Changelog

0.20.0 (2026-11-12)

• Use compatible bitbucket-cli native HTTPS Git authentication for manifest initialization, cloning, fetch and remote-head reads. SSH declarations and stored remotes are preserved; fx never handles token values. Missing or older companions retain ordinary Git, while rejected credentials remain failures.
• Update the companion catalog to bitbucket-cli 0.30.0 and wecom-channel-cli 0.4.0.

0.19.0 (2026-11-12)

• Read pinned version declarations and first-parent field history with repo version, including nested packages and an optional historical baseline. Release checks can include the same evidence without treating a source change as publication.
• Add absolute, timezone-qualified error windows and paginated Sentry issue metadata. Keep lifetime first-seen, window activity, current status and count scope distinct; report partial collection and detail failures explicitly.
• Add evidence_complete and per-repository facet coverage without changing release check verdicts or exit-zero reporting. repo show adds scoped responsibility evidence with date-filtered relationships from the current members document, without choosing an assignee.

0.18.1 (2026-11-12)

• Separate confirmed release failures from incomplete checks, group repeated unavailable causes with recovery, and preserve structured query errors. Missing Sentry projects no longer trigger organization-wide queries or misleading release failures.

0.18.0 (2026-11-12)

• Bind readiness evidence, provider commands, handoffs and probes to the same executable and workspace, including custom binary names.
• Scope tool diagnosis and repair before setup expansion; distinguish startup prerequisites from credential handoffs.
• Support manual personal-token acquisition for Sentry instances without device OAuth while preserving existing authentication.
• Add member selector discovery and workspace-preserving query corrections, with updated companion guidance and regression coverage.

0.17.0 (2026-11-10)

• Declare personal mentors and leaders on people; member list/show expose explicit IDs or null, and validation rejects invalid references and cycles.

0.16.1 (2026-11-10)

• Align npm companion minimum versions with the current published family and upstream releases, including wecom-chat-cli v0.3.0 and official wecom-cli v1.3.4.
• Document Chat 0.2.x watch-output and cursor migration; keep fx readiness scoped to the worker task and Skills managed by the host.

0.16.0 (2026-11-10)

• Require wecom-channel-cli v0.3.0+ for mandatory self-bot observation filtering and use its canonical doctor readiness command. Keep wecom-chat-cli at its latest published v0.2.0 floor, with task-scoped readiness and host-managed Skills.

• Add wecom-chat-cli v0.2.0+ to the tool catalog and as an im provider for isolated workers. Readiness checks only the current task, keeps chat text out of reports, and never issues grants or approves replies. Its packaged Skill remains host-managed.

• Query directory memberships, personal functions and business collaboration separately with members schema 2. Scoped member lookups return the matching relationships, validity dates and sources. Schema 1 remains readable.

0.15.0 (2026-11-02)

Added

• Bind the workspace capability channel to wecom-channel-cli to install the CLI and its Skill. Readiness checks the local broker through the agent socket.

Fixed

• Keep local files that a sync update would replace. When upstream starts tracking a path that a checkout holds as an ignored file, such as a local .env, fx sync fast-forwarded over it and replaced its contents. It now leaves that checkout as it is and skips it with outcome local_files_in_the_way, lists the files in files_in_the_way and holds readiness until they are moved aside. An untracked file in the same position, which Git already refused to replace, is reported the same way instead of as a failed GIT_FAST_FORWARD_FAILED row. The manifest checkout is kept the same way.

0.14.0 (2026-10-28)

Added

• Refresh Skills when tools change. fx tool install and fx tool update have each tool they installed or updated install its Skill again for your coding agents, so the agents read the instructions of the version now on PATH; each row gains skill. fx self update ends by having the newly installed fx refresh the copies of its own Skill that are older than it, reported as skill_refresh, and the new fx skill update does that on its own: it refreshes only recognized copies older than the running fx, in every agent's directory, and installs none. Keeping everything current is fx self update --yes and fx tool update --all.

Changed

• Update only installed tools with fx tool update --all. Outside a workspace it used to take the whole catalog and install every tool not yet there; inside one it installed the workspace's missing tools. It now leaves those to fx tool install --all, and inside a workspace a TOOLS_NOT_INSTALLED notice names them. The FX_OUTDATED notice no longer suggests fx skill install.

