Skip to content

feat: add Google Antigravity ACP harness - #37974

Merged
bolichen97 merged 2 commits into
mainfrom
crew/CommunityCrew/antigravity-harness-26171
Nov 16, 2026
Merged

bolichen97 merged 2 commits into
mainfrom
crew/CommunityCrew/antigravity-harness-26171

Conversation

@bolichen97

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: Let Kiro Crew users select Google Antigravity's standalone ACP runtime alongside Kiro CLI, Claude Code, Codex, OpenCode, and the other supported harnesses.

Google Antigravity publishes an ACP server in the ACP Registry, but Kiro Crew cannot currently expose it as an agent.acp_backend choice, resolve its two-file distribution, declare its host capabilities, or pin its permission behavior. Users who want to drive their Google-backed coding agent from Kiro Crew have to leave Crew and use a separate ACP client.

Why it matters

The existing harness switch already supports multiple vendors. Adding Antigravity makes Google's agent available through the same dashboard, CLI, channels, scheduling, MCP tools, hooks, and approval flow as the other coding agents. It also keeps Crew's harness seam honest: the host contract, install probe, model/effort options, auth mask, and permission routing are declared instead of inferred.

Not a goal

  • Not a new provider abstraction or a second model API path.
  • Not support for the interactive agy terminal UI; this uses the standalone agy_acp_server ACP runtime.
  • Not bundling or auto-installing the Antigravity archive.
  • Not claiming enterprise-policy acceptance without an enterprise tenant on the recording host.
  • Not enabling session sharing, steering, compaction, Side Chat read-only execution, or member capability loading for this harness.

What changed (motivation → approach → change)

In-repo move of #26171. Supersedes #26171 for the in-repo Sensitive Change Review; author LuisBrel. The fork PR touches agent_sdk/tool_gate.py and scripts/, so its Sensitive Change Review cannot pass from a fork. This branch is LuisBrel's single commit (2eeb923502, author kept) cherry-picked onto current main. The only conflict was in docs/system-specs/modules/acp-client.md: main's Claude bullet (with the hook_runtime/hook_manager.py path) is kept and the Antigravity bullet is re-added. No code changed in the move.

Antigravity uses the existing per-session ACP seams. The launcher resolves agy_acp_server and its sibling helper together, sends MCP servers through its own _antigravity_session_mcp_servers seam in the shared session array (gated on Antigravity, like the copilot, deepseek and pi seams), and uses the advertised model and thinking-level options. Defaults stay unchanged. The dashboard catalogs and glossary include Google Antigravity.

Only ACP 1.4.4 is verified; missing or different versions are refused. The backend card, install hint and sign-in remedy say so. Permission mode stays default; modes that suppress asking are refused. Prompt commands are limited to plan and usage.

Antigravity joins the set whose unstartable stdio MCP servers are pruned before session/new: measured on 1.4.4 signed out with an invalid API key, the session is created and nothing on the wire or stderr names a server whose command does not exist.

Antigravity arms main's unasked-call tripwire (AcpClient._tripwire_pi_gate) at spawn, as MiMo Code does: a call that reports completed with no session/request_permission for its toolCallId stops the harness. It acts after completion and cannot undo that call. Its HARNESS_UNASKED_TOOL_IDS row lists only the read-only built-ins (fs_read, grep and glob rows of the tool table), because 1.4.4 runs file reads unasked; harness_tool_name reads the Running <name> title such a read carries, so the recorded name is view_file. The tripwire is fail-closed: any other unasked completion, a call whose name was never recorded included, ends the session, and a call whose frame named an MCP server never passes as a listed read. Replaying the three signed-in captures trips nothing. Main's tripwire state is not reset when the process respawns; that is shared with pi, DeepSeek and MiMo Code and is left as it is.

Path-like argument names feed the sensitive-path check, including nested values and decoded file URIs. Unfamiliar command-like names detect disagreement but never supply a trusted command. Other arbitrary unknown keys are not blanket-refused.

Antigravity's run_workflow runs the Python in its Script argument, which no command or path rule reads, and invoke_subagent starts sub-agents whose own tool calls no capture has shown asking. Such a call, named by its tool_call frame or carrying an argument whose name has the whole word script or subagent (so Description, the edit tool's own argument, does not count), is judged as a shell call with no verifiable command, which the unverifiable-shell tier refuses on every approval policy. Its other arguments are dropped first, so none can stand in for the call as the command.

ACP_BACKENDS_CONFIG_MODEL_SELECT_IS_CANONICAL makes the shared picker use configOptions model ids, because the Antigravity models list combines model and effort (for example gemini-3.8-flash-high). Antigravity stays outside the four launch/reload capability sets: model and effort use session options, MCP uses session/new, and relaunch/reload behavior is unverified. The mirror rules the new toolsSettings concern no-channel, as goose and strands do: the session/new element set has no field for it, and a new chat session whose spec sets it gets one notice naming what goes unapplied. Antigravity is in ACP_BACKENDS_CREW_FIRES_SPEC_OBSERVER_HOOKS beside claude and codex, not in ACP_BACKENDS_CREW_FIRES_SPEC_HOOKS: its file reads raise no permission request, so a spec's preToolUse hooks do not run, a new session whose spec has one is told so, and postToolUse, userPromptSubmit, agentSpawn and stop still fire, and an unasked read is named for a postToolUse matcher by its tool_call event. Session Activity counts Antigravity sessions from its usage rows (_SHARD_COUNTED_HARNESSES), since it writes no kiro-cli or KAS transcript.

