mirror of
https://github.com/OpenHands/OpenHands.git
synced 2026-10-07 16:08:23 +08:00
* feat(acp): add minimal ACP agent UI (parity with OpenHands#14401) Adds a Settings → Agent page so users can switch the conversation between the built-in OpenHands agent and an external ACP (Agent Client Protocol) subprocess (Claude Code, Codex, Gemini CLI, or custom command) without hand-editing settings. Discriminates in agent-server-adapter: when `agent_settings.agent_kind === "acp"`, build an `ACPAgent` payload (kind, acp_command, acp_model) instead of the LLM-shaped Agent, and skip the LLM defaults that would otherwise be rejected as extras. Stamps the provider key onto `tags.acpserver` so the chip can resolve a brand name from a single source. Tag-key constant note: the conventional `acp_server` form is invalid — agent-server validates tag keys against `^[a-z0-9]+$` and returns 422. The flattened `acpserver` form survives validation; the named constant `ACP_SERVER_TAG_KEY` keeps the regex and the key colocated. Gates the LLM and Condenser nav items behind a `disabledByAcp` flag, greys them out with a tooltip, and redirects to `/settings/agent` in the settings loader (not a per-route useEffect, so there is no one- frame flash of the LLM page before bouncing). E2E validated against `ghcr.io/openhands/agent-server:fa29ae2-python`: - PATCH /api/settings with `agent_kind: "acp"` round-trips - POST /api/conversations with the adapter's ACP payload returns 201, `agent.kind=ACPAgent`, `acp_command` preserved, tags stamped. Closes #412 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(acp): wire onboarding ChooseAgent step into ACP settings Drops the "Support for other agents coming soon!" banner now that the support exists. Enables the Claude Code / Codex tiles and adds a Gemini CLI tile so the four options here match ``ACP_PROVIDERS`` from the Settings → Agent page. Selecting an ACP option and clicking Next persists ``agent_kind:"acp"`` plus the registry provider key (``acp_server``) via ``useSaveSettings``, mirroring the diff the Settings page emits. The advance only happens on save success — a failed PATCH stays on the step and surfaces a toast. Skips the embedded LLM-setup step (index 2) on both forward and back navigation when an ACP agent is active: the subprocess owns its own LLM and authenticates through Secrets, so the form has nothing to configure. OpenHands path is untouched. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(acp): seamless Claude Code + Codex CLI auth via dev-docker Live-validated against `ghcr.io/openhands/agent-server:1.22.1-python` (the canvas's default pin, which now ships ACPAgent natively — no SHA override needed). Two changes surfaced by the run: 1. **Mount `~/.claude.json` in dev:docker.** Recent Claude Code CLI versions persist auth + workspace state in `~/.claude.json` next to (not inside) `~/.claude/`. Without this single-file mount, `@agentclientprotocol/claude-agent-acp` can't see the user's existing login and prompts to re-auth inside the sandbox. 2. **Use the new ACP package name in `ACP_PROVIDERS`.** Upstream renamed `@zed-industries/claude-code-acp` → `@agentclientprotocol/ claude-agent-acp`. The old name still works but emits an npm deprecation warning; the agent-server's own OpenAPI example uses the new name. Test fixtures pinning the legacy name updated to match. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * refactor(acp): import ACP_PROVIDERS from typescript-client Canvas was carrying its own copy of the ACP provider registry, which drifted out of sync with the canonical Python SDK source and ended up encoding an invalid Codex invocation (``@openai/codex acp`` — codex CLI has no ``acp`` subcommand, so the spawn deadlocked silently with ``Error: stdin is not a terminal`` and no log line). This change deletes ``src/constants/acp-providers.ts`` and imports the registry from ``@openhands/typescript-client`` instead, which now mirrors the Python SDK (see OpenHands/typescript-client#167). The TS SDK pin in ``package.json`` is bumped to the PR-branch SHA (``45a803c``) for now; once #167 merges and a new tagged release is cut, the pin can flip to the tag in a follow-up commit. Shape changes consumers needed to absorb: - ``ACPProviderConfig[]`` → ``Record<string, ACPProviderInfo>`` (lookup by key replaces ``.find``; ``Object.values`` where an array is needed) - ``display_name`` → ``displayName`` (camelCase matches TS conventions) - ``default_command`` → ``defaultCommand`` (and now ``readonly string[]``; components spread into a fresh array before passing to consumers that expect mutability) ``ACP_CUSTOM_PRESET_KEY`` is the only ACP-related constant that stays canvas-local — it's a synthetic sentinel for the "Custom" dropdown option, not a real provider, so it has no SDK counterpart. Moved to ``src/constants/acp-presets.ts``. