Skip to content

CHANGELOG.md

Latest commit

 

History

History
209 lines (141 loc) · 32.9 KB

CHANGELOG.md

File metadata and controls

209 lines (141 loc) · 32.9 KB
Raw

Changelog

Unreleased

Added

  • MULTI_ACP_INACTIVITY_MS / MULTI_ACP_OVERALL_MS operator env knobs for the ACP watchdog windows (same convention as CODEX_COMPANION_TURN_INACTIVITY_MS), read from the spawn env at turn start. Live-verified end-to-end: a knob set in Claude Code's settings env reaches the forwarder subagent's companion process and bounds a hanging CLI.
  • Forwarder-contract directive in the companion's fatal-error output. Every fatal companion error now ends with a FORWARDER CONTRACT: block instructing a forwarding subagent to return the one-line failure format and NOT substitute its own answer — runtime steering at the exact decision moment for the catalogued #319-class failure.
  • HARD GATE block in all 11 forwarder agent definitions (unconditional forwarding: no task too trivial, Bash for the companion invocation only, failure line is the entire response on error), plus the same substitution ban in the multi-cli-runtime skill. Motivated by live stress tests: with the previous definitions, wrappers sometimes bypassed the companion entirely for trivial-looking questions or "helpfully" did the task themselves after a CLI failure — masking the outage from the caller.
  • Failure-mode regression tests (240 → 243): hang-at-handshake caught by the inactivity watchdog (not the 30-minute overall cap), mid-turn agent crash always yields an explicit error with partial text preserved, and the MULTI_ACP_INACTIVITY_MS knob bounds a silent agent. New fake-agent fixture flags --hang-handshake and --die-mid-turn.

Changed

  • Docs caught up to the ACP transport. README gains a Transports section documenting the MULTI_TRANSPORT_CURSOR / MULTI_TRANSPORT_OPENCODE opt-in; the multi-cli-anything and customize skills and AGENTS.md were corrected — they previously asserted "no shipped adapter uses ACP" and pointed at the legacy lib/acp-client.mjs. They now describe ACP as a live transport for Cursor + OpenCode on the SDK-based lib/acp/client.mjs, with customize documenting the transport toggle as a supported change. Also fixed a stale "Haiku forwarders" reference in AGENTS.md (all forwarders are model: sonnet). /multi:setup needed no change (it's transport-agnostic — the ACP path passes no MCP servers in-protocol, so each CLI still reads its own MCP config). Docs only; no code change.
  • The four read-only forwarder agents (cursor-explore, opencode-explore, antigravity-explorer, codex-review) moved from model: haiku to model: sonnet. Live A/B under an injected CLI outage: the haiku wrapper ignored even a point-blank in-band contract directive and substituted its own answer; the sonnet wrapper honored the failure contract. (Consistent with openai/codex-plugin-cc merging its forwarder as sonnet over a haiku proposal in PR #169; haiku's price advantage is also undercut by subagent model-pin reliability issues and subscription-quota weighting.)

Fixed

  • ACP inactivity watchdog now covers the HANDSHAKE phase. Previously it was first armed immediately before session/prompt, so a CLI that spawned and hung silently at initialize/session-new (lock, auth, network) was only caught by the 30-minute overall cap — reproduced live, then fixed: the watchdog arms at connection start and re-arms after each completed handshake step. A silent hang now errors out after inactivityMs (default 120 s) + the 5 s cancel grace.

  • Mid-turn agent crash can no longer race to a success-shaped result. A post-handshake child exit with no stopReason and no cancel now sets an explicit crash error (whichever of the exit handler or the SDK's connection-closed rejection wins the race), with the stderr tail as detail and partial streamed text preserved.

