[github-actions]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and publish to npm yourself or setup this action to publish automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@rivus/agent-finder-cli@0.3.0

Minor Changes

• #150 14d6ab7 Thanks @nyx-cell! - Detect installed agents with @rivus/agent-kit 0.11.0 instead of the MoonBit scanner, and require Node.js 22.13 or newer (the kit's Node adapters).

  The kit supports 7 agents with built-in probe recipes: claude-code, codex, cursor, gemini-cli, grok, opencode, pi. agent-finder-cli keeps the other 20 catalog entries through its own probe recipes (openhands, cline, codebuddy, command-code, kiro-cli, antigravity, roo-code, github-copilot, amp, openclaw, neovate, qoder, zencoder, kimi-code-cli, windsurf, vscode-copilot, codex-desktop, aider, hermes, trae), so scan, provider list, provider inspect and doctor list 27 agents: the previous 26 plus the kit's grok.

  Differences accepted in the parity comparison, for the same probe inputs:

  • Report order changes: the kit's built-in agents come first (opencode, claude-code, codex, cursor, pi, gemini-cli, grok), then the legacy recipes; the MoonBit scanner used catalog file order.
  • The conservative "provider metadata" warnings on opencode and pi no longer appear; the kit's recipes for them carry no warnings, while the 20 legacy recipes keep theirs verbatim.
  • Cursor: ~/.cursor is now checked as a configuration path, cursor-agent as a command name, and the recipe adds two caveats about not following CURSOR_CONFIG_DIR/XDG_CONFIG_HOME and not probing the command name agent.
  • opencode also checks ~/.config/opencode/opencode.jsonc as an MCP configuration path.
  • Each installation carries the new auth (file-based login state) and problems (checks that could not complete) fields; agent-finder scan --json prints that installation array instead of the old schema_version: "0.1" report.
  • agent-task-loop init detects claude-code and codex with versionProbe: false, so it never executes an agent command; an agent is configured when its command is found on PATH, where discovery previously also required a successful --version run.

• #151 60fb524 Thanks @nyx-cell! - Discover sessions and read transcripts with @rivus/agent-kit 0.11.0 (/sessions + /transcript) instead of the deleted internal @rivus/agent-sessions package. Resume command strings stay application-side: packages/agent-finder-cli/src/sessions/resume.ts and packages/agent-task-loop/src/session/resume.ts are the single sources (codex resume <id>, claude --resume <id>).

  Differences accepted in the parity comparison against the old package, for the same on-disk fixtures:

  • The session list covers more sessions: Codex archived_sessions/ is scanned, CODEX_HOME and CLAUDE_CONFIG_DIR are honored, and grok sessions are included. The old scanner only walked ~/.codex/sessions and ~/.claude/projects.
  • Claude subagent transcripts (<session>/subagents/agent-*.jsonl) are no longer listed as independent sessions; their turns appear inside the parent session's transcript instead.
  • agent-finder sessions list --json prints kit session heads (ref, lastActiveAt epoch ms, sizeBytes, firstPrompt, …) instead of the old Session shape, and --agent claude is accepted as an alias of claude-code. sessions inspect --json prints the transcript as kit transcript events.
  • Timestamps: updatedAt (file mtime, ISO) becomes lastActiveAt (the last record time read from the file's head and tail, epoch ms).
  • A Codex rollout records each turn twice (the CLI's event_msg copy and the model's response_item copy). The preview renders each turn once, attributed to the response_item; the old parser rendered both copies. A regression test pins the single count.
  • The task-loop TUI preview and the sessions browser read transcript events (user/assistant/reasoning/tool_call); tool calls render as ⚙ tool regardless of which agent emitted them.

• #134 778fe6f Thanks @PerfectPan! - Require Node.js 22 or newer; Node.js 20 is no longer supported. The CLIs move to Ink 7, and @rivus/agent-task-loop also to execa 10; both need Node.js 22.

Patch Changes

• #130 68a1d1c Thanks @PerfectPan! - Upgrade the runtime dependencies to Ink 6 with React 19 and citty 0.2. -v prints the version, --help output uses citty 0.2's layout, and a value after a boolean flag is no longer read as the flag's value (--json false turns JSON output on; use --no-json).