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * revert(acp): keep ACP_PROVIDERS local to canvas Reverts the brief detour through `@openhands/typescript-client` for the ACP provider registry. Splitting the registry across two repos adds publish-coordination friction and doesn't actually eliminate the drift problem — it just moves it from "canvas vs. python-sdk" to "ts-sdk vs. python-sdk", with extra steps. Now: - `src/constants/acp-providers.ts` is the canvas-local copy again, with the corrected `codex` command (`@zed-industries/codex-acp`, the real ACP-protocol stdio server — not `@openai/codex acp`, which is the codex CLI's interactive mode and deadlocks the agent handshake when spawned without a TTY). - The package.json pin reverts to `v0.6.0` (the typescript-client release that does not include the unmerged `ACP_PROVIDERS` export from #167, which is now closed). - The split `acp-presets.ts` file is folded back in. Drift risk between this file and the Python SDK source is tracked in #587, with a longer-term plan to address it (TS-SDK mirror, code-gen, or runtime endpoint). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): resolve empty acp_command from registry in adapter PR #416 ships the Settings → Agent page (and onboarding) with a "default preset" shortcut that stores ``acp_command: []`` and trusts the agent-server to resolve it from ``acp_server``. It doesn't. The agent-server's ``ACPAgent`` model has no ``acp_server`` field and no registry resolution — it just hands ``acp_command`` straight to a subprocess spawn. Empty list trips ``acp_agent.py:1013`` with ``IndexError: list index out of range``, the agent loop dies silently inside the agent-server's run thread, and the conversation hangs in ``idle`` with the user's message persisted but never answered. No error reaches the UI; from the user's perspective they sent a message and nothing happened. Caught while exercising the live ``dev:safe`` stack: a fresh conversation seeded from the onboarding "Claude Code" tile produced ``Failed to start ACP server: list / IndexError: list index out of range`` in the agent-server log. The fix is purely client-side — expand ``acp_command`` against ``ACP_PROVIDERS`` (canvas's local mirror of the Python SDK registry, see #587) before the payload leaves the adapter, when the user picked a built-in preset. ``acp_server: "custom"`` and any unknown key are left untouched — those genuinely depend on the user's explicit command, and silently inventing one would mask a real config bug. Three new adapter tests cover: - ``acp_command: []`` + ``acp_server: "claude-code"`` → command resolved to ``["npx","-y","@agentclientprotocol/claude-agent-acp"]`` - ``acp_command`` omitted entirely + ``acp_server: "codex"`` → same resolution path - ``acp_command: []`` + ``acp_server: "custom"`` → left untouched 2273 tests pass, lint + typecheck clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): pass disabled state into SettingsDesktopSidebar When ACP is the active agent, ``useSettingsNavItems`` correctly tags the LLM and Condenser entries with ``disabled: true``. The mobile drawer (rendered via ``SettingsNavLink``) already respected that. The desktop sidebar (rendered via ``SidebarNavLink``, came in with the recent sidebar refactor) was constructing the link without forwarding the flag, so both items stayed fully clickable / styled as enabled while the conversation was running on an ACP subprocess. Two tiny changes: 1. ``SettingsDesktopSidebar`` passes ``renderedItem.disabled`` through to ``SidebarNavLink``. That alone gives the right visual state (``opacity-50``, ``pointer-events-none``) and keyboard behaviour (``tabIndex=-1`` + ``onClick preventDefault``) — both already implemented by ``SidebarNavLink``. 2. ``SidebarNavLink`` additionally sets ``aria-disabled="true"`` when ``disabled``, closing a screen-reader gap that existed independently of this regression (the link sounded actionable to assistive tech even though it wasn't). The ``clientLoader`` redirect in ``routes/settings.tsx`` continues to handle direct URL navigation to a disabled-by-ACP page, so even if someone bookmarks ``/settings/condenser`` and lands there while ACP is active, they get bounced to ``/settings/agent``. Two new tests in ``settings-navigation.test.tsx``: - Disabled-by-ACP items in the desktop sidebar carry ``aria-disabled``. - Enabled items don't. 2275 tests pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): address PR #416 review feedback Addresses both human and all-hands-bot review comments on #416: **Critical bugs fixed** - ``acp_args`` duplication on load (bot critical #1): the textarea is the single source of truth for the launch tokens, but save only wrote ``acp_command`` — any API-set ``acp_args`` survived and concatenated at spawn time. Save now always writes ``acp_args: []``. - ``tokenizeCommand`` corrupted quoted Custom commands (human bug #2): ``bash -c "echo hello"`` got split into ``["bash","-c","\"echo","hello\""]`` and silently misbehaved. New ``src/utils/acp-command.ts`` wraps ``shell-quote`` with selective re-quoting (so ``npx -y @org/pkg`` renders verbatim, not ``\@org/pkg``) and filters non-string entries (redirects, env-var refs) out of the parsed argv. Round-trip tests pin the contract. - Loader/component settings cache mismatch (bot critical #4): loader used ``SETTINGS_QUERY_KEYS.byScope("personal")``; ``useSettings`` used ``[...byScope("personal"), backend.id, orgId]``. They didn't share cache. Aligned + set ``staleTime: 0`` on the loader read so cross-tab kind flips are picked up immediately (the in-render hook keeps its 5-minute stale window). - ``getFirstAvailablePath`` ignored the new agent route (human bug #3): ``/settings/agent`` now precedes the others in the fallback list, so first-time / hide_llm_settings users land