Outpost v11.1.3

Changelog

11.1.3

This release refreshes the documentation and visual identity without changing the library's runtime behavior or public API.

• Present five workflow scenarios on the English and French homepage: email, research, documents, customer requests and software. Link each step to the relevant Guide mechanisms, including System One typed decisions, while keeping the examples understandable on the page.
• Make the API the complete reference for prompts added by Outpost: exact templates, injection points, dynamic values, customization options and fixed instructions. Link the relevant Guide pages to these contracts.
• Make associated API contracts fill each row equally, with at most three columns. Remove experimental badges and generated maturity warnings from API pages; retain execution requirements, opt-in settings and live-validation limits in the contracts and guides.
• Replace the logo across the site, README banner and bilingual sharing previews. Add light/dark logo variants, browser-theme favicons and an Apple touch icon, with reproducible asset generation during documentation builds.

11.1.2

This release adds YAML workflow composition, terminal recipe dialogues and Git-free file workspaces while preserving existing Git APIs and defaults.

• Reorganize the English and French Guide around user tasks, with focused pages, preserved routes and section anchors, clearer setup paths and nine downloadable example projects. Make mobile diagrams readable without zoom and clarify desktop flow conditions. Correct journal/replay side effects, Git path handling and recovery instructions; document the local review report produced by the webhook example.

• Show every public API overload and its own property table, correctly mark required fields from mapped types, and link reference contracts to their practical guides.

• Define private vulnerability reporting through GitHub Security and security-fix support for the latest major version.

• Make recipe contract classification portable across Windows and POSIX source paths, and resolve temporary directory aliases in the macOS recovery test.

• Build public declarations before checking consumer fixture types so validation works on a fresh checkout.

• Prepare the Git-free Node image explicitly for container tests and document the required image pull.

• Create the host UID's account in the Git-free package test image, and adapt container documentation checks to the simplified setup example.

• Fix Windows output publication permission checks and relative file/directory link copies. Keep ownership and escape checks, and make recipe test cleanup, process assertions and host commands portable.

• Add explicit directory-copy, read-only/writable source-mount and ephemeral file workspaces without host Git, configuration version 3, provider workspace capabilities and file dispatch/report contracts. Preserve legacy Git APIs and configuration defaults. Journal protected output publication with conditional rollback and explicit recovery; conserve settled files and conversations for local or explicit Transport-backed portable resume. Cover independent commands, cross-process dialogue, interrupted publication and an installed CLI running with Node, Docker and tar on a host without Git. Native CLI file variants and live cloud/Firecracker validation remain unvalidated.

• Add a runnable Linear development recipe with a separate TypeScript application, validated private API-key caching, issue selection, plan approval, implementation, enforced tests and formatted observation. Extend terminal recipe interaction to unsigned gates with saved evidence and explicit decisions; signed gates retain external proof submission. Test the flow offline with real Git and simulated services; live Linear/Codex remains unvalidated.

• Handle durable recipe questions directly in the CLI: automatic terminal prompts, actor selection, saved answers and continuation without a custom runner. Add interactive/headless flags, keep JSON reporting final-only, and preserve pending questions on cancellation. Cover offline cross-process resume and real terminal input; no paid agent calls.

• Serve recipe and configuration JSON Schemas from documentation URLs under /schemas/, with immutable per-version copies reconstructed from stable Git tags. Preserve archives across GitHub Pages deployments, check JSON content and HTTP responses, and document editor setup in English and French.

• Complete the seven YAML composition lots with explicit experimental speculation, cumulative durable usage receipts, conflict resolvers, native reporters and OpenTelemetry. Keep machine settings in named local components, synchronize and close execution sandboxes before integration, and report resolver usage separately. Generate component documentation and enforce per-export/per-option parity classifications in CI. Add an offline community recipe and example; paid/cloud/remote-SDK validation remains absent.

• Compose YAML SQLite/HTTP/BullMQ queues, recipe and native workflow jobs, queued tasks, cron schedules and authenticated webhook services. Add explicit recipe enqueue/serve commands with native ownership, cancellation, token rotation and job identities. Verify local service composition and real Redis 7 without paid calls.

• Add durable YAML checkpoints with exact Git workspace restoration, artifacts, caches, signed gates, dialogues, quota pauses and cumulative budgets. Add recipe status/resume/answer/decide; require explicit replay and revision-fenced recovery after interruption. Exercise separate processes, preserved uncommitted files and original diff guards offline; live cloud restoration remains unvalidated.

• Compose format-3 workflows with structured JSON inputs, typed result references, conditions, transformations, local task callbacks, native loop and decision tasks, and isolated multi-repository dispatch. Preserve native scheduling and usage accounting, refuse overlapping shared workspaces, and project execution results to JSON. Add an offline example and YAML/TypeScript equivalence tests using real Git repositories.

• Compose the Outpost harness, model and decision providers, tools, skills, hooks, permissions, context strategies, routing, fallback agents and conversation stores from YAML. Format-3 agent steps accept dispatch options and structured response references. Callback extensions declare a static slot contract; schemas distinguish data schemas from component references. Compare delegated YAML and TypeScript runs offline, and exercise the compiled CLI against a local streaming model fixture.

• Extend YAML configuration with named providers, native sandbox and workspace options, portable profiles, CLI harness settings, MCP declarations, local transports and host-side secret selections. Generate option schemas from TypeScript contracts, reject dependency cycles before imports, resolve installed ESM-only extensions, and redact selected values from observation and reports. Cover local lifecycle execution and YAML/TypeScript request equivalence; cloud and paid agent calls remain unvalidated.

• Begin YAML composition parity with configuration format 2, recipe format 3, the /recipes registry/runtime API, explicit observation and final reports, static extension validation and a versioned public-contract inventory. Recipe execution is silent unless outputs or --json are requested; failures still print diagnostics.

• Add reusable YAML recipes with typed inputs, defaults and enums, dependency result references and editor schemas. Run with separate recipe and execution YAML files covering sandbox providers, CLI agents, model selection and explicit credential references; retain literal format 1 and TypeScript configuration compatibility. Add recipe init, allocation-free validate, list and integrity-checked fetch, repeated --input flags, three bundled Git catalogue recipes and third-party HTTPS catalogues. Fix CLI Git integration, bounded outputs and diagnostics, and final reporting after cleanup. Cover real local Git execution and cross-repository reuse; HTTP downloads are tested with simulated responses and paid agent/cloud runs remain unvalidated.

11.1.1

Publication was blocked after coverage passed: the SQLite recipe example deleted its temporary directory before closing its inspection queue on Windows. No package or documentation deployment was published. Version 11.1.2 fixes cleanup order and checks examples and package consumption before coverage in CI.

11.1.0

Publication was blocked by a Windows regression test that incremented an inode beyond JavaScript's safe integer precision. No package or documentation deployment was published. The fixture was corrected for the next release attempt.

11.0.1

This major release adds workspace integration safeguards, portable agent profiles, execution reports and state tracking, monetary budgets, incremental preparation, cloud dependency caches, offline workflow testing and optional secret sources. Update exhaustive consumers for FaultCode (guard, stuck), AgentEvent and ObservationEvent (stuck), stopped reasons (stuck), StorageCategoryName (runs) and WorkflowBudgetExceeded.dimension (cost). Implementations and test doubles of DispatchResult and WarmDispatchResult must provide report(); custom ObservationHub implementations must provide redacting and redact(). Existing entry points remain available. Best-of-N speculation remains experimental, and the live-validation limits below still apply.

• Add SecretSource and bounded fromSecrets() startup resolution with optional adapters for Vault/OpenBao KV v2, 1Password, Infisical, AWS Secrets Manager, Google Cloud Secret Manager and Azure Key Vault. Resolve declared names into immutable variables without writing values to disk or changing the host environment; sanitize vendor errors and retain caller ownership of host SDK clients. Add bilingual Guide/API documen...

Outpost v11.1.2

Changelog

11.1.2

This release adds YAML workflow composition, terminal recipe dialogues and Git-free file workspaces while preserving existing Git APIs and defaults.

• Reorganize the English and French Guide around user tasks, with focused pages, preserved routes and section anchors, clearer setup paths and nine downloadable example projects. Make mobile diagrams readable without zoom and clarify desktop flow conditions. Correct journal/replay side effects, Git path handling and recovery instructions; document the local review report produced by the webhook example.

• Show every public API overload and its own property table, correctly mark required fields from mapped types, and link reference contracts to their practical guides.

• Define private vulnerability reporting through GitHub Security and security-fix support for the latest major version.

• Make recipe contract classification portable across Windows and POSIX source paths, and resolve temporary directory aliases in the macOS recovery test.

• Build public declarations before checking consumer fixture types so validation works on a fresh checkout.

• Prepare the Git-free Node image explicitly for container tests and document the required image pull.

• Create the host UID's account in the Git-free package test image, and adapt container documentation checks to the simplified setup example.

• Fix Windows output publication permission checks and relative file/directory link copies. Keep ownership and escape checks, and make recipe test cleanup, process assertions and host commands portable.

• Add explicit directory-copy, read-only/writable source-mount and ephemeral file workspaces without host Git, configuration version 3, provider workspace capabilities and file dispatch/report contracts. Preserve legacy Git APIs and configuration defaults. Journal protected output publication with conditional rollback and explicit recovery; conserve settled files and conversations for local or explicit Transport-backed portable resume. Cover independent commands, cross-process dialogue, interrupted publication and an installed CLI running with Node, Docker and tar on a host without Git. Native CLI file variants and live cloud/Firecracker validation remain unvalidated.

• Add a runnable Linear development recipe with a separate TypeScript application, validated private API-key caching, issue selection, plan approval, implementation, enforced tests and formatted observation. Extend terminal recipe interaction to unsigned gates with saved evidence and explicit decisions; signed gates retain external proof submission. Test the flow offline with real Git and simulated services; live Linear/Codex remains unvalidated.

