by the HEAD’s builder and complete output with a typed cause in every verify command. - Verify and review receipt valid while HEAD, claims and plan do not change (RM-078): arrived in 0.6.7. The receipt now depends on the proven HEAD, the active claims and the plan, and a GO without a commit that keeps all three repeats neither the verify nor the review. A claim that declares an external dependency on the network or the clock (
ork claims add --dependencia-externa rede,relogio) still asks for a new verify, and only a session with a confirmed commit in the chain counts as an author. - Claims that the fix passes and typed withdrawal with a replacement (RM-079): arrived in 0.6.7. A new CHECK claim states the correct behavior and passes with the fix, and lint warns about an artificial or circular proof. A typed withdrawal with a green replacement does not stop the thread, and a withdrawal without one becomes a request to the owner.
ork fix openreads the CHECK review,ork fix statusshows what remains, and the receipt, the CLI and MCP redact commands and reasons. In 0.6.8, F2 makes a claim run in CI the way the session wrote it: repeated--verificarand--comandooptions are all kept, in order, a new claim with a local base-branch ref is refused with the fix toorigin/<base>, and the CHECK review lives only in the check artifact, in the thread’s state. Redacting commands on the surfaces that already existed, such as the verify text, the baseline, the ledger, the board and the audit output (A7 to A12), is left to RM-082. - Credentials redacted on every command, reason and detail surface (RM-082): planned. It is the first of four conduction findings from November 6 to 8, with no priority set by the owner yet, and it picks up RM-079’s A7 to A12 debt and the CI bundle: no included surface stores or shows a credential as text, and runs stay comparable through the fingerprint of the original command.
- Roadmap state that follows the delivery (RM-083): planned. The delivery PR itself prepares the post-merge state of the item page and the index, and main no longer depends on a manual sync PR.
- Session that continues with many loose Git objects (RM-084): planned. The core packs the objects without loss, outside the session sandbox, or warns before the limit, and the Git metadata guard does not change.
- Verify with causal scope, assertion message and independent CHECK in #Fast (RM-085): planned. It addresses the three gaps that RM-074’s final bench found: verify measures all the work it creates, records the safe detail of the failed assertion and accepts an independent CHECK requested in an existing #Fast.
- Kimi Code and OpenCode validated alongside Codex and Claude Code (RM-086 and RM-087): planned, at the owner’s request on November 9, with one page per runtime, in discovery. Each page has six slices: headless dispatch, session end, profiles and quota, setup with MCP and documentation, a live trial and, after the trial, conducting the factory, as Codex already does. On November 11, the owner decided the authentication: Kimi Code uses the Plus or Pro subscription, with OAuth sign-in; OpenCode uses an API key, with OpenRouter or a local GPU behind an OpenAI-style endpoint, without OpenAI, on the V2 line.
- An older version never becomes npm’s latest (RM-088): arrived in 0.6.9. The publish workflow chooses and validates an explicit dist-tag before publishing: a lower stable version goes to previous, a prerelease goes to its own channel, and latest never moves back.
- Company-wide collective Orkastery Network, with mandatory use (RM-089): in development, at the owner’s request on November 10. Version 0.6.9 shipped the enrollment on the collective network (F1), the mandatory collective reservation (F2) and the CI receipt policy, with observar, avisar and exigir (F4). Version 0.6.10 shipped the conflict warning before the merge, with a suggested order (F3), and the basic verify receipt on the HEAD executed by CI (F5). F6, with
ork adotar, the terminal notice and the auditable exception, is in development, and F7, with the company factory’s signed receipt, is planned. - OpenTelemetry- and Grafana-compatible telemetry for SDLC Analytics (RM-090): all five slices released, at the owner’s request on November 10. Version 0.6.9 shipped the ledger’s OpenTelemetry logs with private opt-in (F1), prompt content without secrets (F2), correlation with Claude Code and Codex sessions (F3) and the Collector and Grafana models (F4). Version 0.6.10 shipped the six agentic SDLC Analytics indicators per maestro and per team, with DORA kept separate (F5).
- Ork Mesh on the site, for companies and teams (RM-091): in development. The companies page presents the company network and telemetry and shows, in each block, the release of each slice or that it is still in development.
- Authenticated ratification of a public contract in #Auto (RM-092): planned, at the owner’s request on November 12. A new public contract only proceeds in #Auto after the owner ratifies its exact names through an authenticated channel and CHECK verifies the receipt.