on the agent picker rather than ``/settings/app``. - ``ACP_SETTINGS_KEYS`` documentation (human #4): pre-empts the "why not trim this list to UI-visible fields" question by spelling out that it serves as both the ACP allow-list and the OpenHands deny-list — trimming would silently leak API-set ``acp_*`` state. **Refactor (human #2 + #3)** - ``description_key`` moves into ``ACP_PROVIDERS``; the onboarding ``AGENT_OPTIONS`` is now derived from the registry so adding a new provider only needs one edit. - One ``buildAcpAgentSettingsDiff`` helper replaces the two near-copies in ``choose-agent-step.tsx`` and ``agent-settings.tsx``; both call sites are now under a single contract for the agent_settings_diff shape. **UX (bot)** - Onboarding progress bar shows the actual visited-step count when the LLM step is skipped (3 segments for ACP, 4 for OpenHands). Previously segment 2 popped "completed" on a slide the user never visited. **Test coverage (bot)** - New ``__tests__/utils/acp-command.test.ts`` covers parseCommand / formatCommand round-trips, quoted args, embedded escapes, shell- operator filtering, package-style tokens. - Adapter: empty ``acp_model: ""``, unknown ``acp_server`` key, ACP→OH→ACP round trip (no field leakage either direction). - agent-settings: cleared input keeps Save disabled, whitespace-only same, full Custom command with quoted args round-trips through shell-quote. - choose-agent-step: provider switching (claude-code → codex) rebuilds the diff cleanly, no leak from the prior selection. **Acknowledged (no action)** - Bot critical #2 (supply chain drift) — same problem as the existing agent-canvas#587, already tracked. - Bot critical #3 (desktop sidebar disabled) — fixed in 27a3e79 a few commits before this review was written; review snapshot was stale. - Translation duplication (human #1) — matches the existing ``translation.json`` convention (every key has all 15 locales). - Option-bag → split functions (human #5) — cosmetic; defer. 2292 tests pass, lint + typecheck clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): pre-bundle shell-quote so the Vite dev server can load it ``shell-quote`` is a CommonJS module that does ``module.exports = { parse, quote }``. The previous commit wired it into ``src/utils/ acp-command.ts`` with a named ESM import, which the dev server rejected on the first ``agent-settings.tsx`` load: SyntaxError: The requested module '/node_modules/shell-quote/ index.js?v=...' does not provide an export named 'parse' Switching to a namespace import (``import * as shellQuote from "shell-quote"; const { parse, quote } = shellQuote;``) makes the named-export check pass, but Vite then served the raw CJS file to the browser unchanged and the next request died with: ReferenceError: exports is not defined This second failure is because ``vite.config.ts`` sets ``optimizeDeps.noDiscovery: true`` — new dependencies must be listed in ``optimizeDeps.include`` or Vite won't run them through its CJS-to-ESM prebundler. Adding ``"shell-quote"`` there fixes it; the existing entry has a comment block explaining the same constraint for other deps. The Rollup-based prod build was unaffected. Verified: dev server boots clean, ``GET /settings/agent`` returns 200, no ``exports is not defined`` in the Vite client log, 9 unit tests in ``__tests__/utils/acp-command.test.ts`` pass on the Node test runner (vitest) where the CJS interop already worked. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): address second-pass review on PR #416 Addresses the second all-hands-bot review's critical + improvements: **Critical: load path mis-merged acp_command + acp_args** Settings stored with the registry-default shortcut (``acp_command: []``, ``acp_server: "claude-code"``) plus a non-empty ``acp_args`` showed only the args in the textarea — no registry prefix. Saving then sent ``acp_command: ["--extra-arg"]`` and flipped the preset to ``custom``, silently losing the ``npx -y @agentclientprotocol/ claude-agent-acp`` prefix. The fix expands the registry default *before* concatenating with args, so the textarea always shows the full launch command and round-trips cleanly. **Improvement: formatCommand drops empty-string args** ``formatCommand(["bash", "-c", ""])`` rendered as ``"bash -c "`` which parsed back to ``["bash", "-c"]``, silently losing the empty slot. Now quotes empty tokens explicitly so they survive. **Improvement: desktop sidebar disabled tooltip parity** Mobile drawer's ``SettingsNavLink`` already showed "Disabled while {agentName} is active" on greyed-out items; the desktop ``SidebarNavLink`` had no explanation. Added a ``disabledReason`` prop (i18n-agnostic; the caller formats the string) and wrap with ``StyledTooltip`` when disabled-with-reason. ``SettingsDesktopSidebar`` now forwards the formatted message — same UX on both surfaces. **Test coverage gaps the bot flagged** - ``agent-settings``: new regression guard for ``acp_command:[]`` + non-empty ``acp_args`` load (would have caught the critical bug above). - ``acp-command``: empty-string round-trip case + explicit assertion; five more shell-operator filters (pipe, ``;``, ``&&``, ``||``, ``>>``). - ``settings-navigation``: desktop sidebar wraps disabled items in StyledTooltip when ``disabledReason`` is supplied; not when omitted. Plus a clean merge from ``origin/main`` (one-line import conflict in ``agent-server-adapter.ts``). 