The credential declaration and packaging fence cover both token files; on a host that adopted aws-control, the fence entries arrive through KiroCrewCaptain/kirocrew-app-aws-control#10. Its own tokens remain readable for sign-in; a maintainer accepted, in review, the host_auth.py change that allows this, and its comment now says so. The unused is_antigravity_backend property is removed. The one-line MCP seam stays: the Strands mirror test fails on a bare _session_mcp_servers() splice in the session arrays, so Antigravity uses a named seam like the other harnesses. Provider labels move unchanged into acp/provider_labels.py and are re-exported from acp/types.py to meet the line cap. Process recognition tests use launcher-shaped bytes; a live 1.4.4 process on Linux, launched with the argv Crew's spawn plan uses (absolute path only), reads back the same /proc/<pid>/cmdline shape, and the production check recognizes it. The turn corpus is one signed-in 1.4.4 session in three files (turn-live-tools, turn-live-refused, turn-live-mcp). The optional batch write guard does not cover every Antigravity native tool name; the permission floor and unasked-call tripwire do not depend on that guard. A separate documentation correction aligns the DeepSeek mirror table with its existing runtime projection and adds a regression test; it is not an Antigravity rebase-conflict fix.

Area Change
src/kiro_crew/acp, src/kiro_crew/acp_backends.py Launch, permission frames, version refusal, tripwire and provider labels
src/kiro_crew/agent_sdk Harness registries, capability choices, install hint, auth and tool gate
src/kiro_crew/providers, src/kiro_crew/session_pid_harness.py Provider mirror and process recognition
src/kiro_crew/apps, scripts Packaging credential fence and parity/deny checks
src/kiro_crew/docs, docs Configuration, host and permission contracts
website/src/api, website/src/i18n, test Dashboard name/catalogs and behavior tests

Backwards compatibility

Compatible: this adds a backend id without changing the default backend or existing launch paths. The credential floor gains Antigravity token leaves, which enforced sibling backends also mask unless their own declarations exclude them; settings files and their parent directory are not fenced as token files.

Antigravity starts with a credential mask that hides other credential stores, including .aws and .ssh. Its own two token files remain readable to the child for sign-in. Git over SSH works there only through the ssh-agent forward the operator consents to: the private keys stay hidden, and ~/.ssh/known_hosts comes back read-only, as for every enforced harness since #34328. A maintainer accepted the own-token exclusion in review.

Reader: src/kiro_crew/security/paths.py:sensitive_path_refusal -- chat -- the shared file gate also refuses Antigravity's two token leaves for existing backends; the settings file and parent directory remain readable.
Reader: src/kiro_crew/agent_sdk/tool_gate.py:adapter_hidden_credential_dirs -- chat -- builds the mask; for antigravity it leaves out the adapter's own acp_token.json and acp_business_token.json and keeps .aws and .ssh in it, and enforced siblings gain the token leaves unless their own declarations exclude them; test_antigravity_credential_mask_spares_only_its_own_token_files
Reader: src/kiro_crew/acp/client_start.py:_spawn -- chat -- takes the mask from the launcher's spawn plan; only the antigravity plan is new; test_antigravity_spawn_sets_helper_and_credential_mask
Reader: src/kiro_crew/acp/client_lifecycle.py:_sandbox_init_failure -- chat -- hands the mask the spawn recorded to the sandbox init probe; the Antigravity mask is one more value of the same type
Reader: src/kiro_crew/acp/transport_errors.py:sandbox_init_failure_for_runtime -- chat -- hands the runtime's recorded mask to the same probe; same type, no new code path

Reader: src/kiro_crew/security/paths_lists.py:_SENSITIVE_HOME_DIRS -- chat -- the shared read and write gates fence the two token leaves for every backend, not the settings file or the whole parent directory.
Reader: src/kiro_crew/security/paths_targets.py:_home_dir_targets_uncached -- subagent -- GEMINI_HOME re-anchors the two exact token leaves; ordinary Gemini settings remain outside this credential addition.
Reader: src/kiro_crew/dashboard/file_api/grep.py:_grep_sensitive_globs -- crew page -- search excludes the credential targets before reading them; per-hit sensitive-path checks remain the authority.
Reader: src/kiro_crew/dashboard/side_readonly_spec.py:credential_denied_paths -- crew page -- Side Chat derives its denied paths from the shared credential targets and therefore also refuses the new token leaves.
Reader: src/kiro_crew/acp/client_claude_session.py:_claude_side_read_deny_rules -- chat -- Claude Side Chat adds denies for the same targets while keeping ordinary project settings available.
Reader: src/kiro_crew/mcp_cron_shell_scan.py:_CRON_CRED_PATH_RE -- cron -- scheduled shell commands also reject literal references to the new token leaves; ordinary scripts retain the same scanner.
Reader: src/kiro_crew/skills_script_validator.py:_SENSITIVE_PATH_RE -- app -- generated skill scripts pick up the two token leaves from the shared floor; no new script syntax is required.
Reader: src/kiro_crew/apps/builtins/aws_control/crew/packaging/pipeline/sensitive.py:_SENSITIVE_RELATIVE_DIRS -- release -- the standalone packaging fence excludes the same two token files without excluding their settings siblings.
Reader: src/kiro_crew/apps/builtins/papyrus/backend/latex.py:_sensitive_hidden_dirs -- app -- the compiler child inherits the home-relative credential floor and hides the new token leaves.
Reader: src/kiro_crew/dashboard/handlers/office_slides.py:_sensitive_hidden_dirs -- crew page -- the document converter derives its home-relative hidden paths from the same floor.
Reader: src/kiro_crew/security_posture_controls.py:_sensitive_path_items -- crew page -- the security posture view derives its credential items from the shared floor rather than a separate list.