• Handle durable recipe questions directly in the CLI: automatic terminal prompts, actor selection, saved answers and continuation without a custom runner. Add interactive/headless flags, keep JSON reporting final-only, and preserve pending questions on cancellation. Cover offline cross-process resume and real terminal input; no paid agent calls.

• Serve recipe and configuration JSON Schemas from documentation URLs under /schemas/, with immutable per-version copies reconstructed from stable Git tags. Preserve archives across GitHub Pages deployments, check JSON content and HTTP responses, and document editor setup in English and French.

• Complete the seven YAML composition lots with explicit experimental speculation, cumulative durable usage receipts, conflict resolvers, native reporters and OpenTelemetry. Keep machine settings in named local components, synchronize and close execution sandboxes before integration, and report resolver usage separately. Generate component documentation and enforce per-export/per-option parity classifications in CI. Add an offline community recipe and example; paid/cloud/remote-SDK validation remains absent.

• Compose YAML SQLite/HTTP/BullMQ queues, recipe and native workflow jobs, queued tasks, cron schedules and authenticated webhook services. Add explicit recipe enqueue/serve commands with native ownership, cancellation, token rotation and job identities. Verify local service composition and real Redis 7 without paid calls.

• Add durable YAML checkpoints with exact Git workspace restoration, artifacts, caches, signed gates, dialogues, quota pauses and cumulative budgets. Add recipe status/resume/answer/decide; require explicit replay and revision-fenced recovery after interruption. Exercise separate processes, preserved uncommitted files and original diff guards offline; live cloud restoration remains unvalidated.

• Compose format-3 workflows with structured JSON inputs, typed result references, conditions, transformations, local task callbacks, native loop and decision tasks, and isolated multi-repository dispatch. Preserve native scheduling and usage accounting, refuse overlapping shared workspaces, and project execution results to JSON. Add an offline example and YAML/TypeScript equivalence tests using real Git repositories.

• Compose the Outpost harness, model and decision providers, tools, skills, hooks, permissions, context strategies, routing, fallback agents and conversation stores from YAML. Format-3 agent steps accept dispatch options and structured response references. Callback extensions declare a static slot contract; schemas distinguish data schemas from component references. Compare delegated YAML and TypeScript runs offline, and exercise the compiled CLI against a local streaming model fixture.

• Extend YAML configuration with named providers, native sandbox and workspace options, portable profiles, CLI harness settings, MCP declarations, local transports and host-side secret selections. Generate option schemas from TypeScript contracts, reject dependency cycles before imports, resolve installed ESM-only extensions, and redact selected values from observation and reports. Cover local lifecycle execution and YAML/TypeScript request equivalence; cloud and paid agent calls remain unvalidated.

• Begin YAML composition parity with configuration format 2, recipe format 3, the /recipes registry/runtime API, explicit observation and final reports, static extension validation and a versioned public-contract inventory. Recipe execution is silent unless outputs or --json are requested; failures still print diagnostics.

• Add reusable YAML recipes with typed inputs, defaults and enums, dependency result references and editor schemas. Run with separate recipe and execution YAML files covering sandbox providers, CLI agents, model selection and explicit credential references; retain literal format 1 and TypeScript configuration compatibility. Add recipe init, allocation-free validate, list and integrity-checked fetch, repeated --input flags, three bundled Git catalogue recipes and third-party HTTPS catalogues. Fix CLI Git integration, bounded outputs and diagnostics, and final reporting after cleanup. Cover real local Git execution and cross-repository reuse; HTTP downloads are tested with simulated responses and paid agent/cloud runs remain unvalidated.

11.1.1

Publication was blocked after coverage passed: the SQLite recipe example deleted its temporary directory before closing its inspection queue on Windows. No package or documentation deployment was published. Version 11.1.2 fixes cleanup order and checks examples and package consumption before coverage in CI.

11.1.0

Publication was blocked by a Windows regression test that incremented an inode beyond JavaScript's safe integer precision. No package or documentation deployment was published. The fixture was corrected for the next release attempt.

11.0.1

This major release adds workspace integration safeguards, portable agent profiles, execution reports and state tracking, monetary budgets, incremental preparation, cloud dependency caches, offline workflow testing and optional secret sources. Update exhaustive consumers for FaultCode (guard, stuck), AgentEvent and ObservationEvent (stuck), stopped reasons (stuck), StorageCategoryName (runs) and WorkflowBudgetExceeded.dimension (cost). Implementations and test doubles of DispatchResult and WarmDispatchResult must provide report(); custom ObservationHub implementations must provide redacting and redact(). Existing entry points remain available. Best-of-N speculation remains experimental, and the live-validation limits below still apply.

• Add SecretSource and bounded fromSecrets() startup resolution with optional adapters for Vault/OpenBao KV v2, 1Password, Infisical, AWS Secrets Manager, Google Cloud Secret Manager and Azure Key Vault. Resolve declared names into immutable variables without writing values to disk or changing the host environment; sanitize vendor errors and retain caller ownership of host SDK clients. Add bilingual Guide/API documentation and a selectable service example. Deterministic HTTP and SDK-boundary tests cover behavior; live service authentication remains unvalidated.
• Add @elie-laloum/outpost/testing with scriptedAgent() and createMemorySandboxProvider() for offline workflow CI: predefined events, usage, real Git commits, response repairs and strict command fixtures. Add bilingual Guide/API documentation and an executable workflow example; Git workspaces remain on disk and transfers, cold conversation resume and arbitrary command execution are unsupported.
• Add opt-in Vercel and Daytona dependency download caches backed by a caller-owned Transport: restore before preparation, archive before disposal, retain compatibility hashing and fence concurrent publishers. Add bilingual Guide/API documentation and a cloud example. Deterministic SDK and filesystem fixtures cover binary transfers, permissions, concurrency and cleanup failures; live cloud/S3 validation remains pending.
• Add incremental lifecycle preparation with when: changed(files): fingerprint exact file contents in the hook environment and reru...

Outpost v11.0.1

Changelog

11.0.1

This major release adds workspace integration safeguards, portable agent profiles, execution reports and state tracking, monetary budgets, incremental preparation, cloud dependency caches, offline workflow testing and optional secret sources. Update exhaustive consumers for FaultCode (guard, stuck), AgentEvent and ObservationEvent (stuck), stopped reasons (stuck), StorageCategoryName (runs) and WorkflowBudgetExceeded.dimension (cost). Implementations and test doubles of DispatchResult and WarmDispatchResult must provide report(); custom ObservationHub implementations must provide redacting and redact(). Existing entry points remain available. Best-of-N speculation remains experimental, and the live-validation limits below still apply.

• Add SecretSource and bounded fromSecrets() startup resolution with optional adapters for Vault/OpenBao KV v2, 1Password, Infisical, AWS Secrets Manager, Google Cloud Secret Manager and Azure Key Vault. Resolve declared names into immutable variables without writing values to disk or changing the host environment; sanitize vendor errors and retain caller ownership of host SDK clients. Add bilingual Guide/API documentation and a selectable service example. Deterministic HTTP and SDK-boundary tests cover behavior; live service authentication remains unvalidated.

• Add @elie-laloum/outpost/testing with scriptedAgent() and createMemorySandboxProvider() for offline workflow CI: predefined events, usage, real Git commits, response repairs and strict command fixtures. Add bilingual Guide/API documentation and an executable workflow example; Git workspaces remain on disk and transfers, cold conversation resume and arbitrary command execution are unsupported.

• Add opt-in Vercel and Daytona dependency download caches backed by a caller-owned Transport: restore before preparation, archive before disposal, retain compatibility hashing and fence concurrent publishers. Add bilingual Guide/API documentation and a cloud example. Deterministic SDK and filesystem fixtures cover binary transfers, permissions, concurrency and cleanup failures; live cloud/S3 validation remains pending.

• Add incremental lifecycle preparation with when: changed(files): fingerprint exact file contents in the hook environment and rerun conditional hooks before operations on a reused sandbox. Save fingerprints only after success and retry failed preparations without closing a warm sandbox. Add bilingual Guide/API documentation and an offline npm ci example. Deterministic local-process and simulated-remote tests cover reuse, failures, cancellation and deadlines; real Docker checks the live sandbox files. Live cloud validation remains pending.

• Add opt-in dispatch repetition watchdogs over decoded tool calls and file changes, with stuck events, explicit stop/warn policies and bounded steering instructions. Preserve sandbox reuse, recovery, response instructions and synchronous policy enforcement independently of observer delivery. Add bilingual Guide/API documentation and an offline example; deterministic process, simulated-model and real Docker tests cover behavior. Live native CLI and cloud validation remains pending.

• Add portable defineAgentProfile() declarations for literal instructions, built-in tool allowlists and MCP servers across CLI presets and the Outpost harness. Project native Claude/Codex instructions, preserve request-prefix guidance on other CLIs, enforce exact shell commands through Claude hooks and intersected harness permissions, and refuse unsupported allowlists explicitly. Add bilingual Guide/API documentation and an offline local/MCP example. Deterministic protocol, command-hook and local-model tests cover behavior; paid CLI/model runs and native CLI continuation remain unvalidated.

• Fix lock-free local transport reads when an atomic replacement grows between size sampling and inspection. Retry changed snapshots while preserving caller byte limits and bounded allocation; a deterministic race regression covers the failure.

• Add synchronous result.report() to cold and warm dispatch results, with Markdown and versioned JSON snapshots of committed Git changes, observed failed commands/tools, duration, tokens and optional estimated cost. Preserve reports after workspace cleanup, aggregate cold passes as a net diff, apply inherited redaction and report collection limits without changing execution outcomes. Add bilingual Guide/API documentation and an offline local example; deterministic Git, process and simulated-model tests cover behavior, with no paid model calls.