- Site as a product, the two public sites held to the core’s rigor: in development, and nothing is published without the owner’s approval.
- CLI in English and pt-BR: in development, with slices F1 to F3e released. In version 0.6.10, the first-run path (from
ork doctorandork inittoork thread newandork phase run),ork verify,ork ship,ork pulse,ork master,ork retry plan,ork retry run,ork roadmap statusand the help, among others, print English whenowner.languageor the locale asks for en. F3e, in 0.6.2, brought the remaining routes into the catalog, such asork master, the Maestro panel and the early project refusals, plus the human presentation of the graph and of security-scan, without changing codes, data or--json; the surfaces outside the CLI are still ahead. - Public repository without private data: in development, with slices F1 and F2 released in 0.6.0: the private-data barrier in
ork privado varrerand, with thedado_privadopolicy, inork verifyandork ci prepare. The evidence file whose name carried a machine name has already been renamed to a generic one. In 0.6.2, F3 makes the Orkastery Network snapshot publish only a generic machine alias, never the real hostname, and F4 gives the PR guide the honest-report template for when private data slips out. - Search by meaning in memory: the owner chose the proof pairs, and the proof passed in every measured run; the item is still in validation.
- Code graph: phase consumption gained citations, a second hop and provenance checked against the tree, with the
grafo.mcpflag still off. - Fast and reliable verify: in pilot. Version 0.6.0 brought the waits, the full output in
ork verifyand the fixes for flaky tests and conduction defects, besides theork verifythat accounts for CPU steal. Version 0.6.2 brought the test-by-test comparison within the same command, claims that are born runnable in CI, the CHANGELOG rule in verify and the asynchronous verify over MCP, which returns arodadaIdat once and is followed withork_verify_status. - Blocker only the owner can clear: it recognizes the runtimes’ current messages and reaches the owner in the summary, with the command. In 0.6.2, the request to conduction became a separate contract: an action the session sandbox cannot run stays with conduction, with the exact command, and does not reach the owner.
- Explicit target project: the second slice of the honest answer was delivered by PR, and the item awaits the owner’s pilot.
- Factory on several machines: in pilot. In 0.6.0, the baseline refuses a dependency missing from the worktree, and a new worktree starts from the current remote base.
- Company Brain: written knowledge has been released since 0.6.0, with all three slices:
ork conhecimentoin the CLI, in MCP, in OpenClaw and in Hermes. - Agile Check and Classic Check: available. The default CHECK has been the Agile Check since 0.6.0, with the same coverage as Classic, a single preparation and parallel lanes; Classic returns with
check.mode: classic orork verify --check classic. Since 0.6.1, the hermetic lane, the scan and the integration share the machine’s same CPU budget, and, under saturation, verify returns to the Classic order and records the reason; in 0.6.5, high steal first reduces capacity in steps. - Security scan in CHECK: available since 0.6.0.
ork security-scanscans HEAD and the base with official tools pinned by digest and classifies every finding as new, existing or dismissed; a new secret and a new high or critical vulnerability fail the run. It turns on withverify.security_scanin the manifest. - DSN password outside the agents’ environment: available since 0.6.0. Memory reads the DSN from a file in the core’s secrets directory, with strict permissions, instead of an environment variable.
- GitLab with parity to GitHub: available since 0.6.0, with its own MR, CI and SHIP transport.
- Exhausted account does not stop the phase: in development, with slices 1 to 3 released in 0.6.0. The quota profile switch passes the original prompt when nothing was produced, or a resume prompt with a checkpoint when the session already produced.
Read the current status
The CLI report groups initiatives, marks owner decisions with #HITL and shows the next step without changing the roadmap. ork network roadmap provides the same report with threads from every machine and the source of each part. Current sources cover short HITL requests and conduction HITL by alternatives, owner blockers that become requests, profile per dispatch, the Company Brain context package and decision dossier, the code graph, search by meaning in memory, explicit target projects, the network roadmap, the Orkastery Network of a person’s machines, the experience pack, the marketplace plugin, conduction fixes, contribution guides and, since October, #Noctum mode, conduction inside Claude Code, the site as a product and the CLI in English and pt-BR, and, in November, the conduction findings from a network, the CHECK that understands any runner, the conduction findings from November 6 to 8, the owner’s requests from November 9, the company-wide collective network and OpenTelemetry-compatible telemetry. Consult the repository roadmap for each initiative’s state: merged or branch code does not establish publication, so check the state dimensions at the source.
ork roadmap status
ork network roadmap