Tests

The PR head is 2eeb923, one commit on main 969de99: 4e0d321 rebased. Two files conflicted. In acp/client_tool_events.py main moved the DeepSeek session code off AcpClient (#37641), so the ACP_BACKEND_DEEPSEEK import has no use left, and this PR's start_event_tool_id replaces the last use of ACP_BACKEND_CODEX, so that import goes too (flake8 F401 found it). In acp/types.py main's steering_wire re-export gained two goose readers (#37567), so the resolution keeps main's four names and this PR's provider_labels re-export. No other file's patch changed. On 2eeb923, 62 related test files ran: 5331 passed, 38 skipped and 4 failed, and the same 4 fail on clean main 969de99 on the same machine (two test_pid_lifecycle.py reclaim tests, test_worker_agent_spawn_gate.py::test_the_kas_projection_reads_nothing_after_the_gate and test_acp_client_stale_turn.py::TestStaleTurnProbeAccounting::test_priming_verdict_defers_then_real_evidence_reaps). 4e0d321 was one commit on main 25860b2: bcb339e rebased. #37602 moved the opencode session code off AcpClient into OpencodeAcpClient, so the client_class.py conflict keeps main's move and this PR's _is_antigravity, and two agent-host-contract.md tables keep main's rows plus the Antigravity cell. aws-control is now pinned for adoption (#37411), so its wheel-copy fence entries are ported in KiroCrewCaptain/kirocrew-app-aws-control#10, named under Related Issues. That PR lands after this one: its fence test compares the list with the shared validator of its pinned Kiro Crew, which has no Antigravity entries until this PR merges. Dropping the two wheel-copy lines here instead fails two of this repository's tests (test_every_harness_owned_credential_leaf_is_fenced_standalone and the wheel copy's test_the_local_subset_fences_every_foreign_credential_store), so they stay. bcb339e was one commit on main faac7ea: be1babb rebased, plus one fix. #37521 added ClaudeHarness to the exports in acp/harness/__init__.py, beside this PR's AntigravityLaunch, so the resolution keeps both names; no other file conflicted, and #37472's per-harness runtime-sharing switches list KAS and Codex only. The fix: this PR's own test_slash_command_no_channel.py expected /help on Antigravity to get the dashboard's notice, but Antigravity had no ACP_BACKEND_PROMPT_HANDLER_COMMANDS row, so /help went to the client, which refused it before the wire, and the / menu listed every kiro-cli command. Antigravity now has a row of /usage, the one Crew harness command in its prompt allowlist, as Copilot has one. The / menu lists that and the dashboard's own commands, any other kiro-cli command gets the notice, and /plan (not a Crew harness command) stays prompt text that the allowlist admits. The slash-commands cell in agent-host-contract.md changes from gap to partial, and harness-parity.md lists the row. On bcb339e, 45 related test files ran: 4451 passed, 19 skipped and 3 failed, and the same 3 fail on main on the same machine (two test_pid_lifecycle.py reclaim tests and test_worker_agent_spawn_gate.py::test_the_kas_projection_reads_nothing_after_the_gate). be1babb was one commit on main fd4c0d3: f59827b rebased. #37535 moved the goose session code off AcpClient into GooseAcpClient, so the client_class.py and test_acp_mcp_startable.py conflicts keep main's move and re-add only this PR's Antigravity lines (the version check binding, and Antigravity's own MCP array in the startable test's per-backend map); one table in agent-host-contract.md keeps main's rows plus the Antigravity cell. The patch's added and removed line counts are unchanged. f59827b was one commit on main 2e56b3b: 4dd8455 rebased, plus the two GPT 6.1 findings on it and one CI fix. #37484 moved the claude-agent-acp code off AcpClient into ClaudeAcpClient, so the client_class.py, acp_launch_capture.py and test_client_bindings.py conflicts keep main's move and re-add only this PR's Antigravity lines; the two tables in agent-host-contract.md keep main's rows plus the Antigravity cell. A tool_call event now names an Antigravity call for its PostToolUse spec hooks (start_event_tool_id, shared with codex), because 1.4.4's reads send no permission request: the recorded Running view_file is antigravity#view_file, which a fs_read matcher meets. An MCP call is named from its trusted _meta.mcp pair on both events (@probe/echo_probe, not the fused title probe_echo_probe). The usage facade names ACP_BACKEND_ANTIGRAVITY for type checkers. 4dd8455 was one commit on main 3397952: cda20d5 rebased, where the only conflict was test/test_client_bindings.py, whose frozen member inventory gained main's _note_claude_adapter_version beside this PR's _refuse_unverified_antigravity_version; both entries are kept, and the patch's added and removed line counts are unchanged. cda20d5 was one commit on main 3887f88. On 3fdb7d4 the review lanes found that the PR's own unasked-call watch (client_unasked_watch.py, ACP_BACKEND_MUST_ASK_TOOL_KINDS) duplicated main's tripwire, so it is deleted and Antigravity uses main's; that a substring match marked Description as a script argument, so every edit was refused; and that Antigravity promised preToolUse on every call. CI found the _dispatch facade missing five forwarded names and Session Activity not placing Antigravity. All are fixed in cda20d5. Before that, 3fdb7d4 was 616cb8b, the signed-in capture (4f045ce, 0179f0c, 9c2f239) and the run_workflow and invoke_subagent refusals (9b58adc, b00a8ee5), squashed with the same tree as 1398028, then rebased onto 3887f88 with the two CI fixes below. 