• Add transport-backed execution projections by ID with createRunObserver(), readRun() and cursor-based watchRun(). Record tasks, dispatch agents, branches, commits, errors and cumulative workflow usage; detect suspected abandonment through heartbeats without changing ownership. Fence writers, preserve cursors across settled workflow resumes and expose detected observation gaps. Deterministic local-process and HTTP S3-fixture tests cover behavior; live AWS validation remains pending.

• Add optional EUR/USD workflow cost budgets, per-model token pricing, explicit Models.dev/OpenRouter catalog loading and inherited observation/conversation redaction. Preserve cumulative model counters through checkpoints; stop strict monetary budgets on unknown charges. Deterministic mock-model, local-process and HTTP-fixture tests cover behavior; public Models.dev and OpenRouter catalog loading was checked; paid-model billing and native CLI resume with masked reasoning remain unvalidated.

• Add experimental best-of-N speculation with select: "best" and a required finite score. Let admitted candidates finish, choose the highest score with declaration-order ties, preserve cancellation, budgets and cleanup ownership, and persist scores for durable recovery. Keep first-valid selection as the default. Deterministic local-agent and crash-recovery tests cover the behavior; live judge-agent and cloud validation remains pending.

• Add opt-in agent conflict resolution through workspace.integrate({ onConflict: createAgentConflictResolver(...) }). Resolve in a separate branch and sandbox, require verification, recheck final diff guards and frozen input commits, and preserve both workspaces on failure. Return resolution usage and verification output separately. Deterministic tests use real temporary Git repositories, a simulated remote provider and real Docker in mounted and isolated modes; live agent and cloud validation remains pending.

• Add workspace diff guards (guard.protectedPaths, guard.maxChangedLines) for named and integration branches, independently of the agent. Check the final committed diff after synchronization and again before merging the inspected commit; reject protected paths, oversized text changes, uncountable binaries with a line limit and incomplete inspection with code guard. Retain rejected branches and worktrees; refuse current before execution. Deterministic tests use real temporary Git repositories, local agents and a simulated remote provider; no paid model calls are required.

• Make the simulated Daytona dependency-cache sandbox provide its own Linux setsid --wait wrapper so its host-process fixtures also run on macOS. Production sandbox process supervision is unchanged.

• Freeze the queue clock in the scheduler deduplication fixture so slow Windows runners cannot expire the first claim between assertions. Queue lease expiry remains covered separately.

The v11.0.0 release was blocked by the macOS dependency-cache fixture and a timing-sensitive Windows scheduler fixture and was not published.

10.0.1

• Breaking: Report gate refusals as workflow status rejected. Add typed terminationCode to workflow results, finish events and queued workflow summaries, and expose it on WorkflowFailure.code; preserve classified faults, deadlines, budget stops and checkpointed rejection audit trails. Consumers checking only failed must also handle rejected. Deterministic regression tests cover the termination semantics; no live agent or provider calls are required.
• Accept explicit repositoryMode: "isolated" on Vercel and Daytona, matching their existing private-checkout default. Reject unsupported modes before allocation. Deterministic provider fixtures verify Git metadata separation, commit/file synchronization and recovery after concurrent host edits; live cloud validation of this explicit option remains outstanding.
• Automatically append tagged final-answer instructions and input JSON Schema for typed responses, including repairs and resumed turns. Breaking: defineJsonResponse() now requires Standard JSON Schema conversion or an explicit jsonSchema; parsing functions and validation-only schemas must declare their input schema. Existing parsing and validation behavior is preserved.
• Update the bilingual guides, API reference and offline examples. Deterministic tests do not establish live agent compliance with the injected instructions.
• Pin the transitive shell-quote dependency to 1.11.0 and documentation sharp to 0.35.5 to resolve the dependency audits blocking CI.
• Resume quota-interrupted workflow fixtures explicitly from durable checkpoints, preserving conversation continuation and fallback checks without relying on process startup completing before a simulated reset.

The v10.0.0 release was blocked by a timing-sensitive Windows fixture and was not published.

9.1.1

• Add typed System One decisions (defineDecision, decide, defineDecisionTask) and one bounded HTTP provider for Jev and compatible Laya endpoints. Preserve native distributions, metadata and input-truncation diagnostics.
• Add per-step model routing to the built-in harness, including subagents, explicit confidence and outage fallback policies, cumulative routing usage and version-2 transcripts with version-1 compatibility.
• Document decis...

Outpost v10.0.1

Changelog

Unreleased

10.0.1

• Breaking: Report gate refusals as workflow status rejected. Add typed terminationCode to workflow results, finish events and queued workflow summaries, and expose it on WorkflowFailure.code; preserve classified faults, deadlines, budget stops and checkpointed rejection audit trails. Consumers checking only failed must also handle rejected. Deterministic regression tests cover the termination semantics; no live agent or provider calls are required.
• Accept explicit repositoryMode: "isolated" on Vercel and Daytona, matching their existing private-checkout default. Reject unsupported modes before allocation. Deterministic provider fixtures verify Git metadata separation, commit/file synchronization and recovery after concurrent host edits; live cloud validation of this explicit option remains outstanding.
• Automatically append tagged final-answer instructions and input JSON Schema for typed responses, including repairs and resumed turns. Breaking: defineJsonResponse() now requires Standard JSON Schema conversion or an explicit jsonSchema; parsing functions and validation-only schemas must declare their input schema. Existing parsing and validation behavior is preserved.
• Update the bilingual guides, API reference and offline examples. Deterministic tests do not establish live agent compliance with the injected instructions.
• Pin the transitive shell-quote dependency to 1.11.0 and documentation sharp to 0.35.5 to resolve the dependency audits blocking CI.
• Resume quota-interrupted workflow fixtures explicitly from durable checkpoints, preserving conversation continuation and fallback checks without relying on process startup completing before a simulated reset.

The v10.0.0 release was blocked by a timing-sensitive Windows fixture and was not published.

9.1.1

• Add typed System One decisions (defineDecision, decide, defineDecisionTask) and one bounded HTTP provider for Jev and compatible Laya endpoints. Preserve native distributions, metadata and input-truncation diagnostics.
• Add per-step model routing to the built-in harness, including subagents, explicit confidence and outage fallback policies, cumulative routing usage and version-2 transcripts with version-1 compatibility.
• Document decisions and routing in English and French, with public reference contracts and deterministic HTTP, workflow and conversation regression tests.
• Accept accumulated four-decimal rounding in native decision distributions and weighted scores without renormalizing them; continue rejecting incoherent responses.
• Add the runnable System One example with direct decisions and a typed workflow against a local Laya server or an authenticated compatible endpoint.
• Announce System One support in a bilingual indigo banner. Use a square logo button, equally sized Guide/API buttons with icons, colored source-link hovers and search aligned with the content border on wide screens.
• Open /reference/ at the first alphabetical API symbol. Remove family overview pages and redirect their old URLs to the corresponding practical Guide pages.
• Retire the repository roadmap and preserve its removal in the documentation migration audit. Keep validation limits in the guides and release notes.
• Pin the documentation's transitive source-map-js dependency to the patched 1.2.2 version.

The v9.1.0 release was blocked by the documentation dependency audit and was not published.

Direct decisions and decision tasks were exercised against a local Laya 0.3.28 CPU server. Live Jev inference and conversational model routing remain unvalidated; the ordinary test suites use deterministic local fixtures.

9.0.8

• Redesign the bilingual documentation home with a full-viewport hero, a four-step journey and a nine-capability grid linking to the Guide. Show a medium-sized CSV-export workflow with parallel tests and review, bounded correction rounds and human approval.
• Add a site favicon, localized social previews, homepage search metadata and structured data, and a robots endpoint.
• Replace direct console output in Guide, homepage and README snippets with reporters and sample-output comments. Typecheck the shared reporter and execute the offline examples in both languages.
• Hide code-tab and snippet scrollbars while preserving horizontal scrolling, mouse dragging and keyboard navigation.
• Include the shared reporter when executing real container documentation examples in both languages.

Public library contracts and runtime behavior are unchanged.

9.0.6

• Rewrite the English and French Guide with clearer titles and explanations. The recommended setup builds an agent image, then imports Outpost in user-written TypeScript files; generated init workflow projects remain optional. Link option and property definitions to their API contracts instead of repeating them in the Guide.
• Limit Guide code blocks to 20 lines, split longer examples into named .ts files in tabs and place their explanations beside them. Code tabs support horizontal scrolling and mouse dragging so every file stays reachable.
• Replace Guide flows with pannable, zoomable canvases, illustrate cards with icons, extend notices across the content section and give page navigation actions equal widths. Reduce spacing in the sidebar.
• Rebuild the bilingual home around what developers can accomplish with Outpost, with a direct starting point, short working examples, illustrated use cases and a canvas. Reuse the Guide's layout and components, and regenerate content when starting or building the site so development and production render the same page.
• Keep only Guide and API in the header. Open the changelog exclusively from the version button, including on mobile, and keep the roadmap in the repository's roadmap.md. The version button's hover fills its whole cell, and Guide/API buttons keep their size while fonts load and the active page changes.

Public library contracts and runtime behavior are unchanged.

9.0.5

• Add a development-workflow use case that takes a ticket to a reviewed branch: an agent questions the owner and plans, the owner approves the plan, then agents write failing tests and the code that makes them pass under a reviewer agent, while the workflow checks every step and makes every commit. It replaces the “Write a specification with a human” recipe, whose URL redirects to it.
• Open four Guide chapters with an overview page — isolation, typed workflows, unattended runs, and observe and recover — that maps the chapter's choices before its detailed pages. Flows drawn in Guide pages can now be panned and zoomed.
• Label the Reference space “API” in the header, sidebar and breadcrumbs in both languages; its URLs stay under /reference/. The sidebar lists every public symbol alphabetically with its kind icon, and the family map with each overview moves to the reference home.
• Show the landing page's install command for each package manager in tabs. Public library contracts are unchanged.