  • ACP cancel dispatch now reports transport: "process-tree" when there is no in-flight ACP turn in the calling process (the cross-process cancel case — the mechanism that actually does the work is the companion's process-tree kill; the ACP child sits inside the worker's tree). Previously it reported "acp" while doing nothing in-process, which was dishonest and also made the suite sensitive to ambient MULTI_TRANSPORT_* env (3 pre-existing headless cancel tests failed when the dogfood flags were set session-wide). transport: "acp" is now reported only when a live in-flight handle was actually cancelled in-protocol. Suite verified green both with and without the ambient flags.

0.1.2 — 2026-06-11

Added

  • ACP transport layer (slice 1): a shared, SDK-backed turn runner for the Cursor + OpenCode ACP path. New plugins/multi/scripts/lib/acp/client.mjs exports runAcpTurn(spec), a policy layer over the official @agentclientprotocol/sdk ClientSideConnection: it spawns one ACP child per turn, runs initialize → newSession → (optional set_mode / set_config_option model) → prompt, accumulates agent_message_chunk text, and resolves to a uniform { text, error, sessionId, stopReason, cancelled, modes, configOptions, usage, exitCode } (error set, never thrown — matching the adapter contract). Permission requests are auto-rejected by default and allowed only with allowWrites. Cancellation (promise.cancel() or spec.signal) sends in-protocol session/cancel, waits a 5 s grace, then process-tree-kills the child via the existing lib/process.mjs helpers; per the verified cancelled-detection rule, a cancel-requested turn is reported cancelled: true regardless of the agent's stopReason (OpenCode mislabels cancel as end_turn). Inactivity (reset on every session/update) and overall watchdogs convert hangs to a timeout error and reap the child; a child that dies before the handshake yields a spawn error with the stderr tail. A requested model that is absent from the live configOptions list fails the turn with a config error before prompting (no silent fallback). The runner imports only the vendored SDK bundle, node built-ins, and existing lib helpers, and never logs prompt content.
  • Win32-first ACP spawn resolution. plugins/multi/scripts/lib/acp/resolve.mjs exports resolveOpenCodeAcp() / resolveCursorAcp() returning { exe, args } (or { exe: null, detail } instead of throwing when not found). OpenCode resolves the platform exe under %APPDATA%\npm\…\opencode-windows-x64\bin (with a -baseline sibling probe); Cursor resolves the bundled node.exe + index.js acp under the lexically-latest ^\d{4}\. version dir in %LOCALAPPDATA%\cursor-agent\versions (it self-updates in the background). Both honor the existing OPENCODE_CLI_PATH / CURSOR_AGENT_PATH overrides and fall back to the bare binary name off-win32.
  • Vendored zero-runtime-dependency SDK bundle. scripts/build-acp-vendor.mjs (run via npm run build:acp-vendor) esbuild-bundles @agentclientprotocol/sdk + its zod peer into a single committed ESM file (plugins/multi/scripts/lib/acp/vendor/acp-sdk.bundle.mjs, external: node:*), so nothing is installed at plugin-install time. The SDK, zod, and esbuild are devDependencies only. A banner records the bundled SDK version; the build is idempotent (a second run is byte-identical).
  • Offline ACP contract test suite. test/unit/acp-client.test.mjs (27 tests) drives runAcpTurn against a fake ACP agent fixture (test/fixtures/fake-acp-agent.mjs, raw newline-delimited JSON-RPC over stdio): happy-path streaming order, set_mode / set_config_option, model-not-in-options, permission reject/allow, all three cancel paths (honored, end_turn-mislabel, ignored→tree-kill with orphan-liveness assertion), inactivity timeout, child-exit-before-handshake, unknown/tolerated sessionUpdate kinds, and a cwd-with-spaces round-trip — plus resolve.mjs unit tests (env overrides, missing-binary {exe:null}, Cursor version-dir selection) and a drift-gate test asserting the vendor bundle exists, imports, exports ndJsonStream/ClientSideConnection/PROTOCOL_VERSION, and records the SDK version pinned in package-lock.json.
  • ACP transport behind per-CLI flags (slices 2 & 3): OpenCode + Cursor adapters can run over ACP instead of headless, opt-in and off by default. Two env vars select the transport per turn (read at invoke() time, not import time, so it can be flipped per case): MULTI_TRANSPORT_OPENCODE and MULTI_TRANSPORT_CURSOR, each acp | headless (default headless). Selection lives entirely inside each adapter — the adapter contract, registry, companion, commands, and skills are untouched, and with no flag set the behavior is byte-identical to today (the headless path). On the ACP path each adapter calls the slice-1 runAcpTurn and maps its options back to the SAME result shape the headless invoke() returns (text/error/sessionId/status/fileChanges/commandExecutions/toolCalls), so the render layer is unchanged; session/update chunks are mapped to the existing onStream phase/message_chunk convention. OpenCode ACP (opencode acp): read-only roles reuse the headless buildOpencodeSpawnEnv verbatim (the OPENCODE_PERMISSION deny floor — the same constant, not duplicated JSON), and the model passes through as the same provider/model id the headless path gives --model, pinned via set_config_option. Cursor ACP (cursor-agent acp): read-only roles (research/explore/…) set session mode ask, write/delegate set agent (writes allowed); the friendly --model name is resolved against the LIVE composite-id list from session/new (exact match → unique prefix match on the segment before [, e.g. composer-2.5 → composer-2.5[fast=true]; ambiguous or no match → a config error listing the available ids, no silent fallback). To resolve the composite id pre-prompt without a second session, runAcpTurn gained a small resolveModel(availableIds, requested) callback (validated after session/new, before the prompt). Cancellation routes to the in-flight ACP turn's in-protocol cancel handle (in-process) and reports transport: "acp"; the authoritative cross-process cancel remains the companion's process-tree kill (the ACP child sits inside the worker's tree). The headless cancel/SIGTERM paths are untouched and never run on the ACP path.
  • Decoupled the slice-1 ACP layer from the legacy pre-retreat files. lib/acp/client.mjs now imports sanitizeDiagnosticMessage from a new lib/acp/diagnostics.mjs (the function copied byte-for-byte from the legacy lib/acp-diagnostics.mjs, which is left untouched), so the ACP-v2 layer no longer depends on a file slated for deletion.
  • Offline ACP-path tests for both adapters. New test/unit/opencode-acp.test.mjs (16 tests) and test/unit/cursor-acp.test.mjs (17 tests) drive each adapter's invoke() against the fake ACP agent via an injected spawnSpec (the role resolve.mjs env overrides serve in production): transport selection honors the env flag and defaults to headless; the OpenCode read-only deny floor + model pass through (asserted via a new --echo-env fixture flag); Cursor session mode by role and model alias resolution (exact, prefix-unique, ambiguous→error, missing→error with the list); result-shape parity with the headless path; and cancel routing through the in-flight handle. acp-client.test.mjs gains 3 tests for the new resolveModel callback. Suite total 204 → 240.

