Skip to content

Open source Grok Bot alternative, Bring your own agent and subscription: message Claude Code, Codex, Gemini CLI, Cursor, Copilot and 40+ AI agents as bots on your own computer, from iPhone, Mac, Windows, Linux or terminal. Agent-agnostic and free. Also a Muse and Dots alternative. Built in Rust.

12.9 KB
Raw Blame

ACP adapter workarounds

Codync drives every agent through its ACP adapter (host/src/agent/acp.rs, host/src/agent/bot/). Some adapters ignore part of the protocol or behave in their own way, and the host patches around them. This page lists those patches, so that each one can be deleted once its adapter is fixed. When enough of an adapter's rows are gone, we can reconsider how Codync uses that adapter, or replace it with another implementation.

Last checked: 2026-11-02. Versions in the ACP registry then: claude-acp 0.99.1, codex-acp 2.5.0, pi-acp 0.0.34, cursor 2026.10.22, gemini 0.66.0, github-copilot-cli 1.0.105, grok-build 1.0.73, kilo 7.8.18, opencode 2.2.5, qwen-code 0.26.5.

Rules

  • A new patch gets a row in the same change. Its code comment names the defect and the upstream issue, if there is one. If upstream has no issue, open one, or note in the row why we didn't.
  • Remove a patch with its row. Do it once the fixed adapter version is the one Codync starts: the registry build (~/.codync/registry.json pins the version), or the oldest CLI we still support for harnesses that run locally. If a fix exists but users can still run an older version, keep the patch and write "fixed in X, keep until Y" in the row.
  • Check upstream when an adapter's registry version moves, and whenever a cited issue or PR closes: gh issue view <n> -R <owner/repo>, gh pr view <n> -R <owner/repo>. Then retest the defect with the new version before deleting anything.
  • Workarounds that only parse adapter-specific fields (error categories, auth _meta) are cheap. They are listed so they don't surprise anyone, but there is no hurry to remove them.

pi-acp (svkozak/pi-acp)

Defect Codync's patch Upstream Drop when
Accepts session/new mcpServers and never hands them to pi, so a pi bot had none of Codync's tools (create_bot, send_message, memory, routines) host/src/agent/bot/pi_tools.rs: PI_ACP_PI_COMMAND points to Codync's pi-codync launcher. The launcher starts pi with codync.js, which registers the servers from CODYNC_PI_MCP_SERVERS (pi.registerMcpServer, pi 0.99+). Set in host/src/agent/bot/session.rs for backend == "pi". The extension does nothing on an older pi, or when pi-acp's own bridge env (PI_ACP_MCP_SERVERS, PR #164) is set, because pi exits when an extension fails to load Open PRs #164, #145; README "Limitations" pi-acp forwards mcpServers itself. Delete host/src/agent/bot/pi_tools.rs and its calls in host/src/agent/bot/session.rs
pi skips a registered server whose name is already in its own mcp.json (disabled entries too) host/src/agent/bot/pi_tools.rs reads pi's and the project's mcp.json and renames ours to codync-<name> (- and _ count as one), through host/src/agent/bot/codex_names.rs none Codync's servers win over pi's config, or the row above is gone and pi-acp handles name clashes
pi gives up on a tool call after 60 s, before a team change's approval card can be answered host/src/agent/bot/pi_tools.rs sets each server's timeout past chat::manage::APPROVAL_TIMEOUT none With the first row
Sends banners and command lists on a timer, after it has already answered session/new host/src/agent/bot/turn.rs: a fresh session waits 400 ms, then drops what came in (for every adapter) none pi-acp sends them before its answer, or not at all

codex-acp (agentclientprotocol/codex-acp)