2350 tests pass, lint + typecheck clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): address third-pass review on PR #416 - parseCommand: try/catch around shell-quote.parse so a malformed command in the textarea can't crash Settings → Agent mid-render - Rewrite shell-metasyntax tests to pin the *actual* shell-quote behaviour (it's a parser, not a security filter) — operators, globs, and comments are dropped; backticks / $VAR / $(...) survive as literal tokens but are NOT expanded at parse time - Add npm URLs + verification date (2026-05-19) to each ACP_PROVIDERS entry so future maintainers can re-check upstream packages - Document the silent preset-switch behaviour on detectPreset (the dropdown follows the textarea; the textarea is the source of truth) - Use the exported ACP_SERVER_TAG_KEY constant in the adapter test so a rename surfaces as a compile error rather than a runtime schema mismatch - Restore the canonical typescript-client lock entry (drop the git+ssh:// + SHA bump that crept in from a local npm install) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): don't expose LLM-switch UI on ACP conversations The SDK's ACPAgent carries a sentinel ``llm`` (``acp-managed``) for cost-attribution only — the real model lives on the ACP subprocess via ``acp_model`` and isn't visible on ``agent.llm.model``. Without this fix, ``toAppConversation`` surfaced the sentinel as the conversation's ``llm_model``, and the chat header's SwitchProfileButton happily let users "change the model" while the running Claude-Code / Codex / Gemini subprocess kept its own. A confusing silent no-op. Two layers of defence so no future consumer has to re-derive the rule: 1. Boundary normalisation: ``toAppConversation`` reads the pydantic discriminator (``info.agent.kind === "ACPAgent"``), surfaces it as ``agent_kind: "acp" | "openhands"`` on AppConversation, and nulls ``llm_model`` for ACP. Mirrors OpenHands PR #14401. 2. UI gate: SwitchProfileButton returns null when ``conversation.agent_kind === "acp"``. The right control for ACP model switching is the ``acp_model`` field on Settings → Agent, not this picker. Tests cover both: a new ``toAppConversation`` case asserts ``agent_kind === "acp"`` + ``llm_model === null`` for an ``{kind: "ACPAgent"}`` payload, and a new SwitchProfileButton case asserts the button hides for an ``agent_kind: "acp"`` conversation even when profiles are present. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): bridge Settings → Secrets into the ACP subprocess env The bare ``payload.secrets`` channel lands in the agent-server's ``secret_registry`` server-side, which the OpenHands ``Agent`` reads directly — but ``ACPAgent._start_acp_server`` builds its subprocess env from ``agent_context.secrets``, not from the registry. Without a bridge, a Settings → Secrets entry like ``ANTHROPIC_API_KEY`` is silently invisible to the ACP CLI (Claude Code, Codex, Gemini), so users hit "authentication failed" with no on-screen hint that their configured secret never reached the subprocess. Mirror the same LookupSecret map onto ``payload.agent.agent_context.secrets`` when ``acpMode === true``, so the agent-server's existing env-injection loop picks them up. The bare ``payload.secrets`` channel is also kept (it serves other consumers + remains the canonical "conversation secrets" wire). The mirroring fires only when there's something to bridge; non-ACP payloads are unchanged. This is a shim. Once canvas pins to an agent-server build that includes software-agent-sdk PR #3299 (which teaches ACPAgent to also read from ``state.secret_registry``), the ``if (acpMode)`` branch can be deleted with no behaviour change. Tests: - New: ACP payload mirrors customSecrets onto agent_context.secrets - New: empty customSecrets does NOT synthesize an empty bridge map - New: non-ACP payload does NOT get an agent_context.secrets bridge - All 42 adapter tests pass Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): hide MCP nav + cloud LLM-model fallback while ACP is active Two ACP-leak fixes the review surfaced: MCP page reachable + editable under ACP - The SDK's ``ACPAgent`` rejects ``mcp_config`` on init (acp_agent.py:845) and the canvas adapter already strips it from start payloads, but the /mcp route and the Extensions nav still let users add / edit / delete MCP servers — silent no-ops against the running subprocess. - Add a ``clientLoader`` on /mcp that bounces to /settings/agent when ``agent_kind === "acp"``. Grey out the MCP item in ExtensionsNavigation with the same explanatory tooltip the LLM / Condenser items already use under ACP. - Extract the redirect into ``utils/acp-route-guard.redirectIfAcpActive`` so /settings and /mcp share one cache-key + redirect-target definition. settings.tsx's clientLoader now calls into it. Cloud chat ``ChatInputModel`` falls back to ``settings.llm_model`` for ACP - ``toAppConversation`` writes ``llm_model: null`` on ACP conversations (commit 8f0efe62), but ChatInputModel did ``conversation?.llm_model ?? settings?.llm_model``, resurrecting the user's default OpenHands model on a Claude-Code conversation and linking to /settings (which is itself ACP-disabled). Gate on ``conversation?.agent_kind === "acp"`` and return null instead. Tests: - New: ExtensionsNavigation