9.0.4

• Make the landing page's typed-response brief ask for the shape it expects. The brief now carries <verdict>{"approved": true, "reasons": []}</verdict>, so the schema, the request and the agent's tagged reply agree: Outpost sends a brief unchanged and adds no format instructions, so a brief that does not show the shape is one the agent cannot satisfy. A comment beside repairs says what the option buys — two more turns to fix an answer the schema rejects, on the same conversation.

9.0.3

• Rebuild the documentation landing page as a path for a developer deciding whether to install: the headline runs on one rule, the install command carries its prerequisites and the first command to run, and a dispatch() call cut from a typechecked snippet shows one task end to end. The page then reads as what you can run, the comparison drawing, the agents and sandboxes it runs on, the argument that drawing opened, typed responses, and where Outpost is not the answer.
• Make the orchestration comparison argue in numbers. Two lanes, “The model orchestrates” and “Your code orchestrates”, play four beats; each step reports its own time — 0.6 s for a model step, 0.9 s for an agent turn, 0.1 s for code, read from one constants module so no figure can disagree with another — and the running total turns the Resume beat into a number: the model climbs from 2.4 s to 3.6 s because it rereads two steps, while the code lane holds at 1.2 s because restored steps run nothing. The model’s context meter turns ember past half and red past three quarters while the code lane stays at a quarter. The chain loops, carries on from a clicked beat, parks on pointer or focus, keeps the explicit pause toggle and shows a still state when motion is reduced.
• Drive the problem register from that figure instead of repeating it: one pain shows at a time beside the beat that makes its case, with a four-tick pager that fills over the beat’s own duration, a visually hidden list of all four for assistive technology and search, and the first pain served as static prose when scripts do not run.
• Show the whole typed-response call rather than its schema alone: the Zod declaration, the dispatch() that passes it as response and the typed value read back, cut at build time from a typechecked snippet, beside the agent’s reply in prose with its tagged block lit at the moment it becomes typed data.
• Rewrite the landing copy in both languages so each reads as written rather than translated, without moving a claim, a proof or a link.
• Stop flagging Firecracker as experimental on the landing page, where the reference already ships it stable; the note there now covers speculation only. Public library contracts are unchanged.

9.0.2

• Rebuild the documentation home page around “Code orchestrates. Agents think.”: an animated comparison of a pipeline orchestrated by an LLM and one orchestrated by code, the pains of LLM-driven pipe...

Outpost v9.1.1

Changelog

9.1.1

• Add typed System One decisions (defineDecision, decide, defineDecisionTask) and one bounded HTTP provider for Jev and compatible Laya endpoints. Preserve native distributions, metadata and input-truncation diagnostics.
• Add per-step model routing to the built-in harness, including subagents, explicit confidence and outage fallback policies, cumulative routing usage and version-2 transcripts with version-1 compatibility.
• Document decisions and routing in English and French, with public reference contracts and deterministic HTTP, workflow and conversation regression tests.
• Accept accumulated four-decimal rounding in native decision distributions and weighted scores without renormalizing them; continue rejecting incoherent responses.
• Add the runnable System One example with direct decisions and a typed workflow against a local Laya server or an authenticated compatible endpoint.
• Announce System One support in a bilingual indigo banner. Use a square logo button, equally sized Guide/API buttons with icons, colored source-link hovers and search aligned with the content border on wide screens.
• Open /reference/ at the first alphabetical API symbol. Remove family overview pages and redirect their old URLs to the corresponding practical Guide pages.
• Retire the repository roadmap and preserve its removal in the documentation migration audit. Keep validation limits in the guides and release notes.
• Pin the documentation's transitive source-map-js dependency to the patched 1.2.2 version.

The v9.1.0 release was blocked by the documentation dependency audit and was not published.

Direct decisions and decision tasks were exercised against a local Laya 0.3.28 CPU server. Live Jev inference and conversational model routing remain unvalidated; the ordinary test suites use deterministic local fixtures.

9.0.8

• Redesign the bilingual documentation home with a full-viewport hero, a four-step journey and a nine-capability grid linking to the Guide. Show a medium-sized CSV-export workflow with parallel tests and review, bounded correction rounds and human approval.
• Add a site favicon, localized social previews, homepage search metadata and structured data, and a robots endpoint.
• Replace direct console output in Guide, homepage and README snippets with reporters and sample-output comments. Typecheck the shared reporter and execute the offline examples in both languages.
• Hide code-tab and snippet scrollbars while preserving horizontal scrolling, mouse dragging and keyboard navigation.
• Include the shared reporter when executing real container documentation examples in both languages.

Public library contracts and runtime behavior are unchanged.

9.0.6

• Rewrite the English and French Guide with clearer titles and explanations. The recommended setup builds an agent image, then imports Outpost in user-written TypeScript files; generated init workflow projects remain optional. Link option and property definitions to their API contracts instead of repeating them in the Guide.
• Limit Guide code blocks to 20 lines, split longer examples into named .ts files in tabs and place their explanations beside them. Code tabs support horizontal scrolling and mouse dragging so every file stays reachable.
• Replace Guide flows with pannable, zoomable canvases, illustrate cards with icons, extend notices across the content section and give page navigation actions equal widths. Reduce spacing in the sidebar.
• Rebuild the bilingual home around what developers can accomplish with Outpost, with a direct starting point, short working examples, illustrated use cases and a canvas. Reuse the Guide's layout and components, and regenerate content when starting or building the site so development and production render the same page.
• Keep only Guide and API in the header. Open the changelog exclusively from the version button, including on mobile, and keep the roadmap in the repository's roadmap.md. The version button's hover fills its whole cell, and Guide/API buttons keep their size while fonts load and the active page changes.

Public library contracts and runtime behavior are unchanged.

9.0.5

• Add a development-workflow use case that takes a ticket to a reviewed branch: an agent questions the owner and plans, the owner approves the plan, then agents write failing tests and the code that makes them pass under a reviewer agent, while the workflow checks every step and makes every commit. It replaces the “Write a specification with a human” recipe, whose URL redirects to it.
• Open four Guide chapters with an overview page — isolation, typed workflows, unattended runs, and observe and recover — that maps the chapter's choices before its detailed pages. Flows drawn in Guide pages can now be panned and zoomed.
• Label the Reference space “API” in the header, sidebar and breadcrumbs in both languages; its URLs stay under /reference/. The sidebar lists every public symbol alphabetically with its kind icon, and the family map with each overview moves to the reference home.
• Show the landing page's install command for each package manager in tabs. Public library contracts are unchanged.

9.0.4

• Make the landing page's typed-response brief ask for the shape it expects. The brief now carries <verdict>{"approved": true, "reasons": []}</verdict>, so the schema, the request and the agent's tagged reply agree: Outpost sends a brief unchanged and adds no format instructions, so a brief that does not show the shape is one the agent cannot satisfy. A comment beside repairs says what the option buys — two more turns to fix an answer the schema rejects, on the same conversation.

9.0.3

• Rebuild the documentation landing page as a path for a developer deciding whether to install: the headline runs on one rule, the install command carries its prerequisites and the first command to run, and a dispatch() call cut from a typechecked snippet shows one task end to end. The page then reads as what you can run, the comparison drawing, the agents and sandboxes it runs on, the argument that drawing opened, typed responses, and where Outpost is not the answer.
• Make the orchestration comparison argue in numbers. Two lanes, “The model orchestrates” and “Your code orchestrates”, play four beats; each step reports its own time — 0.6 s for a model step, 0.9 s for an agent turn, 0.1 s for code, read from one constants module so no figure can disagree with another — and the running total turns the Resume beat into a number: the model climbs from 2.4 s to 3.6 s because it rereads two steps, while the code lane holds at 1.2 s because restored steps run nothing. The model’s context meter turns ember past half and red past three quarters while the code lane stays at a quarter. The chain loops, carries on from a clicked beat, parks on pointer or focus, keeps the explicit pause toggle and shows a still state when motion is reduced.
• Drive the problem register from that figure instead of repeating it: one pain shows at a time beside the beat that makes its case, with a four-tick pager that fills over the beat’s own duration, a visually hidden list of all four for assistive technology and search, and the first pain served as static prose when scripts do not run.
• Show the whole typed-response call rather than its schema alone: the Zod declaration, the dispatch() that passes it as response and the typed value read back, cut at build time from a typechecked snippet, beside the agent’s reply in prose with its tagged block lit at the moment it becomes typed data.
• Rewrite the landing copy in both languages so each reads as written rather than translated, without moving a claim, a proof or a link.
• Stop flagging Firecracker as experimental on the landing page, where the reference already ships it stable; the note there now covers speculation only. Public library contracts are unchanged.

9.0.2

• Rebuild the documentation home page around “Code orchestrates. Agents think.”: an animated comparison of a pipeline orchestrated by an LLM and one orchestrated by code, the pains of LLM-driven pipelines paired with what Outpost does instead, a table showing that only the agent’s work varies between runs of the same workflow beside the real verification check that gates it, and a compact list of agents and sandboxes. Both figures can be paused or replayed, show still states when motion is reduced, and are described for screen readers. Public library contracts are unchanged.

9.0.1

• Rewrite the English and French Guide by reader task: twelve chapters in reading order, six complete use-case recipes, and pages drawn with feature cells, reading paths, flows and file trees. Moved Guide URLs redirect to their new pages.
• Rewrite the Reference: each family overview picks between its entry points with decision and outcome tables, and every symbol and field description was checked against the source and corrected where it was wrong, vague or outdated. Each symbol links to its own Guide page.
• List the values of 35 type aliases such as FaultCode, TaskStatus and SteeringMode, keep placeholders such as <name> visible, and print types from other modules without import("…") qualifiers. Documentation sync now rejects descriptions that no page uses.
• Mark the speculation API experimental in the Reference, as the 9.0.0 notes stated. Add numbered runnable examples under examples/. Public library contracts are unchanged.