Defect Codync's patch Upstream Drop when
Silently leaves out a requested MCP server whose name Codex's own config uses, even a disabled one, and says so only in a notice before the first turn host/src/agent/bot/codex_names.rs: reads the Codex config files and renames the clashing server codync-<name>; host/src/agent/bot/tool_entry.rs tracks the new names Codync #217 Requested servers override the config, or codex-acp returns an error
/compact ignores its instructions host/src/agent/bot/session.rs: CODEX_CONFIG={"compact_prompt": KEEP} none /compact <text> uses the text
Sets no tool timeout for Codync's servers, so Codex's 300 s default drops a call without telling the server host/src/chat/manage.rs: APPROVAL_TIMEOUT is 270 s none A per-server timeout, or Codex sends notifications/cancelled
Sends 429s and overloads as reply text, not as errors Not patched: only usageLimitExceeded moves a Codex bot to its fallback (agent fallbacks) #647 When fixed, map the new categories in host/src/agent/fallback/reason.rs
Keeps a closed session's MCP servers until the agent stops Nothing beyond the idle stop in host/src/agent/bot/lifecycle.rs (conversation runtime) none Upstream fix: nothing to delete
Never closes the thread it titles each new session in, so that thread's MCP servers stay up until the agent stops Nothing beyond the idle stop in host/src/agent/bot/lifecycle.rs (conversation runtime) #628 Upstream fix: nothing to delete
The Codex CLI deletes the current credentials as soon as a new sign-in starts host/src/agent/term.rs: no Login step while signed in none The CLI keeps them until the new login succeeds

claude-agent-acp (agentclientprotocol/claude-agent-acp)

Defect Codync's patch Upstream Drop when
The standard systemPrompt field drops the _meta.claudeCode.options; ACP can't change a session's system prompt host/src/agent/bot/mod.rs, host/src/agent/bot/session.rs: always _meta.systemPrompt; after a compaction, session/load again to swap the prompt in none The adapter merges both, and ACP can update a system prompt
After an internal error with no category, the session stays dead ("The Claude Agent session has ended") host/src/agent/acp.rs, host/src/agent/bot/lifecycle.rs: session/close so the next turn loads it again (any adapter's uncategorized -32603) none The adapter restarts its query stream
Opens a session without checking the login host/src/agent/auth.rs: for agents with _meta.authStatus, only _auth/status_update counts none A session fails when signed out
Gives its own claude auth status only 5 s, too few on many Windows computers host/src/agent/auth.rs, host/src/agent/backends/sign_in.rs: waits 8 s, then runs the check itself for 15 s investigation A longer budget, or it always reports
claude auth status still says signed in when the token was revoked host/src/agent/backends/sign_in.rs, host/src/agent/bot/queue.rs: a prompt refused for auth wins until a sign-in finishes none The status check notices
The process keeps the credentials it started with host/src/agent/bot/queue.rs: restart after a turn that failed signing in none The adapter reads them again
Killed during an OAuth refresh, it loses the new refresh token and with it the login host/src/agent/acp.rs: close stdin first, 10 s EXIT_GRACE none Keep: a graceful stop costs nothing
A usage limit arrives as rate_limit / billing_error text host/src/agent/fallback/reason.rs: matches the Agent SDK's usage-limit prefixes none A distinct errorKind
Why Fast mode is refused is only in the option's description host/src/agent/config.rs, host/src/agent/bot/fast_mode.rs: split on " — " none A structured reason
The compaction summary rides on one of several completed compaction_updates host/src/agent/bot/updates.rs, host/src/agent/bot/compaction.rs: dedupe by compactionId, attach a late summary none One update per compaction

Cursor (cursor-agent acp)

Defect Codync's patch Upstream Drop when
session/load comes back on the last model used in any session host/src/agent/bot/session.rs: select_model again after loading none Each session keeps its own model
Lists a model's options (effort, fast…) only through cursor/list_available_models host/src/agent/auth/models.rs none Standard configOptions per model
Model variants as fixed ids (grok-4.7[reasoning_effort=high,fast=true]) host/src/agent/bot/launch.rs, host/src/agent/auth/agent.rs: _meta.parameterizedModelPicker, saved variant ids split into options none Plain models are the default and no saved variant ids are left