0.1.0 — 2026-06-02

Versioning reset. The project moved to a pre-1.0 0.x scheme to reflect that it is in active development; the earlier v2.0.0/v2.0.1 GitHub releases were removed. Entries under Pre-reset history below predate this reset and are kept as a record (their version numbers do not continue the 0.x line).

Added

  • OpenCode provider. /opencode:delegate, /opencode:research, and /opencode:explore are now shipped commands. Transport: headless opencode run --format json, piped NDJSON. The adapter (lib/adapters/opencode.mjs) parses the NDJSON event stream (step_start, text, step_finish, tool_use, error), derives file changes and command executions from completed tool_use events, and delivers the prompt on stdin (newline-safe). Read-only roles (research, explore) are enforced via injected oc-* primary agents (OPENCODE_CONFIG_CONTENT) with write/edit/bash denied plus an OPENCODE_PERMISSION deny floor — OpenCode has no --read-only flag. Write roles use --dangerously-skip-permissions. --until-done is supported; --effort is not. Default model: opencode/claude-opus-4-8 (Zen, billed separately). Token-offload caveat: anthropic/* models reuse the Claude Code subscription — zero offload; use opencode/*, openai/*, google/*, github-copilot/*, or ollama/* for real offload. MCP servers are read from OpenCode's own opencode.json (not managed by /multi:setup). Set OPENCODE_CLI_PATH to pin a specific binary; set OPENCODE_CLI_DEFAULT_MODEL to override the default model. The adapter is registered in lib/adapters/registry.mjs and the opencode plugin is listed in .claude-plugin/marketplace.json.