greys MCP under ACP, leaves Skills + non-ACP clickable - New: /mcp clientLoader redirects under ACP, returns null otherwise + on settings-fetch errors (no redirect-loop) - New: ChatInputModel returns null for ACP even when settings has a model - All 18 affected tests pass Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): preserve unknown acp_server on no-op saves + reviewer cleanups Fourth-pass review (PR #416 review comment 4486154133). Triage: Fixed: - **Unknown ``acp_server`` demoted to ``"custom"`` on save** (P1, real data corruption). A user with ``acp_server`` set out-of-band to a provider canvas's registry doesn't carry yet (e.g. a future provider, or one removed from the local mirror) would open Settings → Agent and lose the original key on the next Save — ``detectPreset`` routes every unknown server to ``ACP_CUSTOM_PRESET_KEY``. Now we capture the loaded ``acp_server`` + textarea at load time, and on save — when both are unchanged and the loaded key is non-empty, non-``"custom"``, and absent from ``ACP_PROVIDERS`` — pass it back verbatim via a new ``allowUnknownServer`` opt on ``buildAcpAgentSettingsDiff``. Editing the command still demotes to ``"custom"`` (user is configuring a new thing, so the preset name follows the command). - **Dead ``...existingContext`` spread** in the ACP secret bridge. ``createAgentFromSettings`` never populates ``agent_context`` on the ACP branch, so the spread always merged into ``{}``. Direct assignment — and a comment explaining why a deep-merge would be the wrong direction (ACPAgent only treats ``secrets`` as acp_compatible). - **Misleading ``$VAR`` test comment**. Reworded to lead with the no-leak contract (host env values must not end up in the persisted ``acp_command``) rather than the implementation-detail tangent. Documented but not changed: - **``acp_args: []`` "data loss" concern** — false alarm. Load merges ``acp_command + acp_args`` into the textarea before render; save persists the merged tokens as ``acp_command`` with ``acp_args: []``. Round-trip is correct. Added an inline comment on the load merge so the next reviewer doesn't re-flag the reset. - **Silent preset migration without user feedback** — by design. The dropdown re-derives from the textarea so it always reflects what will be saved; adding a toast on every keystroke would be noise. Already documented as intentional on ``detectPreset``. Tracked elsewhere: - Supply-chain drift / npm verification: agent-canvas#587. - Gemini onboarding icon: agent-canvas#621. Tests: - New: ``preserves an unknown loaded acp_server when the user saves without editing`` - New: ``demotes an unknown loaded acp_server to 'custom' when the user edits the command`` - All 73 affected tests pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): stop silently corrupting argv, restore AgentContext, fix home-screen gating Three real review findings, all wired: 1. ``parseCommand`` silently dropped URL tokens with ``?`` query strings (and any other shell-glob metacharacter). Reproducer: node acp.js --endpoint https://example.com/acp?tenant=abc ``shell-quote.parse`` read ``?tenant=abc`` as a glob pattern and emitted a non-string AST node; the ``.filter(string)`` then dropped the URL entirely, persisting ``["node","acp.js","--endpoint"]``. Replaced ``shell-quote.parse`` with a small custom argv tokenizer that handles single/double quotes + backslash escapes and treats every other character — ``?``, ``*``, ``$``, ``|``, ``>``, ``#``, ``&``, ``;``, ``(``, ``)``, backticks — as literal. The agent-server passes the argv straight to ``subprocess.create_subprocess_exec`` anyway (no shell intermediary), so the literal-only model matches what actually happens at spawn time. ``shell-quote.quote`` is still used by ``formatCommand`` for output. 2. The ACP path skipped the ``agent_context`` block that the OpenHands path seeded with ``load_public_skills`` / ``load_user_skills`` / optional ``system_message_suffix``. All three are marked ``acp_compatible: true`` on the SDK ``AgentContext`` model — the ACP CLI renders them via ``ACPAgent._render_suffix`` — so ACP conversations were silently shipping a smaller system prompt than OpenHands ones. ``createAgentFromSettings`` now seeds the same block on both branches. The secret bridge below merges into that block (was overwriting it) so ``{ secrets }`` no longer wipes the skill flags. 3. ``ChatInputModel`` and ``SwitchProfileButton`` only checked ``conversation?.agent_kind``. On the home screen (and during the task-startup window) ``conversation`` is undefined, so the per-conversation check missed and both surfaces fell back to ``settings.llm_model`` / the LLM-profile picker — even when ``settings.agent_settings.agent_kind === "acp"`` made it clear the next-created conversation would be ACP. Added a settings fallback so both controls hide consistently with the rest of the ACP nav gating. Tests: - parseCommand: new "preserves URLs with query strings" + "preserves URLs with multiple query params" + "preserves shell metacharacters as literal argv tokens" cases; the old "filters operator" cases flipped to "preserves operator as literal". 17 parseCommand cases pass. - adapter: assertion on the ACP payload's ``agent_context`` updated to expect the skill flags instead of ``undefined``. - chat-input-model + switch-profile-button: new "hides on the home page when ACP is the default agent" cases. - 83 tests pass across the 5 affected files. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): two chat-rendering UX glitches on streaming ACP tool calls 1. Half-formed ACP tool-call cards flashed in the chat before the final state arrived. ACP servers stream multiple events per ``tool_call_id`` (status flips ``in_progress`` → ``completed`` / ``failed``); the intermediate events carry partial ``raw_input`` / ``raw_output`` / ``title``. The previous gate suppressed only ``in_progress`` and let ``null`` through (a "backwards compat" carve-out for older agent-server builds that no longer apply at our pinned version). Streaming intermediates often arrive without a status set yet, so they leaked. Tighten ``shouldRenderEvent`` to require ``status === "completed" || "failed"``. ``handleEventForUI`` already collapses by ``tool_call_id`` in place, so the terminal event lands at the original position once it arrives — no flash, no double-render. 2. "Reading Read /Users/foo/bar" — Claude Code emits titles like ``"Read /Users/foo/bar"`` for a read tool, and our i18n template ``"Reading <cmd>{{title}}</cmd>"`` then doubles up the verb. Add ``stripRedundantTitlePrefix`` keyed by ``tool_kind``: read → strip ``"Read"``, edit → strip ``"Edit"`` / ``"Write"``, execute → strip ``"Bash"`` / ``"Run"``, fetch → strip ``"Fetch"`` / ``"WebFetch"``. Boundary-checked via trailing whitespace so a token like ``"Reads-from"`` is left alone. English-only on purpose: ACP servers are anglophone and emit english titles regardless of the user's canvas locale; matching translated verbs would go stale the moment a new server is added. Titles already lacking a redundant prefix (the OpenHands ACP wrapper, future servers) round-trip verbatim — the strip is a no-op there. Tests: - ``shouldRenderEvent``: ``null`` status now flips to false + comment explains why (treated as in-flight, not legacy). - New ``stripRedundantTitlePrefix`` describe block covers the four tool kinds, the no-op case, the word-boundary guard, ``tool_kind: null`` (no strip), and empty titles. - 53 tests pass across the conversation-events helpers. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(acp): drop React Router type import on /mcp clientLoader (CI build) CI's ``build:lib`` failed with: src/routes/mcp.tsx(3,23): error TS6059: File '.../.react-router/types/ src/routes/+types/mcp.ts' is not under 'rootDir' '/src'. ``tsconfig.lib.json`` sets ``rootDir: "src"`` and pulls in ``src/components/**/*.tsx``. ``src/components/settings/index.ts`` re-exports from ``routes/mcp-settings``, which imports ``routes/mcp`` — so the lib's typecheck graph reaches ``routes/mcp.tsx`` and trips on the generated ``./+types/mcp`` import that lives under ``.react-router/types/``, outside the lib's rootDir. (``routes/ settings.tsx`` uses the same import pattern but isn't reachable from the lib graph, which is why local typecheck passed.) Drop the type import and declare the loader with no parameters — matches the existing ``index-redirect`` and ``mcp-settings-redirect`` loader pattern. Test calls collapsed to ``clientLoader()`` to match the new signature. ``npm run build:lib`` now passes locally; 248 tests pass across the affected suites. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(acp): tighten adapter comments around SDK refs Two reviewer-flagged comment fixes: - ``ACP_SETTINGS_KEYS`` docblock no longer claims there's a matching ``ACP_SETTINGS_KEYS`` constant in the Python SDK (there isn't; the fields are model attributes on ``ACPAgentSettings``). Reworded to "Keep aligned with the ``acp_*`` fields on ``ACPAgentSettings`` in ``openhands-sdk/openhands/sdk/settings/model.py``" with an explicit "no matching SDK constant — hand-maintained" note, and cross-linked to the existing #587 drift tracker. - ``createAgentFromSettings`` now spells out where the ``acp_compatible`` markers live on each of the three fields we set (``system_message_suffix`` L66, ``load_user_skills`` L80, ``load_public_skills`` L89 in ``openhands-sdk/openhands/sdk/context/agent_context.py``) plus what happens when a future SDK bump drops one (422 at conversation start → drop the demoted field, don't wrap a workaround). Line refs are brittle by design — they're the tripwire that surfaces a regression here rather than in production. No behaviour change; lint + 42 adapter tests pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Debug Agent <debug@example.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
199 lines
8.2 KiB
TypeScript
199 lines
8.2 KiB
TypeScript
import { describe, expect, it } from "vitest";
|
|
import { formatCommand, parseCommand } from "#/utils/acp-command";
|
|
|
|
describe("parseCommand", () => {
|
|
it("splits a simple npx invocation into argv tokens", () => {
|
|
expect(
|
|
parseCommand("npx -y @agentclientprotocol/claude-agent-acp"),
|
|
).toEqual(["npx", "-y", "@agentclientprotocol/claude-agent-acp"]);
|
|
});
|
|
|
|
it("respects double-quoted segments — the headline regression .split fix", () => {
|
|
// The old `.split(/\s+/)` implementation turned this into
|
|
// ``["bash", "-c", "\"echo", "hello", "world\""]`` and the spawn
|
|
// would either misbehave or fail in a confusing place. The
|
|
// quote-aware tokenizer keeps the quoted segment intact.
