Changelog

0.3.37 (2026-09-20)

• Isolated every skill-root archive in its own owner-only staging directory so extraction cannot merge or overwrite sibling skills.

0.3.36 (2026-09-20)

• Required host adapters to select their profile-scoped durable root before writing skills and standardized atomic relative current symlinks.

0.3.35 (2026-09-20)

• Required bootstraps to fetch raw current release metadata, reject every archive-version mismatch before writing files, and never relabel archives.
• Prevented one-shot Hermes turns from leaving OAuth in turn-owned background processes that disappear before approval.

0.3.34 (2026-09-20)

• Kept every OpenClaw install command and durable skill artifact inside the selected profile instead of guessing shared state paths.

0.3.33 (2026-09-19)

• Prevented fail-closed null results from being reported as successful required Workflow pilot outputs.

0.3.32 (2026-09-19)

• Required live Workflow pilots to satisfy the user's requested output, not merely finish execution, before agents report completion.

0.3.31 (2026-09-19)

• Persisted resumable non-secret install checkpoints at human approval and restart boundaries.
• Clarified native OAuth token storage, fresh child-session evidence, and the single bundle version to report as installation identity.

0.3.30 (2026-09-19)

• Required raw, revalidated compatibility metadata and rejected mismatched bundle metadata before installing files.
• Kept Hermes OAuth in a host-owned foreground terminal through approval.

0.3.29 (2026-09-19)

• Required hosted Workflow control-plane calls during the main tool-enabled turn, before optional probes or host text-only finalization.
• Required installers to leave a staging directory before removing it so the remaining setup commands keep a valid working directory.

0.3.28 (2026-09-19)

• Kept OpenClaw skill trees and provenance in its durable profile before native registration.
• Required role gates to run through fixture-tested production selectors before enrichment and added recovery for interrupted Workflow tooling.
• Clarified the compact transport query versus the bare OAuth resource.

0.3.27 (2026-09-19)

• Required literal HTTPS source links and fail-closed identity conflicts in structured research.
• Kept title-based executive discovery independent of provider seniority taxonomies while accepting qualified company-level CEO titles.

0.3.26 (2026-09-19)

• Reused validated Workflow source artifacts instead of retransmitting large inline bundles between check and creation.
• Documented the normalized GTM field and verification shape, plus directory filter behavior that can otherwise hide exact role matches.
• Made final structured research fail closed on requested constraints and conflicting identity evidence.

0.3.25 (2026-09-19)

• Qualified vendor person IDs with the exact matching enrichment adapter so IDs cannot cross provider boundaries inside the generic waterfall.

0.3.24 (2026-09-19)

• Preserved stable provider person IDs through role discovery and enrichment instead of degrading them to partial names or redundant provider calls.
• Made exact role gates and per-row provider failure handling explicit for reusable prospecting Workflows.
• Defined the compatibility record for installed setup skills and simplified the common human-owned path and fresh-session verification boundary.
• Kept requested source URLs and the canonical compact MCP query spelling.

0.3.23 (2026-09-19)

• Preserved complete released skill entrypoints during Hermes activation and kept canonical descriptions within its registration limit.
• Kept requested Workflow creation and testing in one autonomous lifecycle.
• Failed closed when a narrower sales title could not prove the requested company-wide role.

0.3.22 (2026-09-18)

• Kept Hermes OAuth alive for its full browser-approval window.
• Required current-role corroboration before enriching directory candidates.
• Prevented interactive OAuth calls from inventing external_user_id values.

0.3.21 (2026-09-18)

• Made the shared skill tree the editable source for every host adapter and generated the Hermes-compatible view from it.
• Moved the GTM waterfall guide into the canonical plugin bundle so MCP, HTTP, Agent Skills archives, plugins, and downstream importers serve one copy.

0.3.20 (2026-09-17)

• Kept Workflow execution ceilings autonomous by default instead of exposing them as required customer inputs.
• Added a Hermes-compatible Agent Plugins overlay with the Locus MCP server and all three skills, plus pinned validation of the installed runtime.