  • Reworked the Cursor slice into /cursor:delegate, /cursor:research, /cursor:explore (replacing /cursor:execute, /cursor:plan, /cursor:debug). delegate is agentic implementation (Cursor writes code; the calling Claude thread runs the listed ## Verification commands), research is read-only external web/docs research (Cursor's built-in WebSearch with the Exa MCP as a fallback), and explore is read-only codebase Q&A (semantic search + grep). All three default to Cursor's auto model and accept --model. delegate also gains the autonomous --until-done multi-step loop (with --max-turns), previously Codex-only — the loop's stop logic is now a shared, transport-agnostic helper (evaluateAutonomousStop).

  • Implemented the Antigravity slice on Google's headless agy CLI (EXPERIMENTAL). /antigravity:research and /antigravity:explore now run read-only against agy -p (Gemini 3.5 Flash). Because agy's headless stdout is empty upstream (gemini-cli#27466, unfixed as of agy 1.0.3), the adapter spawns agy -p, learns the conversation id from a per-invocation --log-file, and recovers the answer from agy's on-disk transcript JSONL (~/.gemini/antigravity-cli/brain/<id>/.system_generated/logs/transcript.jsonl); the last non-empty PLANNER_RESPONSE step is the answer. Auth is agy's own OAuth keyring (no API key); the desktop app is not required. Cancel is a process-tree kill. Per-call --model, write-delegate, and --until-done are intentionally unsupported on this path. New pure-helper tests in test/unit/antigravity-headless.test.mjs (against captured fixtures).

Fixed

  • Codex broker leak. The reused per-cwd app-server broker daemon now self-terminates after an idle window (CODEX_COMPANION_BROKER_IDLE_MS, default 600000 ms), so brokers spawned for transient or extra workspaces no longer linger forever (and, on Windows, no longer pin their cwd directory open). The SessionEnd hook already reaped the session's primary-cwd broker; this idle timer is the backstop for every other case (app-server-broker.mjs, lib/broker-lifecycle.mjs → shouldIdleShutdown).
  • Unreachable review-gate toggle / dangling /codex:setup references. setup was renamed to /multi:setup, but the companion, the Codex adapter, the stop-review-gate hook, the stop-gate prompt, and the codex-result-handling skill still pointed users at the non-existent /codex:setup — and /multi:setup did not expose --enable-review-gate/--disable-review-gate, so the stop-time review gate could not be toggled from any shipped command. Repointed every reference to /multi:setup and taught /multi:setup to forward the review-gate flags to the companion (which already implemented the toggle).