616cb8b is the commit reviewed as 94a0a9e, rebased onto main fe6b063, e08544c, 5e8741f, 4b820ae and bf9b611 (as 12f22c1, then amended with the watch field bound above as f065d86), then onto a104344 (as e89e2d3) and 1e990f8 (as 2d323b3), then amended to move the unasked-call tripwire into its own module (as a95e158), amended to bound Antigravity's MCP identity (below), rebased onto bab2387 (as 3041371), onto 3535135 (as 63a5da4), onto df553b0 (as cc5739e), onto a398062 (as 6c46cdb), onto b358303 (as 153e78a), onto 3c7981d (as 16d11ce), onto b4b8a01 (as 9386ca2), onto a496a7d (as 75347f2), onto baf55ed (as f2cd8c4), onto b7b7b6c (as b4e6fb9, then amended for CI as 650551b) onto 35dc109 (as 4ffb5fa) onto 45227df (as 64b35ca) onto 478d13a (as 2adf1d2), onto 06d61fe (as a7eceac), onto 1bccaa1 (as d9367bf), onto 048700a (as 460ab00) then onto fbf08dd and de6219e (as 7810a68), onto b3f274d (as b4674eb), onto 9e9de7b (as 62e9656), onto 2cd247a (as 019eaf7), onto 0842788 and 5dc219c (as b7be5a8) and onto 2a8daa3 (as 77f10af7, then amended to complete a test table, below, as 36d3aa60), onto c000f31 (as 56c81aa), onto 911bde8 (as a6bbabe), onto f9ebf3a (as d17eec5) and onto d9840be (as 616cb8b). Each rebase except the one onto c000f31 hit conflicts. #30131 changed one row of the section 5 MCP table and #30237 rewrote the spec-hooks paragraph, both in docs/system-specs/modules/agent-host-contract.md. #30108 added a toolsSettings shell and write allowlist row to the backend table there, and touched client_lifecycle.py (docstring), client_permissions.py (imports), test_slash_command_no_channel.py (imports) and test_tools_settings_notice.py (backend lists). #30086 added a Claude tool table and import to acp/harness_tool_names.py, next to this PR's Antigravity tables, and changed the CC cells of two rows of the harness tool-names table. Each of those resolutions keeps upstream's text and re-applies only this PR's change: the Antigravity table cells (on #30108's new row the cell is "not applied", because Crew enforces those allowlists only for Claude), antigravity in the two spec-hooks backend lists, and the union of both sides in the code and test files. On bab2387, #31812 changed the kiro-cli sign_in_remedy cell of the host identity table in agent-host-contract.md (sign in, then send the message again), and #31571 rewrote the place_single_binary_array docstring in providers/mirrors/opencode.py to cover every single-binary harness. Both resolutions keep main's text: the table keeps this PR's Antigravity column, and the PR no longer changes opencode.py. On 3535135, #32666 changed the Codex cells of the section 6 "Quota API" and "Consumer surface" rows (Codex's 5-hour and weekly windows, read from its own rollout logs), and #32600 changed the section 9 row for MCP servers the spec declares (the Kerberos and ssh-agent pointers relayed to servers). Those resolutions keep main's text in the three rows and re-apply only this PR's existing change to them: the Antigravity cells, and in "Quota API" the shorter kiro-cli cell this PR already had. No other file conflicted, and every other file's patch is unchanged from 3041371. On df553b0, #30937 changed fs_write to write (kiro-cli's own name for it) in the spec-hooks paragraph, the same line where this PR adds antigravity to two backend lists. The resolution keeps both changes, and no other file's patch changed. On a398062, #29509 added three crewmate_harness_* keys to agentBackendTab in en.json and the pseudo-locale, beside this PR's antigravity key in en-XA.json. The resolution keeps both, in the order the pseudo-locale generator writes them (en.json keys, then en.manual.json keys), and every file's added and removed lines are unchanged. On b358303, three files conflicted, each only beside this PR's lines: in acp-client.md the claude launch bullet (KIROCREW_NODE_BIN_DIR) next to the antigravity bullet this PR adds; in agent-host-contract.md the Readiness probe row of the entitlement table, where #32970 renamed the enterprise build to the LTS build; and in en-XA.json #32849's new agentBackendTab card summary keys next to antigravity. Each resolution keeps main's text and re-adds only this PR's lines (the antigravity bullet, the Antigravity cells, the antigravity key in generator order). The Readiness probe row is the only added or removed line that changed, picking up main's LTS wording. On 3c7981d, #33189 added cron-turn sentences to the spec-hooks paragraph, the same line where this PR adds antigravity to two backend lists; the resolution keeps both, and no added or removed line of the patch changed except that paragraph, which now carries main's new sentences. On b4b8a01, #34472 raised the Strands CLI pin from 0.2.3 to 0.2.4, which changed the Strands cell of the "Failure mode when missing" row of the section 8 table in agent-host-contract.md; the resolution keeps main's row and re-adds only this PR's Antigravity cell, and that row is the only added or removed line of the patch that changed. Antigravity stays out of main's new ACP_BACKENDS_CREW_OAUTH_BRIDGE: each member there cites a live capture, and this PR has none for the bridge. On a496a7d, #33510 added a sentence on KAS ask grants (kas_hook_coverage.hook_held_grants) to the spec-hooks paragraph, the same line where this PR adds antigravity to two backend lists; the resolution keeps both, and that paragraph is the only added or removed line of the patch that changed. #33510 also passes governance_app to every harness's session_extras; the Antigravity harness inherits the base method, which takes it. On baf55ed, #34328 gave ~/.ssh/known_hosts back read-only to every enforced harness through tool_gate.adapter_expose_files, changed the "Extra hidden paths" row of the sandbox table in agent-host-contract.md, and added a test that every enforced harness's launch module calls that function. Antigravity is enforced, so its launch now passes extra_expose_files=acp_tool_gate.adapter_expose_files(self.backend, hidden), the same line #34328 added for goose, and its "Extra hidden