Fixed

• Prefer the root VERSION file over dependency metadata in fx release check, so services with a separate release version no longer show a missing or package-only version. Dependency checks continue to inspect the dependency file.

0.13.0 (2026-10-27)

Added

• Set up companion tools without a manifest. fx init with no directory and no source lists the tools in fx's catalog and asks which to set up, by number or name, with Enter taking all of them; fx init --tools <name,...> or --tools all names them for agents and scripts, and --dry-run previews. fx installs the tools that are missing, has each one install its Skill for your coding agents, adds its own fx-workspace Skill, and checks whether each one is signed in. On a terminal it offers to start the setup of a tool that is not signed in (<tool> config init, gh auth login), which asks for the service and your credentials; otherwise, and with --yes, the report hands that command to you. No workspace is registered, and a manifest still decides the tools and presets the services they point at. The JSON report has tools, tool_install and skills, and preparation.scope.workspace is null.
• Let a tool's auth in the catalog or a manifest name the JSON field its status must report (json) and the tool's own interactive setup (login). The bundled catalog names both for the family CLIs and login for gh, and fx tool show reports them.

Changed

• Tell a companion's reuse check in a background hook run how long the run can give it, through the companion's own --timeout, and leave the check unasked when that is under the ten seconds by which a companion bounds the lookups a reuse adds. The check used to be ended when the run's time for it ran out. The reuse step's verify in fx login --list and the init preview gains deadline, with the flag and the least time the check needs (minimumMs).
• Record LOGIN_UNCHECKED in place of LOGIN_REQUIRED when every login check of a background hook run that did not pass was not asked or gave no answer, which fx doctor reports as incomplete coverage, as when the run ran out of time for it.

Fixed

• Count a Jenkins or OpenObserve environment as signed in only when jenkins-cli auth status or openobserve-cli auth status reports authenticated: true. Both exit zero when nobody is signed in, so on a machine without credentials the sign-in step counted as done and fx init and fx sync did not ask for it.
• Leave the declared authentication method out of the OpenObserve and Prometheus commands fx where prints and probes, as checks have done since 0.9.1. A probe of an OpenObserve context that uses basic authentication was sent with --auth-scheme session and refused; the commands now name only the server's URL and organization besides the context.

0.12.0 (2026-10-19)

Added

• Open pull requests on GitHub. fx source push works with a manifest source hosted on GitHub as it does with Bitbucket: it pushes the selected files to your verified personal fork through gh and opens, reuses or retires the pull request against its upstream, and --pr accepts a pull request number or its page on GitHub.
• Recognize checkouts cloned from github.com over HTTPS or SSH in fx import, which binds the github code host with the matching clone prefix as its host. A GitHub Enterprise server is still not guessed from a remote.
• Add wecom-calendar-cli, a local, searchable mirror of your own WeCom calendars with their history, to the tool catalog and as a calendar provider. fx can checks each member's sign-in and hands agents wecom-calendar-cli event list; filling the mirror with wecom-calendar-cli sync stays the member's call.