Changed

  • Revamped the customize and multi-cli-anything skills to match the post-split, headless shape. Both predated the Cursor ACP→headless migration, the Antigravity headless-agy slice, the shared multi-cli-runtime forwarding contract, and the companion monolith split — so they documented a buildPrompt() role→slash-prefix layer and an ADAPTERS map inside multi-cli-companion.mjs that no longer exist, and treated ACP as the default integration path. customize now teaches the real four moving parts (slash command / forwarder framing block / adapter role→flag map / shared multi-cli-runtime contract), the forwarder model-by-role policy (Sonnet for framing roles, Haiku for pure path-bridges), and headless-era escape hatches (CURSOR_AGENT_PATH, AGY_CLI_PATH, per-CLI MCP config), with ACP_TRACE demoted to legacy; its stale /cursor:research "add-a-command" example and cursor-researcher role name are replaced with current ones. multi-cli-anything now leads with headless print-mode (cursor.mjs) as the common path, documents spawn-and-read-artifacts (antigravity.mjs) and ASP (codex.mjs), demotes ACP to a clearly-labeled legacy section, and points registration at lib/adapters/registry.mjs and dispatch at lib/commands/task.mjs (not the companion). Docs-only; no runtime change.
  • Refreshed the banner and README for the four-CLI lineup. New banner art (Cursor · Antigravity · OpenCode · OpenAI Codex, replacing the stale gold banner that still showed the removed Copilot/Gemini/Qwen). The "CLIs supported" badge and intro now include OpenCode, with a note that OpenCode is in active development (its /opencode:* provider commands don't ship yet; today's working providers are Codex, Cursor, Antigravity). Also corrected the README's stale ACP_TRACE troubleshooting note — no shipped provider uses ACP.
  • Migrated the Cursor adapter from agent acp (ACP JSON-RPC) to headless agent -p. Headless fixes what ACP could not on Windows: MCP/web tools fire (they were silently dead in ACP since ~2026.04.17), cancel is a real process-tree kill (was a no-op), progress is parsed from the documented stream-json event stream, and model/mode selection are first-class flags (--model <flat-name>, --mode ask) instead of post-session RPCs — so the stale bracketed-modelId resolution is gone and auto is the default. The prompt is delivered on stdin (newline-safe). ACP (lib/acp-client.mjs) now serves only the antigravity/gemini path. Forwarders run on Sonnet (cursor-delegate, cursor-research) and Haiku (cursor-explore, a pure read-only path-bridge).
  • Forwarder subagent models tuned by role. codex-execute and codex-rescue run on Sonnet: they frame and route the prompt (choosing model/effort and shaping the task) where a more capable model materially improves the work the external CLI then does — matching the official codex-plugin-cc rescue subagent. codex-review stays on Haiku: it does no framing, only bridging the plugin boundary to forward review/adversarial-review to the companion, so the cheapest model is the correct one. (The cursor/antigravity forwarders already run on Sonnet.)
  • Repo structure for multi-agent work (no behavior change). Added AGENTS.md/CLAUDE.md orientation, ARCHITECTURE.md, an explicit adapter CONTRACT.md, a zero-dependency offline test suite (npm test, Node's built-in runner) with a reusable sandbox fixture, and a gitignored .agent/ scratch area. Began splitting the companion monolith: extracted the pure task-option normalizers (model alias, reasoning-effort validation, argv splitting) into lib/task-options.mjs with characterization tests.
  • Split the two monoliths into focused modules (no behavior change). multi-cli-companion.mjs (1462 lines) is now a ~100-line dispatcher; its command handlers moved verbatim into lib/commands/{shared,setup,jobs,review,task}.mjs. lib/adapters/codex.mjs (1110 lines) is now a ~30-line re-export barrel over codex-{roles-prompts,render-parse,transport}.mjs, keeping the public import surface and the adapter object intact. Every function was moved byte-for-byte; the offline suite grew from 27 to 82 characterization tests (all passing), and the live suite plus an adversarial verbatim diff against the prior revision confirm the CLI surface and behavior are unchanged.
  • Antigravity transport: replaced the planned desktop Language Server (ConnectRPC live-attach) with the headless agy CLI. The v3.0.0 stub targeted attaching to the running Antigravity 2.0 desktop app's LS; that approach is dropped — driving the Antigravity desktop/OAuth login from third-party software violates Google's ToS, and Google now ships the standalone agy CLI (which runs without the desktop app). The antigravity.mjs stub is replaced by a real adapter, handleCancel gained an antigravity branch, and /multi:setup now points users at installing agy and signing in (rather than running the desktop app).

Pre-reset history

The entries below predate the move to 0.x versioning and are retained for the record.

v3.0.0 — 2026-05-24

Breaking release. The provider set is now Codex, Cursor, and Antigravity. Three providers were removed and command namespaces were reorganized — there is no in-place behavioral compatibility with v2.x for the dropped CLIs. After upgrading, restart Claude Code so the subagent roster refreshes, then re-run /multi:setup.