• #125 c006fce Thanks @PerfectPan! - License @rivus/agent-finder-cli under MIT instead of GPL-3.0-only. The package now ships its own LICENSE file with the MIT text; the session browser code it bundles from the workspace library @rivus/agent-sessions is MIT as well.

• #139 24fc8aa Thanks @PerfectPan! - Ship dist/cli.d.ts, the declaration file the package's types export points at, and stop shipping the declarations of its tests and Vitest config.

• #140 735a25a Thanks @PerfectPan! - agent-finder -v and agent-finder provider -h print the package's version instead of a hard-coded 0.1.0.

• #149 5cd6419 Thanks @nyx-cell! - Build the published JavaScript with Vite+ (vp pack). The bundle changes form, not behavior.

• #138 f34f632 Thanks @PerfectPan! - Fix the published TUIs (agent-task-loop tui and agent-finder sessions browse) crashing on start with ReferenceError: React is not defined: the build now compiles JSX with the automatic runtime.

@rivus/agent-task-loop@0.12.0

Minor Changes

• #150 14d6ab7 Thanks @nyx-cell! - Detect installed agents with @rivus/agent-kit 0.11.0 instead of the MoonBit scanner, and require Node.js 22.13 or newer (the kit's Node adapters).

  The kit supports 7 agents with built-in probe recipes: claude-code, codex, cursor, gemini-cli, grok, opencode, pi. agent-finder-cli keeps the other 20 catalog entries through its own probe recipes (openhands, cline, codebuddy, command-code, kiro-cli, antigravity, roo-code, github-copilot, amp, openclaw, neovate, qoder, zencoder, kimi-code-cli, windsurf, vscode-copilot, codex-desktop, aider, hermes, trae), so scan, provider list, provider inspect and doctor list 27 agents: the previous 26 plus the kit's grok.

  Differences accepted in the parity comparison, for the same probe inputs:

  • Report order changes: the kit's built-in agents come first (opencode, claude-code, codex, cursor, pi, gemini-cli, grok), then the legacy recipes; the MoonBit scanner used catalog file order.
  • The conservative "provider metadata" warnings on opencode and pi no longer appear; the kit's recipes for them carry no warnings, while the 20 legacy recipes keep theirs verbatim.
  • Cursor: ~/.cursor is now checked as a configuration path, cursor-agent as a command name, and the recipe adds two caveats about not following CURSOR_CONFIG_DIR/XDG_CONFIG_HOME and not probing the command name agent.
  • opencode also checks ~/.config/opencode/opencode.jsonc as an MCP configuration path.
  • Each installation carries the new auth (file-based login state) and problems (checks that could not complete) fields; agent-finder scan --json prints that installation array instead of the old schema_version: "0.1" report.
  • agent-task-loop init detects claude-code and codex with versionProbe: false, so it never executes an agent command; an agent is configured when its command is found on PATH, where discovery previously also required a successful --version run.

• #151 60fb524 Thanks @nyx-cell! - Discover sessions and read transcripts with @rivus/agent-kit 0.11.0 (/sessions + /transcript) instead of the deleted internal @rivus/agent-sessions package. Resume command strings stay application-side: packages/agent-finder-cli/src/sessions/resume.ts and packages/agent-task-loop/src/session/resume.ts are the single sources (codex resume <id>, claude --resume <id>).

  Differences accepted in the parity comparison against the old package, for the same on-disk fixtures:

  • The session list covers more sessions: Codex archived_sessions/ is scanned, CODEX_HOME and CLAUDE_CONFIG_DIR are honored, and grok sessions are included. The old scanner only walked ~/.codex/sessions and ~/.claude/projects.
  • Claude subagent transcripts (<session>/subagents/agent-*.jsonl) are no longer listed as independent sessions; their turns appear inside the parent session's transcript instead.
  • agent-finder sessions list --json prints kit session heads (ref, lastActiveAt epoch ms, sizeBytes, firstPrompt, …) instead of the old Session shape, and --agent claude is accepted as an alias of claude-code. sessions inspect --json prints the transcript as kit transcript events.
  • Timestamps: updatedAt (file mtime, ISO) becomes lastActiveAt (the last record time read from the file's head and tail, epoch ms).
  • A Codex rollout records each turn twice (the CLI's event_msg copy and the model's response_item copy). The preview renders each turn once, attributed to the response_item; the old parser rendered both copies. A regression test pins the single count.
  • The task-loop TUI preview and the sessions browser read transcript events (user/assistant/reasoning/tool_call); tool calls render as ⚙ tool regardless of which agent emitted them.