paths" cell says known_hosts comes back read-only. The table conflict was resolved row by row: main's "Extra hidden paths" row plus the Antigravity cell, and this PR's rows where main changed nothing. That row and the two antigravity.py lines (the import and the argument) are the only added or removed lines of the patch that changed. On b7b7b6c, #34333 made every harness that signs in with its own credential file say that its sign-in runs in a terminal on the gateway host, followed by host_auth.HOST_TERMINAL_NOTE (where the dashboard opens such a terminal), and added a test over every such declaration. Antigravity's declaration is one, so its sign_in_remedy and signed_out_message now say the same, and its sign_in_remedy cell in agent-host-contract.md starts with "in a terminal on the gateway host". The table conflict was again resolved row by row: main's sign_in_remedy row plus the Antigravity cell. That row and the two Antigravity sign-in texts in host_auth.py are the only added or removed lines of the patch that changed. CI on b4e6fb9 then failed test_spec_hook_kiro_cli_payload.py::test_every_native_tool_of_every_harness_gets_its_kiro_cli_name (added by #30937): this PR listed Antigravity's find_file and search_directory under both fs_read and glob, and the first row decides the built-in a spec hook's stdin names, so the glob rows came out as kiro-cli's read. They are now under glob only, kiro-cli's own file finder (payload tags read and @builtin), as no other harness table lists a tool under two rows; test_acp_antigravity_backend asserts the new first name, and that is the amend to 650551b. The other CI failure on b4e6fb9, test_session_markdown.py::test_a_condition_one_opener_is_closed_as_the_parser_reads_it[long-s], is #34654 (markdown-it-py 4.3.0 no longer reads <ſcript> as a script block), not this PR. The rebase onto 35dc109 conflicted only with #33215: it added test/test_kas_subagent_targets.py to the first cell of the section 5 web_search and use_subagent row and rewrote the KAS cell of the section 1 "Selection mechanism" row in agent-host-contract.md, and it added a KAS sub-agent paragraph to agent-spec-fields.md just before the mirrored-harnesses bullet this PR rewrites. The resolution keeps upstream's text and re-adds only the Antigravity cells and that bullet; those two rows are the only added or removed lines of the patch that changed, and the file set is the same. #33215 otherwise changes KAS and the shared runtime (kiro-cli, KAS and Codex), which Antigravity does not run on, #33595 changes MCP gateway, doctor and dashboard-notice code, and no test main added since b7b7b6c iterates over every harness. The rebase onto 45227df conflicted only with #34367, which added a sentence about kirocrew chat asking the spec's own preToolUse hooks to the agent-host-contract.md paragraph where this PR adds Antigravity to the ACP_BACKENDS_CREW_FIRES_SPEC_HOOKS list; the resolution keeps that sentence and this PR's list, and that paragraph is the only added or removed line of the patch that changed. #34367 changes kirocrew chat and its channel gate, not harness code, and its new tests are KAS-only. The rebase onto 478d13a conflicted only with #34668, which renamed the KAS Settings path in the section 3 sign_in_remedy row of agent-host-contract.md ("Settings → Agent Harness → Kiro Agent → Kiro sign-in"), the row where this PR adds the Antigravity cell; the resolution keeps upstream's row and re-adds only that cell, and that row is the only added or removed line of the patch that changed. #34668 also changes _KAS_CARD in host_auth.py and adds a KAS-only kas_in_use flag to acpBackend.ts, both of which merge with this PR's edits to those files without conflict; its test changes are KAS-only. The locale files main changed alongside this PR still parse, with no duplicate key. On 64b35ca and on 2adf1d2, the first Fork PR Description and PR Readiness runs found the description complete but could not post their result, because the repository bot's GitHub API rate limit was spent ("API rate limit exceeded for installation"); both re-run when the description is edited. On a7eceac and on 460ab00, Fork PR Description passed, and the first two PR Readiness runs stopped on the same spent limit before they could post a verdict. The rebase onto 06d61fe conflicted only in two rows of agent-host-contract.md, and each resolution keeps upstream's row and re-adds only the Antigravity cell; those two rows are the only added or removed lines of the patch that changed. #34555 added to the CC cell of the section 10 slash-commands row (/context, /help, /agent ...) that the / menu also lists what the chat's session advertised; the menu keeps that list only for ACP_BACKENDS_HARNESS_OWNED_SLASH_COMMANDS (Claude alone), so Antigravity's cell (only /plan and /usage reach the harness) stands, and its three new AcpClient members, which test_client_bindings.py lists beside this PR's five, share no name with them. #34475 added to the CC cell of the hook and deny-rule tool-name row that a crewmate worker's native-tool excludes ride an allow-list; its new ACP_BACKENDS_MATE_WORKER_NATIVE_TOOL_LIMITS holds KAS and Claude only, Antigravity is in no mate-worker set (a worker is refused there before spawn, ACP_BACKENDS_MATE_WORKER_KIT), and the Antigravity mirror's SessionProjection leaves the new builtin_tools field at its default of None. Of the harness code, main changed only acp/harness/kas.py since 478d13a, its only new backend set is #34475's, and the new real-adapter MCP probe (test_real_adapter_harness_probe.py) runs only the backends its expectations.json measured, which does not include Antigravity. The rebase onto 1bccaa1, 241 commits later, conflicted only in agent-host-contract.md. #35278 (the KAS cell of the section 1 selection-mechanism