Changed

• Judge a CI build by the commit it checked out when the platform names it. fx release check passes a successful build of the branch head and fails one of another commit, instead of comparing when the build started with when the head was written. GitHub Actions names the commit, and Jenkins does with jenkins-cli 0.12.0 or later; without it the times still decide. CI evidence gains commit.
• Let a github-actions ci environment leave out branch. fx release check then counts the runs of the branch it is checking on each side.
• Keep MCP, rules and hook rows that report untrusted, missing_variables, invalid or busy in the summary of an init, sync or doctor JSON report. They were counted under omitted, while a foreign or conflict row of the same stage was shown.
• Report TOOL_MISSING from a companion query, sentry or git only when the executable could not be started. One that started and exited 127 itself now reads as the failed query or Git operation it is, in place of a request to install a tool that is already there.
• Say "<id>" is not an agent fx knows for an unknown --agent of fx can mcp:<name>, as fx agent pick and fx skill install do. fx release check reports an unknown repository in the words of the other repository commands, lists the known names when none is close, and keeps the workspace in its next step.
• Quote a directory, tool or repository name that needs it in the command an error suggests: the imported directory of IMPORT_INSIDE_ROOT, the tools of a withheld installation or a missing companion Skill tool, and the repository of PROBE_RECORD_INPUT_REQUIRED. A path with a space can be pasted as printed.
• Quote for the configured command_shell the commands that follow a workspace relocate or rebind, a registration fx init cannot take or resume, a profile import or source contribution whose workspace changed, a probe that could not be recorded, a manifest import aimed inside the directory it describes, and a --workspace that names nothing registered. These were printed in POSIX quoting whatever the setting.
• Run git in the C locale when reading a manifest repository or a local-git code host, as repository synchronization does. Under another locale a file missing from a branch was reported as a failed query, because git's translated message went unrecognised. A git that cannot be started there is now TOOL_MISSING.
• Close fx sync and fx init with one line per stage before the verdict: the repositories updated, already current, skipped and failed, by name; what the manifest refresh did; and the counts other stages report. In a terminal each success is counted on the status line instead of printed, and so is the SYNC_PARTIAL notice the closing view repeats; --verbose prints them, and a pipe or JSON still receives every record. The status line printed under a Git call no longer repeats itself while other lines keep coming.
• Group sync and init blockers by what resolves them: a sign-in or credentials, access refused, or a server this machine cannot reach, which now gets the command that checks it again instead of fx login; a Prometheus server that does not answer counts as unreachable. A check that waited on another blocker, directly or through a third, is named under it instead of counted apart, a companion's error code follows its message, and a preparation step whose companion says a retry will not help shows the companion's own advice instead of suggesting fx doctor --fix. Preparation checks gain cause and waits_on.
• Run the bundled step that reuses an existing login in fx login without asking and off the terminal, as init and sync do, also with --no-input. Its failure keeps the companion's message and code instead of exit 3, a terminal shows the companion's advice in a LOGIN_STEP_FAILED notice, and the companion's JSON no longer lands on the terminal; fx asks only before a sign-in.
• Show only non-zero counts and no empty problems under the human fx login table, and leave the LOGIN_SUMMARY notice that repeats them to JSON and pipes. A passing check says what it established, such as authenticated is true, instead of repeating its command.

Fixed

• Refuse a recorded harness or contribution path that has a .git segment in any letter case or a delete character. Harness ownership metadata accepted .GIT/..., which a case-insensitive volume resolves into a repository's own directory; both stores now hold a recorded path to the rule fx source push applies to a selection.
• Report bundled_tools in fx self info as a path the file system accepts when fx is installed under a directory whose name has a space or a non-ASCII character, or on Windows. The value was a percent-encoded URL path there.
• Leave a registration that records no manifest source as it is when fx init --resume continues it. Resuming wrote source: "-" into config.yaml: fx sync then took the workspace for one prepared from stdin and stopped saying that it has no source to refresh, and fx init <dir> --from <source> refused the directory with SOURCE_MISMATCH in place of recording the source.
• Skip a repository whose upstream is still empty, such as one registered with fx repo add before its first push, instead of failing the whole sync. fx asks the remote before calling a branch absent: an empty upstream is a skip with outcome empty_remote that does not hold readiness back, while an upstream with other branches but not the tracked one is a skip with outcome no_remote_branch that still does, since the manifest may name the wrong branch. An unborn local branch fast-forwards once upstream has commits, and a manifest checkout in either state is left as it is.
• Bring earlier environments up to date as soon as fx login signs in to one. The tool's earlier steps that failed are checked again and those that waited on a step now passing are attempted, so environments sharing a stored login no longer stay failed or blocked in the summary. No setup action runs a second time; a check that still fails keeps its result and says what it found under recheck.
• Say what became of the manifest checkout in its sync row instead of always "refreshed": refreshed, already current, or kept with the reason, adding when the commits of the branch it was kept on are already upstream.
• Read the reason a failing check prints inside a JSON object after a prefix, such as kubectl's Error from server (Forbidden): {…}, instead of quoting the object.