9.0.0

This major release renames public factories and declarations and removes the previous names, changes conversation and reporter contracts, and extends public event, fault and stop-reason unions. Replace previous names with their create*/define* equivalents, pass a store instead of a format name to createTransportConversations(), use each agent's native store instead of the format-keyed conversations helpers, and update exhaustive handling of AgentEvent (fallback, `s...

Outpost v9.0.8

Changelog

9.0.8

• Redesign the bilingual documentation home with a full-viewport hero, a four-step journey and a nine-capability grid linking to the Guide. Show a medium-sized CSV-export workflow with parallel tests and review, bounded correction rounds and human approval.
• Add a site favicon, localized social previews, homepage search metadata and structured data, and a robots endpoint.
• Replace direct console output in Guide, homepage and README snippets with reporters and sample-output comments. Typecheck the shared reporter and execute the offline examples in both languages.
• Hide code-tab and snippet scrollbars while preserving horizontal scrolling, mouse dragging and keyboard navigation.
• Include the shared reporter when executing real container documentation examples in both languages.

Public library contracts and runtime behavior are unchanged.

9.0.6

• Rewrite the English and French Guide with clearer titles and explanations. The recommended setup builds an agent image, then imports Outpost in user-written TypeScript files; generated init workflow projects remain optional. Link option and property definitions to their API contracts instead of repeating them in the Guide.
• Limit Guide code blocks to 20 lines, split longer examples into named .ts files in tabs and place their explanations beside them. Code tabs support horizontal scrolling and mouse dragging so every file stays reachable.
• Replace Guide flows with pannable, zoomable canvases, illustrate cards with icons, extend notices across the content section and give page navigation actions equal widths. Reduce spacing in the sidebar.
• Rebuild the bilingual home around what developers can accomplish with Outpost, with a direct starting point, short working examples, illustrated use cases and a canvas. Reuse the Guide's layout and components, and regenerate content when starting or building the site so development and production render the same page.
• Keep only Guide and API in the header. Open the changelog exclusively from the version button, including on mobile, and keep the roadmap in the repository's roadmap.md. The version button's hover fills its whole cell, and Guide/API buttons keep their size while fonts load and the active page changes.

Public library contracts and runtime behavior are unchanged.

9.0.5

• Add a development-workflow use case that takes a ticket to a reviewed branch: an agent questions the owner and plans, the owner approves the plan, then agents write failing tests and the code that makes them pass under a reviewer agent, while the workflow checks every step and makes every commit. It replaces the “Write a specification with a human” recipe, whose URL redirects to it.
• Open four Guide chapters with an overview page — isolation, typed workflows, unattended runs, and observe and recover — that maps the chapter's choices before its detailed pages. Flows drawn in Guide pages can now be panned and zoomed.
• Label the Reference space “API” in the header, sidebar and breadcrumbs in both languages; its URLs stay under /reference/. The sidebar lists every public symbol alphabetically with its kind icon, and the family map with each overview moves to the reference home.
• Show the landing page's install command for each package manager in tabs. Public library contracts are unchanged.

9.0.4

• Make the landing page's typed-response brief ask for the shape it expects. The brief now carries <verdict>{"approved": true, "reasons": []}</verdict>, so the schema, the request and the agent's tagged reply agree: Outpost sends a brief unchanged and adds no format instructions, so a brief that does not show the shape is one the agent cannot satisfy. A comment beside repairs says what the option buys — two more turns to fix an answer the schema rejects, on the same conversation.

9.0.3

• Rebuild the documentation landing page as a path for a developer deciding whether to install: the headline runs on one rule, the install command carries its prerequisites and the first command to run, and a dispatch() call cut from a typechecked snippet shows one task end to end. The page then reads as what you can run, the comparison drawing, the agents and sandboxes it runs on, the argument that drawing opened, typed responses, and where Outpost is not the answer.
• Make the orchestration comparison argue in numbers. Two lanes, “The model orchestrates” and “Your code orchestrates”, play four beats; each step reports its own time — 0.6 s for a model step, 0.9 s for an agent turn, 0.1 s for code, read from one constants module so no figure can disagree with another — and the running total turns the Resume beat into a number: the model climbs from 2.4 s to 3.6 s because it rereads two steps, while the code lane holds at 1.2 s because restored steps run nothing. The model’s context meter turns ember past half and red past three quarters while the code lane stays at a quarter. The chain loops, carries on from a clicked beat, parks on pointer or focus, keeps the explicit pause toggle and shows a still state when motion is reduced.
• Drive the problem register from that figure instead of repeating it: one pain shows at a time beside the beat that makes its case, with a four-tick pager that fills over the beat’s own duration, a visually hidden list of all four for assistive technology and search, and the first pain served as static prose when scripts do not run.
• Show the whole typed-response call rather than its schema alone: the Zod declaration, the dispatch() that passes it as response and the typed value read back, cut at build time from a typechecked snippet, beside the agent’s reply in prose with its tagged block lit at the moment it becomes typed data.
• Rewrite the landing copy in both languages so each reads as written rather than translated, without moving a claim, a proof or a link.
• Stop flagging Firecracker as experimental on the landing page, where the reference already ships it stable; the note there now covers speculation only. Public library contracts are unchanged.

9.0.2

• Rebuild the documentation home page around “Code orchestrates. Agents think.”: an animated comparison of a pipeline orchestrated by an LLM and one orchestrated by code, the pains of LLM-driven pipelines paired with what Outpost does instead, a table showing that only the agent’s work varies between runs of the same workflow beside the real verification check that gates it, and a compact list of agents and sandboxes. Both figures can be paused or replayed, show still states when motion is reduced, and are described for screen readers. Public library contracts are unchanged.

9.0.1

• Rewrite the English and French Guide by reader task: twelve chapters in reading order, six complete use-case recipes, and pages drawn with feature cells, reading paths, flows and file trees. Moved Guide URLs redirect to their new pages.
• Rewrite the Reference: each family overview picks between its entry points with decision and outcome tables, and every symbol and field description was checked against the source and corrected where it was wrong, vague or outdated. Each symbol links to its own Guide page.
• List the values of 35 type aliases such as FaultCode, TaskStatus and SteeringMode, keep placeholders such as <name> visible, and print types from other modules without import("…") qualifiers. Documentation sync now rejects descriptions that no page uses.
• Mark the speculation API experimental in the Reference, as the 9.0.0 notes stated. Add numbered runnable examples under examples/. Public library contracts are unchanged.

9.0.0

This major release renames public factories and declarations and removes the previous names, changes conversation and reporter contracts, and extends public event, fault and stop-reason unions. Replace previous names with their create*/define* equivalents, pass a store instead of a format name to createTransportConversations(), use each agent's native store instead of the format-keyed conversations helpers, and update exhaustive handling of AgentEvent (fallback, steer), FaultCode (steering) and stop reasons (steered). Firecracker is no longer experimental; speculation remains experimental. Live validation campaigns listed in the roadmap remain outstanding.

• Promote the Firecracker provider out of experimental: createFirecrackerSandboxProvider() and FirecrackerOptions no longer carry the experimental warning and icon. Host preparation, networking and the root supervisor of the jailer mode remain the operator's responsibility; control-plane resource accounting and adversarial validation remain open.
• Redraw the documentation navigation bar on the page grid: the logo cell lines up with the sidebar, search fills the centre, GitLab, the GitHub mirror and npm are linked beside the current version, and a single link switches between English and French. Narrow screens move the source links into the sidebar and menu footer.
• Give each built-in CLI agent its own src/adapters/agents/<agent>/ folder and a descriptor in one agent catalog: its label, executable, pinned version, harness export, installer, image environment, doctor diagnostics, protocol fixtures and outpost init authentication choices. agentVersions, DoctorAgent, remote bootstrap, outpost doctor, the outpost init prompts and validation, the generated run.ts, the image recipe and the agent image lock check now derive from it, and a test checks every descriptor against the public exports and the image lock. Generated projects, images and bootstrap commands are unchanged. agentVersions entries are now typed as string, and Antigravity model errors use the Antigravity label.
• Open native conversation formats to any CLI harness. createTranscriptConversations(layout) builds a native store for CLIs that keep one JSONL transcript per conversation, and createSessionBundleConversations(profile) for CL...

Outpost v9.0.6

Changelog

9.0.6

• Rewrite the English and French Guide with clearer titles and explanations. The recommended setup builds an agent image, then imports Outpost in user-written TypeScript files; generated init workflow projects remain optional. Link option and property definitions to their API contracts instead of repeating them in the Guide.
• Limit Guide code blocks to 20 lines, split longer examples into named .ts files in tabs and place their explanations beside them. Code tabs support horizontal scrolling and mouse dragging so every file stays reachable.
• Replace Guide flows with pannable, zoomable canvases, illustrate cards with icons, extend notices across the content section and give page navigation actions equal widths. Reduce spacing in the sidebar.
• Rebuild the bilingual home around what developers can accomplish with Outpost, with a direct starting point, short working examples, illustrated use cases and a canvas. Reuse the Guide's layout and components, and regenerate content when starting or building the site so development and production render the same page.
• Keep only Guide and API in the header. Open the changelog exclusively from the version button, including on mobile, and keep the roadmap in the repository's roadmap.md. The version button's hover fills its whole cell, and Guide/API buttons keep their size while fonts load and the active page changes.

Public library contracts and runtime behavior are unchanged.

9.0.5

• Add a development-workflow use case that takes a ticket to a reviewed branch: an agent questions the owner and plans, the owner approves the plan, then agents write failing tests and the code that makes them pass under a reviewer agent, while the workflow checks every step and makes every commit. It replaces the “Write a specification with a human” recipe, whose URL redirects to it.
• Open four Guide chapters with an overview page — isolation, typed workflows, unattended runs, and observe and recover — that maps the chapter's choices before its detailed pages. Flows drawn in Guide pages can now be panned and zoomed.
• Label the Reference space “API” in the header, sidebar and breadcrumbs in both languages; its URLs stay under /reference/. The sidebar lists every public symbol alphabetically with its kind icon, and the family map with each overview moves to the reference home.
• Show the landing page's install command for each package manager in tabs. Public library contracts are unchanged.