row), #29221 (the Codex cell of the section 8 additional-runtime row) and #34696 (the Codex cell of the section 10 slash-commands row) changed rows to which this PR only appends the Antigravity cell, so each resolution keeps upstream's row and re-adds that cell. #35268 added a section 10 row on a shipped agent's job tools staying out of tool-search deferral; its Antigravity cell reads unknown, as for every harness but KAS and kiro-cli, because the Antigravity code has no tool-search path and no signed-in capture. #34736 rewrote the cron sentence of the KAS spec-hooks paragraph, and a word-level merge keeps it beside this PR's two antigravity list entries, which share no word with it. Those rows and that paragraph are the only added or removed lines of the patch that changed; every other file main changed beside this PR took the PR's lines unchanged. Main's new ACP_BACKENDS_CHAT_RUNTIME_SHARING (kiro-cli and KAS) and ACP_BACKEND_PROMPT_HANDLER_COMMANDS (Codex) do not include Antigravity, #35278's new harness seam start_mode_ran_servers has a base default, the two launch helpers the Antigravity harness calls (_run_preflight_bounded, _sandbox_preflight) are unchanged, and every kiro_crew name the PR's Python files import resolves on the new main. The rebase onto 048700a conflicted only in the section 8 failure-mode row, whose Codex cell #35485 extended with the kirocrew doctor row; the resolution keeps upstream's row and re-adds the Antigravity cell. #35637's new ACP_BACKENDS_CREW_LOADS_SPEC_STEERING (Claude and Codex) and ACP_BACKENDS_NATIVE_PROJECT_AGENTS_MD (Codex) leave Antigravity out, which matches its mirror's withheld ruling on resources. The rebase onto fbf08dd conflicted in six Python files and agent-host-contract.md, all from #31950, which fires a spec's non-gating hooks on Codex sessions. #31950 passes the backend to the session-update parser as harness_backend, adds a Codex tool table and reads a frame's tool name through stated_harness_tool_name. This PR had passed the same value as backend in session_update_parser.py and session_handle_dispatch.py; the resolution keeps main's code, so neither file is in the patch any more (92 files), and stated_harness_tool_name hands the backend on to harness_tool_name, so Antigravity's Run <name>? title is still unwrapped on Antigravity only. harness_tool_names.py keeps the Codex table and the Antigravity table side by side, both in HARNESS_TOOL_TABLES, and the two Antigravity test calls to the parser now use harness_backend=. #31950 also filled the Codex cell of the section 10 row on hook matchers written in kiro-cli tool names and changed the first and Codex cells of the row on a spec's own hooks, rows to which this PR only appends the Antigravity cell, so each resolution keeps upstream's row and re-adds that cell; its new Codex sentences in the spec-hooks paragraph share no word with this PR's two antigravity entries. Antigravity was already in ACP_BACKENDS_CREW_FIRES_SPEC_OBSERVER_HOOKS through ACP_BACKENDS_CREW_FIRES_SPEC_HOOKS, #31950's new tool-id reader in the parser runs on Codex only, and every kiro_crew name the PR's Python files import resolves on de6219e. The rebase onto de6219e conflicted only in client_permissions.py, where #29127 now reads the permission value through permission_value_for, which prefers a value the adapter advertises; its one preferred row is Codex's external-sandbox, so Antigravity writes the same value as before, and the resolution keeps the Antigravity version check ahead of both calls. The rebase onto b3f274d conflicted only in the section 2 session-persistence table, where #35260 rewrote the Codex cells of the replay-store, store-root and directory-naming rows; the resolution keeps upstream's rows and re-adds the Antigravity cell. #35260's new harness seam delete_disposable_record defaults to None, and #34763's new ACP_BACKENDS_TOOL_SILENCE_TIMER holds kiro-cli and KAS only, so neither reaches Antigravity. The rebase onto 9e9de7b conflicted in agent-host-contract.md and acp/client_lifecycle.py. #33578, which runs a trusted folder agent's own hooks on KAS, changed the KAS cell of the section 5 delivery-channel row, to which this PR only appends the Antigravity cell, and the KAS sentence of the spec-hooks paragraph, which shares no word with this PR's two antigravity entries; each resolution keeps upstream's text. In _reset_state, #35698 now forgets the pictures a session already sent and this PR clears the unasked-call watch, so the resolution keeps both lines. #35698's new ACP_BACKENDS_SENT_IMAGE_RECORD holds kiro-cli, KAS, Claude and Codex only, so Antigravity inlines pictures as before. The rebase onto 2cd247a conflicted only in the section 10 row on a spec with "tools": [], where #35508, which refuses Codex permission requests on such a spec, changed the first and Codex cells; the resolution keeps upstream's row and re-adds the Antigravity cell. #35508 changes only Codex's runtime session handle and leaves ACP_BACKENDS_HONOR_ZERO_TOOL_BAN as it was, so the Antigravity cell still holds. The rebase onto 0842788 conflicted only in the re-export lists of acp/types.py and acp_backends.py, where #35756 adds ACP_BACKENDS_COMPACTION_REPORTED beside this PR's ACP_BACKENDS_CONFIG_MODEL_SELECT_IS_CANONICAL; the resolution keeps both names in order. That set is ACP_BACKENDS_SENT_IMAGE_RECORD (kiro-cli, KAS, Claude and Codex), so Antigravity, outside it, is sent the reply-format rules on every message, as every other harness outside it is. The rebase onto 5dc219c met #24732, which onboards MiMo Code as an enforced backend and adds it at most of the places this PR adds Antigravity: 11 Python files and four docs conflicted, each where both sides add lines at the same spot. Each resolution keeps main's MiMo Code lines and