0.11.0 (2026-10-12)

Added

• Keep a go.work at the workspace root when the manifest sets go.workspace: true. fx init and fx sync list every selected checkout that has a go.mod at the newest go version they require, so each module builds against its siblings' checkouts. A repository not cloned yet is left out, fx replaces or removes only the go.work it wrote, and fx sync --only go refreshes it alone. go.work and go.work.sum no longer count as stray root entries in such a workspace.
• Bind a workspace to GitHub. The github code-host provider and the github-actions ci provider both drive the GitHub CLI gh, which joins the tool catalog with one sign-in per host for both. fx release check reads branches, commits, files and the last workflow run of a branch from GitHub just as it does from Bitbucket and Jenkins; a ci environment can narrow the runs it counts to one branch and one trigger. Opening pull requests through fx is not available on GitHub yet.

Changed

• Rename fx release to fx release check, so the name says it checks a release rather than making one. fx release keeps working as a compatibility name.
• Say what each release would ship. fx release check opens each repository with the version its target branch declares, the head commits of both branches and how many commits on the source branch are not merged into the target yet, by how many people; repositories whose checks all passed close the output one line each with the same. JSON reports gain repos[] with these facts, summaries included; every existing field is unchanged.
• Name the commit a CI build is judged against, say "not merged into main" and "not merged back into dev" in the merge check, and use each platform's own name for the job or project in CI and error messages. CI evidence gains head.
• Show a repository's release.note in the release report, as the manifest schema describes.

Fixed

• Resolve every relative link in the source AGENTS.md that names a file of the manifest source, such as CONTRIBUTING.md, in the guide derived at the workspace root. Only links into README.md, docs/ and skills/ resolved before, so the others pointed at files the root does not have.
• Stop fx manifest lint from asking for a code-host provider when every repository names its upstream, since such a workspace reaches no code host.
• Pass every argument intact to a companion that Windows starts through a batch file (.cmd, as npm, Scoop and Chocolatey install them). cmd.exe no longer interprets characters such as &, |, % or ^ in an argument.
• Read a local Git remote given as a Windows path, such as C:\host\fx\api.git, as project and name, so import, adoption, forks and fx source push work with it. Manifest documents are named with forward slashes in stdin bundles and source contributions on every platform, and import writes repository paths with them.
• Stop the PowerShell hook from writing task results into the agent's context, passing fx its input with a byte order mark, and writing progress records to stderr.
• Link source files with forward slashes in guides derived on Windows, and report the bundled tools catalog in fx self info as a file path, also when the install path contains spaces.

0.10.1 (2026-10-12)

Added

• Publish the public @angelmsger/fx package to npmjs and attach its tarball to a GitHub Release for each version tag. The internal @fx/devkit channel is unchanged.

Changed

• Group the human output of fx release by repository. Each repository with a failing or warning check gets its own block listing all of its checks, a problem every repository shares, such as a tool nobody has signed in to, is stated once above them with the command that resolves it, and the repositories whose checks all passed share one closing line. --details lists their checks too. JSON and NDJSON output are unchanged.

0.10.0 (2026-10-10)

Added

• List what a workspace can be asked about with fx can and no target: one row per capability a manifest binds, and per platform instance for an environment-scoped one, with its scope, provider, companion tool and what it is for. A capability nothing is bound to keeps its row without a provider. The list reads bindings only; asking about one capability still checks readiness.

Changed

• Say what a failed readiness check answered. Each pending step in a can, where, doctor or login handoff now carries check: what the tool said, or the step this one waits on. Where fx can read the failure it also names the cause, and the step's instructions then say what resolves it.
• Tell a cluster that did not establish authentication from one that answered no and from one that could not be reached. A request the cluster took for nobody's now leads to a step about the kubeconfig and its sign-in, with the page an updated kubeconfig comes from, rather than a request to the cluster's owner for a role.
• Check that a kubeconfig names its declared context without contacting the cluster, and give each request of the cluster access check five seconds.
• Answer unknown from can and where when a check gives no answer, because its tool did not start or it outlived its deadline, and list it under unfinished. where no longer fails for every platform of a repository because one of them did not answer. doctor and login report a tool that did not start as a missing tool in place of handing its step to a person.
• Give <env> declares no <platform> platform as the reason an environment does not cover a platform. The environment's note, which may be about another platform, is no longer that reason.
• Stop asking which environments a person uses during fx login. The personal environment selection is set with fx env select alone: until it is, every declared environment is in scope, and a login that names none with --env says so and how to narrow it. The question was asked even when --env named the environments, came back on every login for a person who answered "all", and narrowed what fx doctor and fx sync cover for one who answered with only the environment at hand.
• Leave the capabilities fx wires itself out of what an unknown fx can target is offered, and point its next steps at fx can and fx tool list.