• #130 68a1d1c Thanks @PerfectPan! - Upgrade the runtime dependencies to Zod 4, Ink 6 with React 19, and citty 0.2. User-visible changes:

  • A value after a boolean flag is rejected: cleanup --force false used to turn --force off, but citty 0.2 never reads a value for a boolean flag and would turn it on. Turn a flag off with --no-force or --force=false. Every command also rejects other positional arguments it does not take.
  • --max-rounds on start and reject now takes effect; it used to be ignored in favour of the default of 5 unless written --maxRounds.
  • -v prints the version, and --help output uses citty 0.2's layout.
  • A failing command prints only its error message instead of the whole error object.
  • Validation errors from create, source add and config loading use Zod 4's wording, for example Invalid --priority: Too big: expected number to be <=9.

• #152 a7be2e7 Thanks @nyx-cell! - Occupy Tasks with @rivus/agent-kit-collab 0.11.0's /lease (createLeaseManager + a scoped manager.acquire) instead of @rivus/agent-orchestration's LeaseManager and FileLeaseStore. effect@4.0.1 is an exact dependency. The old orchestration facade (src/orchestration/orchestration.ts, node-factory.ts, task-orchestration.ts) is deleted; the run baton (Run, the classic-delivery template, the run-state stores, the seat policy) stays application-side and the seat lifecycle is wired to the kit lease handle. The occupancy key stays task:<taskId>.

  Differences accepted in the old/new comparison, recorded in the lease tests:

  • The lease heartbeat is the kit's managed fiber (15 s renewals, 120 s TTL, each renewal a compare-and-swap); the old per-mutation lease CAS and per-spawn heartbeat timer are gone, and run-state writes check the observed loss instead.
  • The lease record is the kit's schema (<key>.lease.json with a schema version and a tombstone on release). Old lease.lock files are not migrated; an existing holder keeps running and a later run re-acquires, which continues the fencing generation.
  • A lease lost mid-run now surfaces as TaskLeaseLostError (taken over, released, missing or expired) instead of the old OrchestrationConflictError from the heartbeat; a refused fenced write still reports the orchestration conflict. A second start of the same Task keeps the old message: Task <id> already has an active orchestration (pid <pid>).
  • An unreadable or newer-schema lease record is refused without being overwritten (repository failure) where the old store treated it as absent and conflicted.

• #134 778fe6f Thanks @PerfectPan! - Require Node.js 22 or newer; Node.js 20 is no longer supported. The CLIs move to Ink 7, and @rivus/agent-task-loop also to execa 10; both need Node.js 22.

• #142 f5a6a3e Thanks @PerfectPan! - Support the Rivus 0.17 plugin SDK with the optional @rivus/agent peer range >=0.17.0 <0.18.0. Plugin projects must upgrade Core to the supported minor; CLI-only installations still do not require Rivus Core.

Patch Changes

• #134 778fe6f Thanks @PerfectPan! - A failing command also prints the message of each error that caused it, the unexpected-argument error mentions boolean flags only when the stray value is true or false, and the ./rivus-plugin tools report a field missing from their input as is required, for example Invalid task-create title: is required, instead of has an invalid type.

• #149 5cd6419 Thanks @nyx-cell! - Build the published JavaScript with Vite+ (vp pack). The published file set changes from fixed chunks to hashed chunks plus a separate task-provider declaration file; imports and exports are unchanged.

• #138 f34f632 Thanks @PerfectPan! - Fix the published TUIs (agent-task-loop tui and agent-finder sessions browse) crashing on start with ReferenceError: React is not defined: the build now compiles JSX with the automatic runtime.