0.3.19 (2026-09-17)

• Added the reviewed GTM enrichment Workflow asset to native and skill-only packages and made it the preferred starting point for enrichment workflows.
• Grounded task accounting in durable call and run receipts instead of global balance deltas.

0.3.18 (2026-09-17)

• Kept simple Workflows focused on the shortest checked, fixture-tested, saved, and piloted path using reviewed assets.
• Made per-run charges and receipts authoritative under concurrent workspace activity.

0.3.17 (2026-09-17)

• Scoped narrow enrichment calls to requested fields instead of broad profiles.
• Hardened cross-host skill archive validation, profile-scoped hydration, and secret-environment inspection.

0.3.16 (2026-09-17)

• Added the required SSH loopback forwarding path when OpenClaw OAuth approval runs in a browser on a different machine.

0.3.15 (2026-09-17)

• Front-loaded people, company, and verified-work-email intent so ordinary outcome prompts activate Locus without naming a provider or tool.
• Scoped filesystem installs to the selected environment and host profile, and made Hermes/OpenClaw OAuth verification recover from completed or timed out login processes before starting a second authorization.
• Made hosted Workflow quality defaults fail closed and preserved durable run IDs across client timeouts while reducing unnecessary guide and search load.

0.3.13 (2026-09-16)

• Clarified that the execute meta-tool takes idempotency_key only as a top-level parameter, while _locus applies only to advertised direct tools.
• Cleaned up malformed execution-recovery line wrapping in the Locus skill.

0.3.12 (2026-09-15)

• Corrected OpenClaw OAuth recovery so managed shells keep the native login listener durable and restart expired authorization state instead of assuming a callback-paste prompt exists.
• Defined precedence between the live compatibility record, bootstrap, and installed setup skill without making OpenClaw the shared setup default.
• Documented enabled-first catalog discovery and the explicit disabled-endpoint filter used when an agent needs to request access.

0.3.11 (2026-09-15)

• Made environment-level adapter availability authoritative so an unavailable stage CLI cannot override a working native MCP path.
• Documented durable OpenClaw OAuth startup, local-directory archive installs, and request timeouts suitable for long-running paid calls.
• Clarified deferred fresh-session verification and the shared decoded-JSON Workflow response contract.

0.3.10 (2026-09-15)

• Made the Hermes OAuth setup path declarative so an unauthenticated discovery probe cannot prevent the server from being registered before login.
• Clarified that OpenClaw CLI authentication and fresh-session skill discovery do not require its Gateway, while retaining the foreground OAuth-listener requirement.
• Bound generated-CLI setup to the compatibility record's environment-specific base URL and documented the explicit owner-only file-storage fallback.

0.3.9 (2026-09-15)

• Documented the exact hosted row.call raw-body contract, input/output parse hooks, provider-boundary probing, targeted source edits, response inspection, and execution-versus-output pilot readiness.
• Kept shared setup host-neutral while moving OpenClaw and Hermes commands, storage, callback, reload, and lifecycle behavior into optional host references.
• Tightened reuse so an existing adapter is healthy only when its environment, authenticated grant, and installed instruction bundle all match.

0.3.8 (2026-09-15)

• Made ordinary Locus API and Workflow execution autonomous instead of adding Locus-specific confirmation gates for billed calls.
• Made compact and legacy tool vocabularies explicit while treating the live connection inventory as authoritative, including outcome and access-request tools.
• Improved Workflow authoring guidance for manifest kinds, digest-carried validation, path-scoped draft patches, consistent revisions, readiness, real executable fixtures, and compact agent behavior.
• Clarified Agent Skills archive installation, user-scoped source-of-truth and workspace mirrors, main-session tool verification, and OpenClaw timeout and stale-runtime-state recovery.