0.9.1 (2026-10-09)

Changed

• Show one row for checks that say the same thing and are resolved by the same commands in the human view of an incomplete init, sync or doctor, and mark a check that could not run apart from one that failed.
• Stop handing a sign-in to a check that could not finish. A timed-out companion check now only asks to be checked again.
• Quote why a bundled companion says it is not ready, in place of expected authenticated to be true: OpenObserve returned HTTP 401, for example, or a failed request to a server that takes no sign-in.
• Offer fx doctor --fix for Skill copies fx can bring up to date itself.
• Accept a companion context that reaches the declared service however its owner signed in. A service declaration's auth_scheme now only shapes a context fx creates and the sign-in fx login offers; preparation, checks and queries hold a context to the declared URL, flavor and organization and no longer pass --auth-scheme. An existing context that lacks one of those fields gains it without the declared method being named, which used to conflict with a context that signs in another way.
• Associate a login a person already has on every explicit fx sync, not only during fx init. Sync now applies local prepare-phase setup steps the way init does, so an adoptable login no longer waits for someone to run fx doctor --fix; personal sign-in still stays with the person, and background hooks only check.
• Say what an unmet companion check means where fx knows it. A login that fx doctor --fix can associate reads an existing login can be reused, and the three WeCom checks, now all judged by the response's error code, read the WeCom API returned an error.

0.9.0 (2026-10-09)

Changed

• Derive AGENTS.md alone into the workspace root. fx no longer writes a CLAUDE.md pointing at it, and fx manifest render no longer creates one in the manifest repository: Claude Code 2.1.277 and later reads AGENTS.md by itself, and reads a CLAUDE.md in its place whenever one exists. fx sync removes the pointer an earlier version derived; fx doctor reports the guide as stale until it has. A CLAUDE.md you wrote stays where it is and no longer counts as a guide conflict; start it with @AGENTS.md so Claude Code still reads the guide. Delete the pointer from a manifest repository that still carries it.
• Leave CLAUDE.md out of the manifest source payload. A directory source's copy is no longer materialized, and a stdin bundle that names it is rejected with MANIFEST_SOURCE_INVALID.
• Show that a command is still working. After half a second a terminal gets one status line that is redrawn in place: the time elapsed, how far counted work has come and what has been under way the longest, such as doctor 23s checks 69/82 kubectl.access@dev 3s. Every command has it, because each external call fx waits on is reported where it is started; doctor, sync, init, status, release and login also count their checks, repositories and steps. Results, notices and outcomes are written above the line, a question or a sign-in command takes the terminal over until it is done, and the line is gone when the command ends or is interrupted. While git or an installer runs, which may ask for a host key or a passphrase on the same terminal, nothing is redrawn under it and the line is printed every five seconds instead. On a terminal that shows the line, the per-item start lines and the repository heartbeat are no longer printed, since the line says the same. A pipe, a file, CI, TERM=dumb, --quiet and JSON or NDJSON output never receive the line and read exactly what they did.

0.8.0 (2026-10-09)

Changed

• Run the commands fx defines for bundled tools without asking for trust, including the service preparation, login reuse and credential checks that a manifest service declaration fills in. Only command vectors the team's tools.yaml patch authors still require confirmation; a companion decides where its stored credentials may be sent.
• Rework the human view of an incomplete init, sync or doctor: each blocker appears under its check id, sign-in commands are joined per tool across environments, commands are spelled fx when PATH resolves it, and a terminal no longer receives the PREPARATION_INCOMPLETE envelope after the view. JSON output is unchanged.
• Report companion Skill installation progress once per tool instead of once per agent directory.
• Approve a team-authored companion Skill command once. The agent name or directory fx appends no longer makes each target a separate vector to confirm, and existing approvals stay valid.
• Say why a companion check failed in the tool's own words: read the message of an error object it printed, set log records and warnings aside, keep the line that follows where it failed, and state the expected value of a JSON check, as in expected authenticated to be true.
• Make human output readable on a terminal. Headings, state words and zero counts are told apart by colour, check lists and progress carry status symbols, commands to type stand out, tables fit the terminal width, a blocker's long reason wraps under its own column, and an error prints as text with its hint, next steps and hand-off instead of a JSON envelope. Only a terminal is decorated, each stream on its own: whatever the environment says, a pipe, a file and CI receive plain text and the JSON error envelope, and JSON and NDJSON output is unchanged. NO_COLOR removes the colour and keeps the symbols.
• Align human tables by terminal columns so rows holding CJK text line up, set a list's trailer apart with a blank line, drop trailing spaces after absent values, and print a list inside a record one field per line once an item runs long, leaving out empty fields.