9.0.4

• Make the landing page's typed-response brief ask for the shape it expects. The brief now carries <verdict>{"approved": true, "reasons": []}</verdict>, so the schema, the request and the agent's tagged reply agree: Outpost sends a brief unchanged and adds no format instructions, so a brief that does not show the shape is one the agent cannot satisfy. A comment beside repairs says what the option buys — two more turns to fix an answer the schema rejects, on the same conversation.

9.0.3

• Rebuild the documentation landing page as a path for a developer deciding whether to install: the headline runs on one rule, the install command carries its prerequisites and the first command to run, and a dispatch() call cut from a typechecked snippet shows one task end to end. The page then reads as what you can run, the comparison drawing, the agents and sandboxes it runs on, the argument that drawing opened, typed responses, and where Outpost is not the answer.
• Make the orchestration comparison argue in numbers. Two lanes, “The model orchestrates” and “Your code orchestrates”, play four beats; each step reports its own time — 0.6 s for a model step, 0.9 s for an agent turn, 0.1 s for code, read from one constants module so no figure can disagree with another — and the running total turns the Resume beat into a number: the model climbs from 2.4 s to 3.6 s because it rereads two steps, while the code lane holds at 1.2 s because restored steps run nothing. The model’s context meter turns ember past half and red past three quarters while the code lane stays at a quarter. The chain loops, carries on from a clicked beat, parks on pointer or focus, keeps the explicit pause toggle and shows a still state when motion is reduced.
• Drive the problem register from that figure instead of repeating it: one pain shows at a time beside the beat that makes its case, with a four-tick pager that fills over the beat’s own duration, a visually hidden list of all four for assistive technology and search, and the first pain served as static prose when scripts do not run.
• Show the whole typed-response call rather than its schema alone: the Zod declaration, the dispatch() that passes it as response and the typed value read back, cut at build time from a typechecked snippet, beside the agent’s reply in prose with its tagged block lit at the moment it becomes typed data.
• Rewrite the landing copy in both languages so each reads as written rather than translated, without moving a claim, a proof or a link.
• Stop flagging Firecracker as experimental on the landing page, where the reference already ships it stable; the note there now covers speculation only. Public library contracts are unchanged.

9.0.2

• Rebuild the documentation home page around “Code orchestrates. Agents think.”: an animated comparison of a pipeline orchestrated by an LLM and one orchestrated by code, the pains of LLM-driven pipelines paired with what Outpost does instead, a table showing that only the agent’s work varies between runs of the same workflow beside the real verification check that gates it, and a compact list of agents and sandboxes. Both figures can be paused or replayed, show still states when motion is reduced, and are described for screen readers. Public library contracts are unchanged.

9.0.1

• Rewrite the English and French Guide by reader task: twelve chapters in reading order, six complete use-case recipes, and pages drawn with feature cells, reading paths, flows and file trees. Moved Guide URLs redirect to their new pages.
• Rewrite the Reference: each family overview picks between its entry points with decision and outcome tables, and every symbol and field description was checked against the source and corrected where it was wrong, vague or outdated. Each symbol links to its own Guide page.
• List the values of 35 type aliases such as FaultCode, TaskStatus and SteeringMode, keep placeholders such as <name> visible, and print types from other modules without import("…") qualifiers. Documentation sync now rejects descriptions that no page uses.
• Mark the speculation API experimental in the Reference, as the 9.0.0 notes stated. Add numbered runnable examples under examples/. Public library contracts are unchanged.

9.0.0

This major release renames public factories and declarations and removes the previous names, changes conversation and reporter contracts, and extends public event, fault and stop-reason unions. Replace previous names with their create*/define* equivalents, pass a store instead of a format name to createTransportConversations(), use each agent's native store instead of the format-keyed conversations helpers, and update exhaustive handling of AgentEvent (fallback, steer), FaultCode (steering) and stop reasons (steered). Firecracker is no longer experimental; speculation remains experimental. Live validation campaigns listed in the roadmap remain outstanding.

• Promote the Firecracker provider out of experimental: createFirecrackerSandboxProvider() and FirecrackerOptions no longer carry the experimental warning and icon. Host preparation, networking and the root supervisor of the jailer mode remain the operator's responsibility; control-plane resource accounting and adversarial validation remain open.
• Redraw the documentation navigation bar on the page grid: the logo cell lines up with the sidebar, search fills the centre, GitLab, the GitHub mirror and npm are linked beside the current version, and a single link switches between English and French. Narrow screens move the source links into the sidebar and menu footer.
• Give each built-in CLI agent its own src/adapters/agents/<agent>/ folder and a descriptor in one agent catalog: its label, executable, pinned version, harness export, installer, image environment, doctor diagnostics, protocol fixtures and outpost init authentication choices. agentVersions, DoctorAgent, remote bootstrap, outpost doctor, the outpost init prompts and validation, the generated run.ts, the image recipe and the agent image lock check now derive from it, and a test checks every descriptor against the public exports and the image lock. Generated projects, images and bootstrap commands are unchanged. agentVersions entries are now typed as string, and Antigravity model errors use the Antigravity label.
• Open native conversation formats to any CLI harness. createTranscriptConversations(layout) builds a native store for CLIs that keep one JSONL transcript per conversation, and createSessionBundleConversations(profile) for CLIs that keep each session as a directory, with self-contained validate, bucket and relocate hooks that run inside the sandbox. createClaudeConversations(), createCodexConversations(), createCopilotConversations() and createKimiConversations() return the built-in stores, which implement the new NativeConversationStore contract with directory() and destination(). Built-in presets now set storage to their native store by default, and AgentAdapter.conversations is removed. Breaking: createTransportConversations() takes a base store instead of a format name, and the format-keyed helpers conversations.native, locate, capture, restore, directory, destination and claudePath are removed; use each agent's native store. ConversationFormat is now string and StoredConversationFormat is removed. Persisted format names, tr...

Outpost v9.0.5

Changelog

9.0.5

• Add a development-workflow use case that takes a ticket to a reviewed branch: an agent questions the owner and plans, the owner approves the plan, then agents write failing tests and the code that makes them pass under a reviewer agent, while the workflow checks every step and makes every commit. It replaces the “Write a specification with a human” recipe, whose URL redirects to it.
• Open four Guide chapters with an overview page — isolation, typed workflows, unattended runs, and observe and recover — that maps the chapter's choices before its detailed pages. Flows drawn in Guide pages can now be panned and zoomed.
• Label the Reference space “API” in the header, sidebar and breadcrumbs in both languages; its URLs stay under /reference/. The sidebar lists every public symbol alphabetically with its kind icon, and the family map with each overview moves to the reference home.
• Show the landing page's install command for each package manager in tabs. Public library contracts are unchanged.

9.0.4

• Make the landing page's typed-response brief ask for the shape it expects. The brief now carries <verdict>{"approved": true, "reasons": []}</verdict>, so the schema, the request and the agent's tagged reply agree: Outpost sends a brief unchanged and adds no format instructions, so a brief that does not show the shape is one the agent cannot satisfy. A comment beside repairs says what the option buys — two more turns to fix an answer the schema rejects, on the same conversation.

9.0.3

• Rebuild the documentation landing page as a path for a developer deciding whether to install: the headline runs on one rule, the install command carries its prerequisites and the first command to run, and a dispatch() call cut from a typechecked snippet shows one task end to end. The page then reads as what you can run, the comparison drawing, the agents and sandboxes it runs on, the argument that drawing opened, typed responses, and where Outpost is not the answer.
• Make the orchestration comparison argue in numbers. Two lanes, “The model orchestrates” and “Your code orchestrates”, play four beats; each step reports its own time — 0.6 s for a model step, 0.9 s for an agent turn, 0.1 s for code, read from one constants module so no figure can disagree with another — and the running total turns the Resume beat into a number: the model climbs from 2.4 s to 3.6 s because it rereads two steps, while the code lane holds at 1.2 s because restored steps run nothing. The model’s context meter turns ember past half and red past three quarters while the code lane stays at a quarter. The chain loops, carries on from a clicked beat, parks on pointer or focus, keeps the explicit pause toggle and shows a still state when motion is reduced.
• Drive the problem register from that figure instead of repeating it: one pain shows at a time beside the beat that makes its case, with a four-tick pager that fills over the beat’s own duration, a visually hidden list of all four for assistive technology and search, and the first pain served as static prose when scripts do not run.
• Show the whole typed-response call rather than its schema alone: the Zod declaration, the dispatch() that passes it as response and the typed value read back, cut at build time from a typechecked snippet, beside the agent’s reply in prose with its tagged block lit at the moment it becomes typed data.
• Rewrite the landing copy in both languages so each reads as written rather than translated, without moving a claim, a proof or a link.
• Stop flagging Firecracker as experimental on the landing page, where the reference already ships it stable; the note there now covers speculation only. Public library contracts are unchanged.

9.0.2

• Rebuild the documentation home page around “Code orchestrates. Agents think.”: an animated comparison of a pipeline orchestrated by an LLM and one orchestrated by code, the pains of LLM-driven pipelines paired with what Outpost does instead, a table showing that only the agent’s work varies between runs of the same workflow beside the real verification check that gates it, and a compact list of agents and sandboxes. Both figures can be paused or replayed, show still states when motion is reduced, and are described for screen readers. Public library contracts are unchanged.

9.0.1

• Rewrite the English and French Guide by reader task: twelve chapters in reading order, six complete use-case recipes, and pages drawn with feature cells, reading paths, flows and file trees. Moved Guide URLs redirect to their new pages.
• Rewrite the Reference: each family overview picks between its entry points with decision and outcome tables, and every symbol and field description was checked against the source and corrected where it was wrong, vague or outdated. Each symbol links to its own Guide page.
• List the values of 35 type aliases such as FaultCode, TaskStatus and SteeringMode, keep placeholders such as <name> visible, and print types from other modules without import("…") qualifiers. Documentation sync now rejects descriptions that no page uses.
• Mark the speculation API experimental in the Reference, as the 9.0.0 notes stated. Add numbered runnable examples under examples/. Public library contracts are unchanged.