Removed (breaking)

  • Gemini, Copilot, and Qwen providers — their plugins, adapters, subagents, and commands are gone. Gemini CLI access was cut during the gap (Gemini CLI sunset); Copilot was dropped after MSFT's billing change; Qwen was unused in practice. The Antigravity provider replaces Gemini-family access via a different transport.
  • The Gemini ACP broker lifecycle (gemini-broker-lifecycle.mjs) and the /gemini:*, /copilot:*, /qwen:* command surfaces.

Added

  • Antigravity provider — /antigravity:research (Gemini 3.1 Pro) and /antigravity:explore (Gemini 3.5 Flash), both read-only, reached through the running Antigravity 2.0 desktop app's Language Server. This release ships a stub adapter: process detection works (Windows-first), but the Language Server transport (ConnectRPC live-attach) lands in a follow-up (Phase 2). /antigravity:* commands currently return a clean "not implemented (Phase 2)" message; macOS/Linux discovery is also Phase 2.
  • Forked-and-merged the official OpenAI codex-plugin-cc into our codex slice: new /codex:rescue, /codex:review, and /codex:adversarial-review commands, the codex-rescue and codex-review subagents (with disjoint-trigger descriptions so Claude's auto-dispatch stops confusing them), and three vendored helper skills (codex-cli-runtime, gpt-5-4-prompting, codex-result-handling). All routed through our multi companion. Attribution recorded in NOTICE (Apache-2.0).

Fixed

  • Latent ENOENT in the review / stop-gate paths. The companion dispatched review/adversarial-review and the stop-review-gate hook, but the data files they read (schemas/review-output.schema.json, prompts/adversarial-review.md, prompts/stop-review-gate.md) were missing from the repo, so those paths threw at runtime. Restored the schema and prompt templates.

Changed

  • Modernized the Cursor adapter. Model selection now uses session/set_config_option (Cursor 2026.04.13+ ignores session/new.model and session/set_model). Dropped the ~/.cursor/cli-config.json allowlist injection entirely — the 2026.04.17 MCP/Terminal regression that required it was fixed upstream (forum #155544/#155516). Refreshed the current-model reference list and the known-broken-version warning.
  • Command-namespace policy. Provider plugins own their own command namespaces (/codex:*, /cursor:*, /antigravity:*); /multi:* is reserved for cross-cutting operations (setup, status, result, cancel).
  • /multi:setup detection now probes Codex, Cursor, and Antigravity (and reports a running Antigravity desktop) instead of the removed CLIs. The companion's setup report enumerates the live provider set via each adapter's isAvailable().
  • Skills (customize, multi-cli-anything, multi-cli-runtime, multi-plan-handoff, multi-result-handling) updated for the new inventory; multi-cli-anything now documents Antigravity's non-ACP Language Server (ConnectRPC) transport as a worked example of a non-ACP adapter.

Migration from v2.x

  1. Update the marketplace: /plugin marketplace update cc-multi-cli-plugin.
  2. Uninstall the dropped provider plugins if you had them: /plugin uninstall gemini@cc-multi-cli-plugin (and copilot, qwen).
  3. Reinstall the hub and the providers you want: /plugin install multi@cc-multi-cli-plugin --force, then codex / cursor / antigravity.
  4. Restart Claude Code (subagent definitions are cached at session start), then run /multi:setup.

v2.0.1 — 2026-04-26

Bug-fix release. Real-world prompts beyond a one-shot text reply silently broke before this — agents stalled, errors vanished, the forwarding subagents reported success on empty output. This release fixes the entire ACP traffic path.