Fixed

• Root help on a terminal titles the everyday commands EVERYDAY again, and the FIRST WORKSPACE, AGENT USAGE and COMPATIBILITY headings are emphasised like the others.

0.7.1 (2026-10-09)

Added

• Return next_steps from fx repo show that select and synchronize a shared repository not on disk yet, and say in the generated repository index that the catalog includes such repositories.
• Print each personal sign-in command once in human preparation output, followed by the command that continues afterwards.

Changed

• Synchronize repositories without a code-host username, and without the notice about one, when the manifest uses explicit Git URLs and binds no code host.
• Install the bundled Sentry tool with npm (sentry).

0.7.0 (2026-10-08)

Added

• Name .workbench/ as the place for output that belongs to no repository. The derived guide gains a workspace-layout region with its rules, fx doctor reports root entries that nothing accounts for in a root group, fx status returns them as stray_entries, and the session-start context carries the latest observation. fx never writes inside the workbench.

0.6.0 (2026-10-08)

Added

• Implement metrics with a prometheus-cli provider and add the companion to the bundled catalog. A service's coordinate is its scrape job; where prints an up{job="..."} query pinned to the declared server and offers no probe, because a selector that matches nothing still succeeds.
• Accept auth_scheme: none and bearer in a service declaration. A server declared without sign-in is ready when it answers and is never handed a login.
• Add the workspace capability meeting for the conferencing system that holds online meetings, recordings and minutes, with a tencent-meeting provider driven by tmeet, now in the bundled catalog. calendar keeps schedules and availability.

0.5.0 (2026-09-24)

Added

• Reuse verified native companion logins in canonical team contexts before requiring personal authentication; retain native setup errors and verification details. Updated family CLI release floors require their auth reuse command.

0.4.0 (2026-09-21)

• Remove the self-reported Skill-loaded handshake and its notice setting; keep explicit installation/version discovery and unverified host activation.
• Prepare authorized local setup through init, hand off personal authentication together, and resume without repeating completed work.
• Replace preparation booleans with a scoped verdict and executable recovery actions; incomplete preparation exits 12, optional activation stays separate.
• Reject invalid declarations before workspace effects and complete generated guides without changing shared sources. Tools schema 3 requires run phases prepare/authenticate.
• Reuse npm scope/default registries and native authentication for self-update and version checks; make FX_REGISTRY optional and preserve public dependency routing.

0.3.0 (2026-09-20)

• Preserve login partial results, summaries and structured recovery information after handoffs or exceptions, including on nonzero exit.

• Protect manifest and guide edits from stale plans; move approved originals to backups atomically and retain per-file outcomes after document failures.

• State overall initialization/sync preparation clearly and stop interpreting descriptive metadata as failure states.

• Keep self-update installer output out of JSON stdout; clarify recovery and the difference between source documentation and published features.

• Add the calendar, task, worklog and metrics capabilities, with Lark and WeCom providers for the first two and a Jira provider for worklogs; metrics is named but unimplemented, so a manifest can describe it and binding it awaits a provider. im now means messaging alone.

• Add a Jira issue tracker and a Lark wiki knowledge base, so those capabilities each have a second implementation; one companion CLI serving several capabilities is installed and checked once.

• Derive a workspace's toolchain from the providers it binds, so a bundled tool no longer has to be named twice; tool list says whether a binding or the manifest put each tool there, and --catalog shows the rest.

• Let a capability bind a provider fx does not implement when the manifest declares a tool of that name, and correct an id that only resembles a bundled tool's name instead of obeying it.