9.0.0

This major release renames public factories and declarations and removes the previous names, changes conversation and reporter contracts, and extends public event, fault and stop-reason unions. Replace previous names with their create*/define* equivalents, pass a store instead of a format name to createTransportConversations(), use each agent's native store instead of the format-keyed conversations helpers, and update exhaustive handling of AgentEvent (fallback, steer), FaultCode (steering) and stop reasons (steered). Firecracker is no longer experimental; speculation remains experimental. Live validation campaigns listed in the roadmap remain outstanding.

• Promote the Firecracker provider out of experimental: createFirecrackerSandboxProvider() and FirecrackerOptions no longer carry the experimental warning and icon. Host preparation, networking and the root supervisor of the jailer mode remain the operator's responsibility; control-plane resource accounting and adversarial validation remain open.
• Redraw the documentation navigation bar on the page grid: the logo cell lines up with the sidebar, search fills the centre, GitLab, the GitHub mirror and npm are linked beside the current version, and a single link switches between English and French. Narrow screens move the source links into the sidebar and menu footer.
• Give each built-in CLI agent its own src/adapters/agents/<agent>/ folder and a descriptor in one agent catalog: its label, executable, pinned version, harness export, installer, image environment, doctor diagnostics, protocol fixtures and outpost init authentication choices. agentVersions, DoctorAgent, remote bootstrap, outpost doctor, the outpost init prompts and validation, the generated run.ts, the image recipe and the agent image lock check now derive from it, and a test checks every descriptor against the public exports and the image lock. Generated projects, images and bootstrap commands are unchanged. agentVersions entries are now typed as string, and Antigravity model errors use the Antigravity label.
• Open native conversation formats to any CLI harness. createTranscriptConversations(layout) builds a native store for CLIs that keep one JSONL transcript per conversation, and createSessionBundleConversations(profile) for CLIs that keep each session as a directory, with self-contained validate, bucket and relocate hooks that run inside the sandbox. createClaudeConversations(), createCodexConversations(), createCopilotConversations() and createKimiConversations() return the built-in stores, which implement the new NativeConversationStore contract with directory() and destination(). Built-in presets now set storage to their native store by default, and AgentAdapter.conversations is removed. Breaking: createTransportConversations() takes a base store instead of a format name, and the format-keyed helpers conversations.native, locate, capture, restore, directory, destination and claudePath are removed; use each agent's native store. ConversationFormat is now string and StoredConversationFormat is removed. Persisted format names, transport keys and bundle contents are unchanged.
• Rebuild the documentation layout: Guide, Reference and Project pages share one header, a sidebar for the current space and breadcrumbs, and every section places its code, contract or entries beside its prose. Reference pages list parameters and properties as rows beside a pinned signature, /reference/ becomes a map of sections and families instead of redirecting to the first symbol, and the changelog and roadmap keep each heading beside its entries.
• Replace the documentation home with a landing page in English and French: the tagline with the install command and Get started, a code window with run.ts, brief.md and an example workflow, the agents and sandboxes with their conversation support, and a workflow walkthrough that highlights the code of each step. Its snippets are typechecked with the guide examples, and the documentation brand now shows the Outpost logo and links to this page.
• Name constructors create* and declarations define*. Functions that build runtime objects become createAgent, createFallbackAgent, createReplayAgent, createHarness, createClaudeHarness, createCodexHarness, createAntigravityHarness, createCopilotHarness, createKimiHarness, createHarnessFileTools, createHarnessEditTools, createHarnessSearchTools, createHarnessGitTools, createHarnessShellTools, createDockerSandboxProvider, createPodmanSandboxProvider, `createLoc...

Outpost v9.0.4

Changelog

9.0.4

• Make the landing page's typed-response brief ask for the shape it expects. The brief now carries <verdict>{"approved": true, "reasons": []}</verdict>, so the schema, the request and the agent's tagged reply agree: Outpost sends a brief unchanged and adds no format instructions, so a brief that does not show the shape is one the agent cannot satisfy. A comment beside repairs says what the option buys — two more turns to fix an answer the schema rejects, on the same conversation.

9.0.3

• Rebuild the documentation landing page as a path for a developer deciding whether to install: the headline runs on one rule, the install command carries its prerequisites and the first command to run, and a dispatch() call cut from a typechecked snippet shows one task end to end. The page then reads as what you can run, the comparison drawing, the agents and sandboxes it runs on, the argument that drawing opened, typed responses, and where Outpost is not the answer.
• Make the orchestration comparison argue in numbers. Two lanes, “The model orchestrates” and “Your code orchestrates”, play four beats; each step reports its own time — 0.6 s for a model step, 0.9 s for an agent turn, 0.1 s for code, read from one constants module so no figure can disagree with another — and the running total turns the Resume beat into a number: the model climbs from 2.4 s to 3.6 s because it rereads two steps, while the code lane holds at 1.2 s because restored steps run nothing. The model’s context meter turns ember past half and red past three quarters while the code lane stays at a quarter. The chain loops, carries on from a clicked beat, parks on pointer or focus, keeps the explicit pause toggle and shows a still state when motion is reduced.
• Drive the problem register from that figure instead of repeating it: one pain shows at a time beside the beat that makes its case, with a four-tick pager that fills over the beat’s own duration, a visually hidden list of all four for assistive technology and search, and the first pain served as static prose when scripts do not run.
• Show the whole typed-response call rather than its schema alone: the Zod declaration, the dispatch() that passes it as response and the typed value read back, cut at build time from a typechecked snippet, beside the agent’s reply in prose with its tagged block lit at the moment it becomes typed data.
• Rewrite the landing copy in both languages so each reads as written rather than translated, without moving a claim, a proof or a link.
• Stop flagging Firecracker as experimental on the landing page, where the reference already ships it stable; the note there now covers speculation only. Public library contracts are unchanged.

9.0.2

• Rebuild the documentation home page around “Code orchestrates. Agents think.”: an animated comparison of a pipeline orchestrated by an LLM and one orchestrated by code, the pains of LLM-driven pipelines paired with what Outpost does instead, a table showing that only the agent’s work varies between runs of the same workflow beside the real verification check that gates it, and a compact list of agents and sandboxes. Both figures can be paused or replayed, show still states when motion is reduced, and are described for screen readers. Public library contracts are unchanged.

9.0.1

• Rewrite the English and French Guide by reader task: twelve chapters in reading order, six complete use-case recipes, and pages drawn with feature cells, reading paths, flows and file trees. Moved Guide URLs redirect to their new pages.
• Rewrite the Reference: each family overview picks between its entry points with decision and outcome tables, and every symbol and field description was checked against the source and corrected where it was wrong, vague or outdated. Each symbol links to its own Guide page.
• List the values of 35 type aliases such as FaultCode, TaskStatus and SteeringMode, keep placeholders such as <name> visible, and print types from other modules without import("…") qualifiers. Documentation sync now rejects descriptions that no page uses.
• Mark the speculation API experimental in the Reference, as the 9.0.0 notes stated. Add numbered runnable examples under examples/. Public library contracts are unchanged.

9.0.0

This major release renames public factories and declarations and removes the previous names, changes conversation and reporter contracts, and extends public event, fault and stop-reason unions. Replace previous names with their create*/define* equivalents, pass a store instead of a format name to createTransportConversations(), use each agent's native store instead of the format-keyed conversations helpers, and update exhaustive handling of AgentEvent (fallback, steer), FaultCode (steering) and stop reasons (steered). Firecracker is no longer experimental; speculation remains experimental. Live validation campaigns listed in the roadmap remain outstanding.

• Promote the Firecracker provider out of experimental: createFirecrackerSandboxProvider() and FirecrackerOptions no longer carry the experimental warning and icon. Host preparation, networking and the root supervisor of the jailer mode remain the operator's responsibility; control-plane resource accounting and adversarial validation remain open.
• Redraw the documentation navigation bar on the page grid: the logo cell lines up with the sidebar, search fills the centre, GitLab, the GitHub mirror and npm are linked beside the current version, and a single link switches between English and French. Narrow screens move the source links into the sidebar and menu footer.
• Give each built-in CLI agent its own src/adapters/agents/<agent>/ folder and a descriptor in one agent catalog: its label, executable, pinned version, harness export, installer, image environment, doctor diagnostics, protocol fixtures and outpost init authentication choices. agentVersions, DoctorAgent, remote bootstrap, outpost doctor, the outpost init prompts and validation, the generated run.ts, the image recipe and the agent image lock check now derive from it, and a test checks every descriptor against the public exports and the image lock. Generated projects, images and bootstrap commands are unchanged. agentVersions entries are now typed as string, and Antigravity model errors use the Antigravity label.
• Open native conversation formats to any CLI harness. createTranscriptConversations(layout) builds a native store for CLIs that keep one JSONL transcript per conversation, and createSessionBundleConversations(profile) for CLIs that keep each session as a directory, with self-contained validate, bucket and relocate hooks that run inside the sandbox. createClaudeConversations(), createCodexConversations(), createCopilotConversations() and createKimiConversations() return the built-in stores, which implement the new NativeConversationStore contract with directory() and destination(). Built-in presets now set storage to their native store by default, and AgentAdapter.conversations is removed. Breaking: createTransportConversations() takes a base store instead of a format name, and the format-keyed helpers conversations.native, locate, capture, restore, directory, destination and claudePath are removed; use each agent's native store. ConversationFormat is now string and StoredConversationFormat is removed. Persisted format names, transport keys and bundle contents are unchanged.
• Rebuild the documentation layout: Guide, Reference and Project pages share one header, a sidebar for the current space and breadcrumbs, and every section places its code, contract or entries beside its prose. Reference pages list parameters and properties as rows beside a pinned signature, /reference/ becomes a map of sections and families instead of redirecting to the first symbol, and the changelog and roadmap keep each heading beside its entries.
• Replace the documentation home with a landing page in English and French: the tagline with the install command and Get started, a code window with run.ts, brief.md and an example workflow, the agents and sandboxes with their conversation support, and a workflow walkthrough that highlights the code of each step. Its snippets are typechecked with the guide examples, and the documentation brand now shows the Outpost logo and links to this page.
• Name constructors create* and declarations define*. Functions that build runtime objects become createAgent, createFallbackAgent, createReplayAgent, createHarness, createClaudeHarness, createCodexHarness, createAntigravityHarness, createCopilotHarness, createKimiHarness, createHarnessFileTools, createHarnessEditTools, createHarnessSearchTools, createHarnessGitTools, createHarnessShellTools, createDockerSandboxProvider, createPodmanSandboxProvider, createLocalSandboxProvider, createVercelSandboxProvider, createDaytonaSandboxProvider, createFirecrackerSandboxProvider, createMountedSandboxProvider, createRemoteSandboxProvider, createLocalTransport, createS3Transport, createArtifactStore, createWorkflowCheckpointStore, createTaskCacheStore, createTransportConversations, createHarnessConversations, createSqliteTaskQueue, createHttpTaskQueue, createBullMQTaskQueue, createOpenAIModelProvider, createAnthropicModelProvider, createEd25519DecisionVerifier and createOpenTelemetryObserver. Workflow declarations become defineWorkflow, defineTask, defineAgentTask, defineIsolatedTask, defineCommandTask, defineApprovalTask, definePauseTask, defineArtifactTask, defineQueuedTask, defineInteractiveAgentTask and defineLoopTask; response.text/response.json become defineTextResponse/defineJsonResponse, and artifact.json/artifact.binary become defineJsonArtifact/defineBinaryArtifact. Breaking: the previou...