this PR's Antigravity lines side by side (backend ids and sets, routing, models, host-auth declarations, the tool tables in harness_tool_names.py, the client_class.py bindings, test_agent_sdk_capabilities.py and _MCP_SEAM_HOOKS in test_harness_parity.py). In the agent-host-contract.md tables main's MiMo Code column stays after OpenCode and the Antigravity column stays last, and the opening counts now include both (eleven backends, nine foreign hosts). #24732 also gave MiMo Code a provider label in acp/types.py, whose label block this PR moves to acp/provider_labels.py; PROVIDER_LABEL_MIMOCODE and its PROVIDER_LABEL_BY_BACKEND row now sit there with the others, and acp/types.py still re-exports it. Seven section 10 rows in which main's MiMo Code cell and this PR's Antigravity cell hold the same text (unknown or a gap) had merged without a conflict but one cell short; they now carry both cells. MiMo Code's new ACP_BACKENDS_UNASKED_NATIVE_TOOLS, HARNESS_UNASKED_TOOL_IDS and HARNESS_PARAM_ALIASES hold MiMo Code only, and its unasked-call tripwire in client_permissions.py and this PR's unasked-call watch in client_unasked_watch.py (ACP_BACKEND_MUST_ASK_TOOL_KINDS, Antigravity only) each run on their own backend. Every capability set and backend-keyed table holds the same backends other than Antigravity as on main, and the only added or removed Python lines of the patch that changed are the MiMo Code provider-label lines (moved with the block) and "mimocode" in _MCP_SEAM_HOOKS. The rebase onto 2a8daa3 met #36462: the client now keeps the process adapter for the life of the process and asks it for the host's own session/new / session/load entries (ProcessAdapter.session_mcp_servers, read through one _adapter_session_mcp_servers splice) instead of one gated splice per harness. This PR's three Antigravity splices in client_start.py are therefore gone: AntigravityLaunch overrides session_mcp_servers to return session._antigravity_session_mcp_servers() and calls super().__init__() for the adapter's new read_back, the ProcessSession protocol in acp/harness/base.py names the Antigravity hook beside the others, and _ADAPTER_MCP_SEAMS in test_harness_parity.py declares AntigravityLaunch. The acp-client.md adapter row keeps main's new text with this PR's file list (Antigravity and Copilot added). A check after that rebase found that COMPACTION_SEEN in test_reply_format_once.py (from #35756), which must name every selectable harness, did not list Antigravity, so test_the_table_names_every_selectable_harness could not pass on b7be5a8; it now lists Antigravity as False, since Antigravity is outside ACP_BACKENDS_COMPACTION_REPORTED. With acp/harness/base.py and that test, the patch touches 94 files. The rebase onto c000f31 had no conflict and leaves the patch unchanged (same git patch-id). #36579 there adds test_every_value_crew_writes_reads_back_as_no_drift, which reads every ACP_BACKEND_FORBIDDEN_MODES row, Antigravity's included: Crew writes mode=default with or without the option advertised, and session_config_drift_issue reads that back as no drift. The rebase onto 911bde8 conflicted in one agent-host-contract.md paragraph, the one on the spec's hooks field, where #36829 rewrote the sentence on a checkout's own hook files (the chat now shows that notice once); the resolution keeps main's sentence and this PR's two Antigravity mentions in that paragraph (the ACP_BACKENDS_CREW_FIRES_SPEC_HOOKS members and the spec-key notice). #36550's new ACP_BACKENDS_CWD_BOUND_SESSIONS holds opencode only, and Antigravity is not added to it: nothing measured shows that its session/load ties a session to the folder it was created in. Apart from that paragraph, the patch's added and removed lines are the same as 56c81aa's. The rebase onto f9ebf3a conflicted in the same paragraph, where #36955 adds a sentence (the other session-start spec notices now show once per chat too); it was resolved the same way, main's text with this PR's two Antigravity mentions, and the patch's other lines are unchanged. #36983 now rules autoApprove withheld on a broker-only harness (pi), as every mirror already does; Antigravity is a mirror, and its card already lists autoApprove as not taking effect. The rebase onto d9840be conflicted in two places, both from #37105: it updated the Pi cell of the section 8 "Failure mode when missing" row (PI_MIN_VERSION is now 0.99.2), and it dropped the harness-parity.md intro sentence about session_pid_orphan_scan.py::_BROWSER_PLAUSIBLE_OWNER_NAMES, a set main no longer has. Each was resolved to main's text with the Antigravity column or name kept. #36993 now names a dropped remote MCP oauth block at session start on every harness outside ACP_BACKENDS_REMOTE_MCP_OAUTH, which includes Antigravity; section 10 already lists that sign-in as a gap for Antigravity. The rebase onto 3887f88 conflicted in three files. #36118 added Copilot to ACP_BACKENDS_MODEL_VIA_CONFIG_OPTION beside this PR's Antigravity line, and #36083 rewrote the kiro-cli and KAS rows of NATIVE_ALLOW_STORES_NONE beside this PR's Antigravity row; both resolutions keep main's lines and this PR's. In agent-host-contract.md #36118 changed the Copilot cell of two rows, "Entitlement discovery, model side" and "The model picker lists the harness's models"; both tables conflicted whole because this PR appends the Antigravity column to every row. Every row now holds main's text plus this PR's Antigravity cell, and every data row of the PR matched its base row except for that cell. The patch's added and removed line counts are unchanged. Two CI findings on 1398028 are fixed in the same push: the portability gate read the new is_shell = True as a subprocess shell=True, so the branch now assigns the computed flag, and a test query spelled the product name in lower case.