• Accept install_on_init in a tools document's defaults block, which an entry can still override.

• Report a bound provider whose companion tool no tools document declares as TOOL_UNDECLARED, instead of claiming the capability has no provider.

• Check a platform's service block against the provider its name implies, so a valid declaration is no longer rejected and an invalid one is no longer skipped.

• Record the background phases a session-start run commits to, so phases a terminated worker never reached read as HOOK_PHASE_UNREPORTED instead of disappearing.

• Keep a background session-start run going when one companion command exhausts its budget, so every later phase still reports; a foreground run a host waits on still ends there.

• Record how a detached background run ended as the workspace's run observation, which its supervisor could previously write only to a stream nobody reads.

• Preserve explicit report fields in all output formats, workspace scope in recovery commands, and source identity through contribution locking.

• Import verified local Git coordinates, safely add matching upstream remotes to existing clones, and align preparation with enabled/required tool declarations.

• Accept all declared setup verification forms and preserve nullable comparison values through manifest overlays.

• Compatibility: fx skill show reports a missing bundled Skill as BUNDLED_SKILL_MISSING; SKILL_MISSING now names only the agent-facing notice.

• Add global workspace discovery, health, unregister, relocation and source rebinding while preserving business checkouts and preventing duplicate directory identities.

• Add repo/env select with explicit, dynamic-all and empty scopes, first-init repository selection and inspected local-source previews.

• Fix untracked work omitted from status filters, required manifests incorrectly passing validation, and stale readability caches blocking recovered repositories.

• Scope default doctor tool checks to declared tools and active provider requirements; retain explicit tool inspection.

• Compatibility: init registered now means actual registration state; intended effects are in plan. Profile import merges registrations by name and retains unmentioned workspaces; use workspace unregister for removal.

0.2.1 (2026-09-18)

• Shorten the repository entry and bundled Skill; keep detailed recovery in the command reference.
• Align report, readiness, Skill-loading and service-version guidance with current behavior across all four languages.
• Correct website release labels and preview examples, and document scoped Registry routing for public dependencies.

0.2.0 (2026-09-17)

• Breaking: init, sync, doctor and release JSON defaults to summaries. Use --details for all rows and evidence; --fields still projects full results.

• Preserve explicit empty agent selections across CLI, hooks and cleanup.

• Add advisory FX_SKILL_VERSION verification and bundled_skill_version discovery.

• Add operation / preparation_complete and SKILL_WRITE_DENIED recovery while preserving shared locks.

• Distribute public service presets during init and explicit sync, with scoped personal login and native conflict recovery.

• Reject personal credentials in shared service declarations and preserve existing companion configuration.

• Add standalone companion Skill install, uninstall, path, status and show commands.

• Generate team Skill discovery links and trigger descriptions in workspace guides.

• Separate explicit Skill refresh from automatic maintenance and protect shared sources.

• Track companion installation history and origin across registered workspaces; report unused fx-installed packages without uninstalling them.

• Prefer canonical resource commands in help/completion while preserving compatibility aliases.

0.1.2

Available through the internal Registry. Public publication remains on hold.

• Correct explicit input, readiness, Git-source recovery and concurrent configuration updates.
• Preserve partial initialization results, summarize human output and report repository heartbeats.
• Isolate blocked Skill commands, protect shared user copies and document host activation limits.
• Bound optional code-host probes and report preflight progress while preserving cancellation.
• Accept workspace root aliases and retain per-item installation and resource state after interruptions.

0.1.1

Available through the internal Registry. Public publication remains on hold. Git tags, registry publication and local installation are separate states. The project is fx-devkit, its binary is fx, and its license is Apache-2.0. Internal distribution uses @fx/devkit; public distribution uses @angelmsger/fx.

• Opt-in installation of missing companion tools during initialization (install_on_init, fx ≥0.1.1).
• Manifest-driven workspace initialization, synchronization and repository selection.
• Structured configuration, discovery, readiness and release reports.
• Companion tool integration, agent Skills, lifecycle hooks and MCP projections.
• Source contribution planning with trust, ownership and recovery boundaries.
• Native Windows adapters; native execution and host activation remain unverified.