Outpost v9.0.3

Changelog

9.0.3

• Rebuild the documentation landing page as a path for a developer deciding whether to install: the headline runs on one rule, the install command carries its prerequisites and the first command to run, and a dispatch() call cut from a typechecked snippet shows one task end to end. The page then reads as what you can run, the comparison drawing, the agents and sandboxes it runs on, the argument that drawing opened, typed responses, and where Outpost is not the answer.
• Make the orchestration comparison argue in numbers. Two lanes, “The model orchestrates” and “Your code orchestrates”, play four beats; each step reports its own time — 0.6 s for a model step, 0.9 s for an agent turn, 0.1 s for code, read from one constants module so no figure can disagree with another — and the running total turns the Resume beat into a number: the model climbs from 2.4 s to 3.6 s because it rereads two steps, while the code lane holds at 1.2 s because restored steps run nothing. The model’s context meter turns ember past half and red past three quarters while the code lane stays at a quarter. The chain loops, carries on from a clicked beat, parks on pointer or focus, keeps the explicit pause toggle and shows a still state when motion is reduced.
• Drive the problem register from that figure instead of repeating it: one pain shows at a time beside the beat that makes its case, with a four-tick pager that fills over the beat’s own duration, a visually hidden list of all four for assistive technology and search, and the first pain served as static prose when scripts do not run.
• Show the whole typed-response call rather than its schema alone: the Zod declaration, the dispatch() that passes it as response and the typed value read back, cut at build time from a typechecked snippet, beside the agent’s reply in prose with its tagged block lit at the moment it becomes typed data.
• Rewrite the landing copy in both languages so each reads as written rather than translated, without moving a claim, a proof or a link.
• Stop flagging Firecracker as experimental on the landing page, where the reference already ships it stable; the note there now covers speculation only. Public library contracts are unchanged.

9.0.2

• Rebuild the documentation home page around “Code orchestrates. Agents think.”: an animated comparison of a pipeline orchestrated by an LLM and one orchestrated by code, the pains of LLM-driven pipelines paired with what Outpost does instead, a table showing that only the agent’s work varies between runs of the same workflow beside the real verification check that gates it, and a compact list of agents and sandboxes. Both figures can be paused or replayed, show still states when motion is reduced, and are described for screen readers. Public library contracts are unchanged.

9.0.1

• Rewrite the English and French Guide by reader task: twelve chapters in reading order, six complete use-case recipes, and pages drawn with feature cells, reading paths, flows and file trees. Moved Guide URLs redirect to their new pages.
• Rewrite the Reference: each family overview picks between its entry points with decision and outcome tables, and every symbol and field description was checked against the source and corrected where it was wrong, vague or outdated. Each symbol links to its own Guide page.
• List the values of 35 type aliases such as FaultCode, TaskStatus and SteeringMode, keep placeholders such as <name> visible, and print types from other modules without import("…") qualifiers. Documentation sync now rejects descriptions that no page uses.
• Mark the speculation API experimental in the Reference, as the 9.0.0 notes stated. Add numbered runnable examples under examples/. Public library contracts are unchanged.

9.0.0

This major release renames public factories and declarations and removes the previous names, changes conversation and reporter contracts, and extends public event, fault and stop-reason unions. Replace previous names with their create*/define* equivalents, pass a store instead of a format name to createTransportConversations(), use each agent's native store instead of the format-keyed conversations helpers, and update exhaustive handling of AgentEvent (fallback, steer), FaultCode (steering) and stop reasons (steered). Firecracker is no longer experimental; speculation remains experimental. Live validation campaigns listed in the roadmap remain outstanding.

• Promote the Firecracker provider out of experimental: createFirecrackerSandboxProvider() and FirecrackerOptions no longer carry the experimental warning and icon. Host preparation, networking and the root supervisor of the jailer mode remain the operator's responsibility; control-plane resource accounting and adversarial validation remain open.
• Redraw the documentation navigation bar on the page grid: the logo cell lines up with the sidebar, search fills the centre, GitLab, the GitHub mirror and npm are linked beside the current version, and a single link switches between English and French. Narrow screens move the source links into the sidebar and menu footer.
• Give each built-in CLI agent its own src/adapters/agents/<agent>/ folder and a descriptor in one agent catalog: its label, executable, pinned version, harness export, installer, image environment, doctor diagnostics, protocol fixtures and outpost init authentication choices. agentVersions, DoctorAgent, remote bootstrap, outpost doctor, the outpost init prompts and validation, the generated run.ts, the image recipe and the agent image lock check now derive from it, and a test checks every descriptor against the public exports and the image lock. Generated projects, images and bootstrap commands are unchanged. agentVersions entries are now typed as string, and Antigravity model errors use the Antigravity label.
• Open native conversation formats to any CLI harness. createTranscriptConversations(layout) builds a native store for CLIs that keep one JSONL transcript per conversation, and createSessionBundleConversations(profile) for CLIs that keep each session as a directory, with self-contained validate, bucket and relocate hooks that run inside the sandbox. createClaudeConversations(), createCodexConversations(), createCopilotConversations() and createKimiConversations() return the built-in stores, which implement the new NativeConversationStore contract with directory() and destination(). Built-in presets now set storage to their native store by default, and AgentAdapter.conversations is removed. Breaking: createTransportConversations() takes a base store instead of a format name, and the format-keyed helpers conversations.native, locate, capture, restore, directory, destination and claudePath are removed; use each agent's native store. ConversationFormat is now string and StoredConversationFormat is removed. Persisted format names, transport keys and bundle contents are unchanged.
• Rebuild the documentation layout: Guide, Reference and Project pages share one header, a sidebar for the current space and breadcrumbs, and every section places its code, contract or entries beside its prose. Reference pages list parameters and properties as rows beside a pinned signature, /reference/ becomes a map of sections and families instead of redirecting to the first symbol, and the changelog and roadmap keep each heading beside its entries.
• Replace the documentation home with a landing page in English and French: the tagline with the install command and Get started, a code window with run.ts, brief.md and an example workflow, the agents and sandboxes with their conversation support, and a workflow walkthrough that highlights the code of each step. Its snippets are typechecked with the guide examples, and the documentation brand now shows the Outpost logo and links to this page.
• Name constructors create* and declarations define*. Functions that build runtime objects become createAgent, createFallbackAgent, createReplayAgent, createHarness, createClaudeHarness, createCodexHarness, createAntigravityHarness, createCopilotHarness, createKimiHarness, createHarnessFileTools, createHarnessEditTools, createHarnessSearchTools, createHarnessGitTools, createHarnessShellTools, createDockerSandboxProvider, createPodmanSandboxProvider, createLocalSandboxProvider, createVercelSandboxProvider, createDaytonaSandboxProvider, createFirecrackerSandboxProvider, createMountedSandboxProvider, createRemoteSandboxProvider, createLocalTransport, createS3Transport, createArtifactStore, createWorkflowCheckpointStore, createTaskCacheStore, createTransportConversations, createHarnessConversations, createSqliteTaskQueue, createHttpTaskQueue, createBullMQTaskQueue, createOpenAIModelProvider, createAnthropicModelProvider, createEd25519DecisionVerifier and createOpenTelemetryObserver. Workflow declarations become defineWorkflow, defineTask, defineAgentTask, defineIsolatedTask, defineCommandTask, defineApprovalTask, definePauseTask, defineArtifactTask, defineQueuedTask, defineInteractiveAgentTask and defineLoopTask; response.text/response.json become defineTextResponse/defineJsonResponse, and artifact.json/artifact.binary become defineJsonArtifact/defineBinaryArtifact. Breaking: the previous names, including the provider, OpenTelemetry, S3 and BullMQ subpath names and the response and artifact objects, are no longer exported; old reference URLs redirect to the new pages. Breaking: createReporter() now creates the terminal reporter formerly named reporter(), and the handler-based reporter becomes createCustomReporter(handlers, options). Generated starters use the new names.
• Add triggers that start workflows through a queue. createCronSchedule() evaluates five-field cr...