|
|
expect(parseCommand('bash -c "echo hello world"')).toEqual([
|
|
"bash",
|
|
"-c",
|
|
"echo hello world",
|
|
]);
|
|
});
|
|
|
|
it("respects single-quoted segments and embedded whitespace", () => {
|
|
expect(parseCommand("env FOO='bar baz' npx -y my-acp")).toEqual([
|
|
"env",
|
|
"FOO=bar baz",
|
|
"npx",
|
|
"-y",
|
|
"my-acp",
|
|
]);
|
|
});
|
|
|
|
it("preserves URLs with query strings — the headline shell-quote-glob fix", () => {
|
|
// Regression guard for the silent-corruption bug:
|
|
//
|
|
// node acp.js --endpoint https://example.com/acp?tenant=abc
|
|
//
|
|
// ``shell-quote.parse`` used to read ``?tenant=abc`` as a glob
|
|
// pattern and drop the entire URL token, so the saved
|
|
// ``acp_command`` became ``["node", "acp.js", "--endpoint"]``.
|
|
// The spawn would then fail with a confusing "missing endpoint"
|
|
// error far from the Settings → Agent page that caused it.
|
|
//
|
|
// The custom tokenizer treats ``?`` as a literal — same for
|
|
// every other shell metacharacter. The URL round-trips intact.
|
|
expect(
|
|
parseCommand("node acp.js --endpoint https://example.com/acp?tenant=abc"),
|
|
).toEqual([
|
|
"node",
|
|
"acp.js",
|
|
"--endpoint",
|
|
"https://example.com/acp?tenant=abc",
|
|
]);
|
|
});
|
|
|
|
it("preserves URLs with multiple query params", () => {
|
|
// ``&`` is also literal — same reason.
|
|
expect(parseCommand("curl https://x.com?a=1&b=2")).toEqual([
|
|
"curl",
|
|
"https://x.com?a=1&b=2",
|
|
]);
|
|
});
|
|
|
|
it("preserves shell metacharacters as literal argv tokens", () => {
|
|
// Pipes, redirects, semicolons, glob chars, ``$``, backticks,
|
|
// ``#`` all round-trip as literal characters within the surrounding
|
|
// token. The agent-server uses ``subprocess.create_subprocess_exec``
|
|
// (no shell intermediary), so a user typing ``foo | bar`` is
|
|
// configuring two literal argv entries — not a shell pipeline.
|
|
// The user's helper text steers them to ``bash -c '…'`` if they
|
|
// actually want shell features.
|
|
expect(parseCommand("foo | bar")).toEqual(["foo", "|", "bar"]);
|
|
expect(parseCommand("foo > log.txt")).toEqual(["foo", ">", "log.txt"]);
|
|
expect(parseCommand("foo *.txt")).toEqual(["foo", "*.txt"]);
|
|
expect(parseCommand("foo $X")).toEqual(["foo", "$X"]);
|
|
expect(parseCommand("foo `bar`")).toEqual(["foo", "`bar`"]);
|
|
expect(parseCommand("foo # comment")).toEqual(["foo", "#", "comment"]);
|
|
expect(parseCommand("foo && bar")).toEqual(["foo", "&&", "bar"]);
|
|
expect(parseCommand("foo; bar")).toEqual(["foo;", "bar"]);
|
|
});
|
|
|
|
it("treats blank input as an empty argv", () => {
|
|
expect(parseCommand("")).toEqual([]);
|
|
expect(parseCommand(" \t\n ")).toEqual([]);
|
|
});
|
|
|
|
it("honors backslash escapes outside quotes", () => {
|
|
// ``foo\ bar`` is one token containing a literal space — the same
|
|
// contract POSIX shells provide. Lets the user type paths with
|
|
// spaces without reaching for quotes.
|
|
expect(parseCommand("foo\\ bar")).toEqual(["foo bar"]);
|
|
// An escaped quote becomes a literal quote in the token.