CodeBuddy and WorkBuddy

Defect Codync's patch Upstream Drop when
A failed prompt is stopReason: "refusal" with the error in _meta["codebuddy.ai/errorMessage"] host/src/agent/acp/codebuddy.rs: turned back into an error (host/src/agent/bot/turn.rs, host/src/chat/oneshot.rs) none A JSON-RPC error
WorkBuddy's session/new succeeds with no token host/src/agent/backends/workbuddy.rs, host/src/agent/auth.rs: asks _codebuddy.ai/getUserInfo none session/new fails when signed out
WorkBuddy's sign-in logs out a saved session, and that logout cancels the login it just started host/src/agent/backends/workbuddy.rs: moves {id}.info aside first none Signing in over a saved session works
The desktop app ships only the lite bundle; --acp alone exits at once host/src/agent/backends/workbuddy.rs: CODEBUDDY_FORCE_LITE_WB_BUNDLE=1, --serve --acp none The app ships an ACP bundle

Others

Adapter Defect Codync's patch Drop when
Hermes Starts MCP servers with a filtered environment host/src/mcp.rs: CODYNC_HOME in each server's env Keep: within the spec
Kilo _meta.terminal-auth names opencode host/src/agent/auth.rs: that method is left out Kilo names its own CLI
Qwen Code Terminal sign-in in _meta: {type: "terminal", args} instead of the spec's fields host/src/agent/auth.rs reads it The spec's fields
Antigravity Signed out, it starts its browser sign-in as soon as a session opens, so a usage reading would open a browser at every poll host/src/usage/acp_read.rs: started with BROWSER=true, which its Python honors. Windows has no such command, so host/src/usage/antigravity.rs needs a saved login file first. No issue opened yet: this only surfaced reading usage session/new doesn't start a sign-in
Antigravity Each new session leaves a conversation database behind, about 400 KB host/src/usage/acp_read.rs: readings share one session, resumed each time (usage.session.<id> in the kv). No issue opened yet: keeping sessions may be intended A session with no turn leaves nothing behind
Grok Build Not every build answers ACP's extension name _x.ai/billing, and the 1.0.75 binary names x.ai/billing host/src/usage/grok.rs: asks x.ai/billing, then _x.ai/billing. No issue opened: seen only in the binary's strings, with no login to run it The oldest build we support answers one name
oh-my-pi Tool call titles are prose ("Creating probe bot"), not the tool's name, so Codync's own calls aren't recognized as its bookkeeping Not patched (host/src/agent/bot/tool_entry.rs matches names only) Titles or _meta carry the tool's name

Gaps in ACP itself

Not one adapter's bug: the protocol leaves these open, and the host covers them for every agent.

  • No structured MCP server and tool name on a tool call: host/src/agent/bot/tool_entry.rs parses titles (mcp__x__y, mcp.x.y).
  • Every adapter names its error category differently (errorKind, category, codexErrorInfo): host/src/agent/acp.rs.
  • Agents that ignore session/cancel are killed after 15 s; a hung session/new, session/load or session/resume after 3 min; an old CLI that hangs on initialize gets 20 s when another way to start it is left (host/src/agent/bot/mod.rs, host/src/agent/bot/session.rs).
  • Agents that send no compaction_update: a drop in context use by more than half stands in for one (host/src/agent/bot/compaction.rs).
  • Mode ids have no fixed meaning: Auto is found by id or name, and Goose's auto (approves everything) is hidden (host/src/agent/modes.rs).
  • No way in ACP v1 to hand a running turn another message: Send now uses the agents' own _session/steering (advertised in the top-level _meta.steering.supported; claude-acp, codex-acp, omp) with idleBehavior: "promptRequired", and stops the turn for agents without it (host/src/agent/bot/steer.rs). Move to the standard once ACP v2's prompts during active work and the delivery RFD (agent-client-protocol#1261) are stable.