Fixed

  • ACP session hangs across all CLIs. The shared ACP client now responds to incoming JSON-RPC requests from the agent (previously dropped). buildAutoApproveRequestHandler services session/request_permission, cursor/ask_question, and the full terminal/* family — without these, agents stalled forever waiting for our response.
  • Silently-dropped errors. Non-codex adapter branches now exit 0 on in-protocol errors (with the failure message in rendered output). Previously, exit 1 tripped the forwarding subagent's "if Bash fails, return nothing" rule and the user saw nothing at all.
  • Cursor agent acp Terminal hang. Plugin now auto-injects a permissive allowlist (Shell(*), Read/Write/Edit(**), MCP(*)) into ~/.cursor/cli-config.json before each Cursor invocation. Without this, Cursor's out-of-band permission gate silently stalls every execute tool call.
  • Gemini --model auto hang. Companion now treats auto as "skip session/set_model" so the CLI's native alias resolver picks a real model id. Calling set_model("auto") over ACP was silently accepted but caused session/prompt to hang.
  • MCP server schema. env is now an array of {name, value} per ACP spec (was a Record<string, string>).

Added

  • MCP wiring (Exa + Context7) into ACP session/new for all four ACP adapters (Gemini, Cursor, Copilot, Qwen). Reads keys from ~/.claude/plugins/cc-multi-cli-plugin/config.json (already populated by /multi:setup).
  • Client-side ACP terminal services (scripts/lib/acp-terminals.mjs) — terminal/create, terminal/output, terminal/wait_for_exit, terminal/kill, terminal/release backed by child_process.spawn with a 1 MiB output ring buffer. Handshake declares clientCapabilities.terminal: true.
  • Yolo / max-permission defaults. Gemini approval mode is now always yolo; Codex sandbox for --write tasks is danger-full-access; Cursor spawn includes --yolo --approve-mcps acp and explicitly sets ACP mode based on role.
  • ACP_TRACE=1 env var for full incoming-message tracing — single most useful diagnostic when an agent silently hangs.
  • One-time stderr warning when Cursor 2026.04.17-787b533 (the build with the documented MCP/Terminal regression) is detected. Auto-quiet on other versions.
  • Operator escape hatches: CURSOR_AGENT_PATH env var is now honored for pinning a specific Cursor build. Documented in the customize skill.

Changed

  • All 10 multi/agents/*.md loosened forwarding contract: capture stderr (2>&1), forbid ad-hoc polling/sleep/cat, return a structured one-line failure summary on Bash failure (instead of silently returning nothing). --write defaults added to writer-style agents (cursor-debugger, cursor-writer, qwen-writer).
  • Skills (multi-cli-anything, customize) now document the ACP gotchas we hit empirically — out-of-band permission gates, terminal capability semantics, MCP wiring quirks, mode-setting variance, version sensitivity. cursor.mjs is cited as the worked example.
  • README Known Issues section with documented Cursor 2026.04.17 upstream regressions (forum links).

Known issues (upstream, not fixable from the plugin)

  • Cursor 2026.04.17 agent acp does not send session/request_permission over the wire and silently stalls Terminal/MCP tool calls. Workaround: pre-approval via cli-config.json allowlist (auto-applied) keeps simple shell exec working; complex multi-tool runs may still hang. Pin an older build via CURSOR_AGENT_PATH if needed.

v2.0.0 — 2026-04-24

Breaking — renamed from skill-gemini to cc-multi-cli-plugin

This release fully replaces the former skill-gemini plugin. The plugin has a new name, a new repo URL (github.com/greenpolo/cc-multi-cli-plugin), a new scope (4 CLI providers, not just Gemini), and new commands. There is no in-place upgrade path.

Migration from v1 (skill-gemini):

  1. In Claude Code: /plugin uninstall skill-gemini
  2. In Claude Code: /plugin install cc-multi-cli-plugin (from github.com/greenpolo/cc-multi-cli-plugin)
  3. Run /multi:setup to configure MCPs on each CLI
  4. The old skills/gemini SKILL is gone. Its functionality is absorbed by /gemini:research and the gemini-researcher subagent.