|
|
expect(parseCommand('foo\\"bar')).toEqual(['foo"bar']);
|
|
});
|
|
|
|
it('honors ``\\\\`` and ``\\"`` inside double-quoted segments', () => {
|
|
expect(parseCommand('bash -c "echo \\"hi\\""')).toEqual([
|
|
"bash",
|
|
"-c",
|
|
'echo "hi"',
|
|
]);
|
|
expect(parseCommand('"foo\\\\bar"')).toEqual(["foo\\bar"]);
|
|
});
|
|
|
|
it("does not env-expand $VAR refs — keeps them as literal", () => {
|
|
// The forbidden outcome would be the tokenizer reading
|
|
// ``process.env.ANTHROPIC_API_KEY`` and inlining its value into
|
|
// the persisted ``acp_command`` — that would leak a host env var
|
|
// into settings on every save. The tokenizer reads ``$NAME`` as
|
|
// a literal substring of the token, so the user's typed text
|
|
// survives verbatim. Users who actually want env vars in the
|
|
// subprocess should set ``acp_env`` instead of inlining them.
|
|
const result = parseCommand("npx $ANTHROPIC_API_KEY");
|
|
expect(result).toEqual(["npx", "$ANTHROPIC_API_KEY"]);
|
|
// Pin the no-leak contract: no ``sk-…`` token sneaks through
|
|
// from the host env (which is also unset here, but still).
|
|
expect(result.some((t) => /sk-ant-/.test(t))).toBe(false);
|
|
});
|
|
|
|
it("does not run subshells: $(…) and backticks become literal tokens", () => {
|
|
// The forbidden outcome would be executing ``date`` and inlining
|
|
// today's timestamp into the persisted command. The tokenizer
|
|
// never invokes anything; both forms round-trip verbatim.
|
|
expect(parseCommand("echo $(date)")).toEqual(["echo", "$(date)"]);
|
|
expect(parseCommand("echo `date`")).toEqual(["echo", "`date`"]);
|
|
});
|
|
|
|
it("survives unterminated quotes without throwing", () => {
|
|
// EOF closes the open quote; the partially-built token gets
|
|
// pushed. A throw here would crash the Settings → Agent page
|
|
// mid-render. The Save button gates on a non-empty argv anyway,
|
|
// so a recoverable miss can't be silently persisted.
|
|
expect(parseCommand('bash -c "unterminated')).toEqual([
|
|
"bash",
|
|
"-c",
|
|
"unterminated",
|
|
]);
|
|
expect(parseCommand("foo 'unterminated single")).toEqual([
|
|
"foo",
|
|
"unterminated single",
|
|
]);
|
|
});
|
|
});
|
|
|
|
describe("formatCommand", () => {
|
|
it("renders package-style tokens verbatim, no escaping of @ or /", () => {
|
|
// The textarea is the only consumer of formatCommand. Escaping the
|
|
// ``@`` in ``@org/pkg`` produces a hostile read-back (the user
|
|
// copies their existing command, the textarea now shows
|
|
// ``\@org/pkg``, they think we corrupted it). The agent-server
|
|
// execs argv directly so the escape isn't load-bearing for
|
|
// behaviour — only for display.
|
|
expect(
|
|
formatCommand(["npx", "-y", "@agentclientprotocol/claude-agent-acp"]),
|
|
).toBe("npx -y @agentclientprotocol/claude-agent-acp");
|
|
});
|
|
|
|
it("shell-quotes tokens that contain whitespace", () => {
|
|
expect(formatCommand(["bash", "-c", "echo hello world"])).toBe(
|
|
"bash -c 'echo hello world'",
|
|
);
|
|
});
|
|
|
|
it("round-trips arbitrary argv arrays through parseCommand", () => {
|
|
const cases: string[][] = [
|
|
["npx", "-y", "@agentclientprotocol/claude-agent-acp"],
|
|
["npx", "-y", "@zed-industries/codex-acp"],
|
|
["npx", "-y", "@google/gemini-cli", "--acp"],
|
|
["bash", "-c", "echo hello world"],
|
|
["env", "FOO=bar baz", "npx", "-y", "my-acp"],
|
|
["./bin/my-agent", "--flag=value"],
|
|
// URL with query string — the headline silent-corruption case.
|
|
["node", "acp.js", "--endpoint", "https://example.com/acp?tenant=abc"],
|
|
// URL with multiple params.
|
|
["curl", "https://x.com?a=1&b=2"],
|
|
// Empty-string tokens are rare but valid (some CLIs treat an
|
|
// empty positional as "no argument supplied" rather than missing).
|
|
// Without explicit quoting in formatCommand they round-trip back
|
|
// as fewer tokens, silently dropping the empty slot.
|
|
["bash", "-c", ""],
|
|
["program", "", "--flag"],
|
|
];
|
|
for (const argv of cases) {
|
|
expect(parseCommand(formatCommand(argv))).toEqual(argv);
|
|
}
|
|
});
|
|
|
|
it("renders an empty argv as an empty string", () => {
|
|
expect(formatCommand([])).toBe("");
|
|
});
|
|
|
|
it("explicitly quotes empty-string tokens so they survive the round trip", () => {
|
|
// Direct assertion on the rendered form — without this rule,
|
|
// formatCommand(["bash","-c",""]) would render ``"bash -c "`` and
|
|
// parseCommand would return ``["bash", "-c"]``, losing the empty arg.
|
|
expect(formatCommand(["bash", "-c", ""])).toBe("bash -c ''");
|
|
});
|
|
});
|