0.3.7 (2026-09-14)

• Made MCP and the generated CLI peer installation adapters selected from the user's explicit choice, a healthy existing setup, and actual host support.
• Made one-line setup completion include adapter authentication, complete versioned skill delivery, a free readiness probe, and fresh-session discovery instead of stopping after copying SKILL.md.
• Added safe filesystem hydration metadata to the guide manifest and improved Workflow guidance for inline validation, deterministic repair, truthful readiness, and integrity counts.

0.3.6 (2026-09-14)

• Released the corrected environment-aware remote-guide serving contract so stage URLs and immutable historical guide versions can be verified without changing the already published 0.3.5 artifacts.

0.3.5 (2026-09-14)

• Added a persistent-filesystem instruction tier between native skill registration and remote MCP guide retrieval.
• Required complete, digest-verified, versioned skill trees and honest fresh-session activation reporting for hosts without a native skill registry.
• Setup skill (v1.2.1) states the path rule: human-owned accounts are the default, the agent-owned path is only for headless runtimes or an explicit request and requires telling the user what such an account cannot do, an already-authenticated connection without an agent credential is human-owned, and a funding request never triggers agent-native signup. Usage skill (v1.1.2) and README send funding on human-owned accounts to the dashboard.

0.3.3 (2026-09-13)

• Corrected the hosted Workflow lifecycle so agents save a checked immutable version before starting a live pilot.
• Documented the exact compact workflow_run request and host-approval boundary, including the distinction from workflow_runs actions.
• Required live contract digests for saved bindings and clarified bounded remote-guide pagination.

0.3.1 (2026-09-13)

• Made hosted Workflows automatically eligible for live Locus-reviewed Tools, maintained Recipes, and explicitly verified external buyer-rail listings; exceptional approvals and emergency denies remain server-controlled.
• Clarified that catalog verification never bypasses connection scope, tool enablement, availability, contract validation, or run budgets.

0.3.0 (2026-09-11)

• Added provider-agnostic category and curated-pack discovery, including live group lookup, scoped pagination, and remote guide retrieval.
• Added the locus-workflows skill and a fixed-runtime TypeScript template for checked, fixture-tested, budget-bounded hosted Workflows with durable run inspection, cancellation, and resume guidance.
• Added one canonical versioned guide registry and generated JSON/Markdown bundle metadata so agents without persistent files can retrieve the same reviewed instructions served by native plugin packages.
• Extended manifest, release, and ChatGPT submission checks for the new guide and conditional hosted-Workflow tools.

0.2.1 (2026-09-04)

• Grok manifest precedence fixed: the root .mcp.json (the file Grok actually consumes while a root plugin.json exists) now uses Grok's headers dialect with its own attribution; the Codex-native definition moved to agents/codex/.mcp.json and the Codex manifest points there.
• README documents every network endpoint and credential the optional agent-owned setup path can touch, and the repository-layout table states the real per-host manifest resolution.
• The usage skill (v1.0.2) requires an estimate plus user confirmation before the first billed call a task did not explicitly authorize, and describes the tool surface scope-independently.

0.2.0 (2026-09-04)

• Skills aligned with production behavior (v1.0.1): capability-tool guidance, AgentMail inbox contract mirrored as a scoped reference, and updated connection steps.
• ChatGPT app submission bundle: all ten tools a full-scope OAuth session exposes, with server-declared annotations, review-facing justifications, and reproducible reviewer test cases; guarded by check-manifests.
• Codex manifest MCP definition moved to the native root .mcp.json; production smoke script and stronger CI validation.
• OpenClaw install simplified to the bundle's own server mapping.

0.1.0 (2026-09-02)

Initial release.

• Locus MCP server connection (streamable HTTP, OAuth 2.1 discovery) for Claude Code, Codex, Cursor, Grok, OpenClaw, and Agent Plugins clients.
• The locus skill: tool routing, cost quotes, idempotent billing, error recovery, and spend safety.
• The locus-setup skill: self-serve onboarding — account creation (human or agent-owned via AgentID), capability selection, and a user-requested Stripe funding handoff.