Added

Four CLI transport adapters, three protocols:

  • Codex via App Server Protocol (ASP) — codex --app-server
  • Gemini via Agent Client Protocol (ACP) — gemini --acp
  • Cursor via ACP — agent acp
  • GitHub Copilot via ACP — copilot --acp --stdio

Eight slash commands:

  • /multi:setup — one-shot Claude-driven wizard that detects installed CLIs and configures Exa + Context7 MCPs on each
  • /gemini:research — deep research / exploration with Gemini's 1M-token context (read-only)
  • /codex:execute — delegate a specific plan step to Codex for rigorous implementation
  • /cursor:write — bulk / multi-file code writing in Cursor Agent mode
  • /cursor:plan — Cursor Plan mode for approach design (read-only)
  • /cursor:debug — Cursor Debug mode for hypothesis-driven root-cause investigation
  • /copilot:research — Copilot's /research (GitHub + web investigation)
  • /copilot:review — Copilot's /review code review agent

Four auto-dispatch subagents (Claude proactively delegates via the Agent tool):

  • gemini-researcher, codex-execute, cursor-writer, cursor-debugger

Two extension skills:

  • customize — guides Claude through rewiring which CLI handles which role (swap, disable, restrict, etc.)
  • multi-cli-anything — guides Claude through adding brand-new CLI providers (ACP, ASP, or subprocess paths)

Companion runtime (ported from OpenAI's codex-plugin-cc):

  • Shared CLI adapter registry with --cli <name> dispatch
  • Background job control (--background / --wait)
  • Session state persistence under ~/.claude/plugins/cc-multi-cli-plugin/state/
  • Session lifecycle hooks
  • Windows-safe spawn() pattern for .cmd-wrapped CLIs (Cursor, Gemini, Copilot on npm global installs)

Changed

  • Plugin name: skill-gemini → cc-multi-cli-plugin
  • License: unchanged (Apache 2.0) but LICENSE and NOTICE files added with full upstream attribution
  • Repo layout: flattened from marketplace format (plugins/skill-gemini/) to a single-plugin layout at the repo root

Removed

  • The old Gemini-only skills/gemini/SKILL.md — functionality absorbed by gemini-researcher + /gemini:research
  • The repo's former plugins/skill-gemini/ nested directory
  • The former .claude-plugin/marketplace.json marketplace manifest

Known limitations

These are explicit v2.0.0 deferrals. Filed for a future release.

  • Background task worker untested for non-Codex CLIs. The cli field is stored in the job request and threaded through executeTaskRun, so Gemini/Cursor/Copilot background jobs should work — not yet verified end-to-end.
  • --resume-last is Codex-only. Gemini/Cursor/Copilot receive the flag but have no session-resumption logic wired to the adapter. Per-invocation ACP sessions work; cross-invocation resume does not yet.
  • job-observability integration between the shared runtime and non-Codex adapters is partial. recordObserverEvent is a no-op in the Gemini/Cursor/Copilot paths. Doesn't affect correctness, does affect introspection.
  • /codex:review and /codex:adversarial-review remain in the official openai-codex plugin; our plugin has no review path for non-Codex CLIs yet. Gemini/Cursor/Copilot reviews can be invoked through each CLI's native slash command via the companion runtime but not through top-level plugin commands.
  • Setup wizard's MCP probes. /multi:setup configures MCPs on each CLI but doesn't deeply verify Exa / Context7 are reachable after configuration. Users should do a sanity check by running /gemini:research test or similar after setup.

Attribution

Apache 2.0 licensed. Major portions derived from:

  • OpenAI's codex-plugin-cc (Apache 2.0) — runtime architecture, Codex adapter, hooks
  • sakibsadmanshajib/gemini-plugin-cc (Apache 2.0) — Gemini ACP transport pattern
  • blowmage/cursor-agent-acp-npm (MIT) — Cursor ACP adapter reference

See NOTICE for the full attribution.

v1.0.0 — 2026-03 (as skill-gemini)

Original Gemini-only read-only consultation skill. See v1.0.0 git tag for history. Superseded by v2.0.0.