#30616 then made Crew expand ${VAR} and ${env:VAR} in the shared session projection, which the Antigravity mirror uses. The Antigravity cell of the section 5 "Env expansion" row therefore changes from "literal pairs, not measured" to: Crew expands both forms before the element is built; whether Antigravity writes the array anywhere itself is not measured. test_session_mcp_env_refs.py now lists Antigravity among the expanding backends (4 new cases: the three value spellings resolve, and a credential-named variable is never dereferenced). git range-diff shows no other change.

#30874 then added a sentence to the KAS cell of the section 5 "Per-tool MCP deny" row in agent-host-contract.md (the switched-off tools the dashboard writes to the global mcp.json), the row where this PR adds the Antigravity cell. The resolution keeps upstream's row and re-adds only the Antigravity cell; git range-diff f065d86e e89e2d36 shows no other change. #30874 changes only KAS code, so Antigravity's behaviour is unchanged.

#30308 then reworded the CC cells of the section 7 "Auto mode" and "Inherited-config hazard" rows in agent-host-contract.md, and the claude bullet in acp-client.md that the Antigravity launch bullet follows. The resolution keeps upstream's text and re-adds only the Antigravity cells and bullet; git range-diff e89e2d36 2d323b3f shows no other change, and the PR's other 90 file patches are identical. #31016 gives kirocrew-core an agent-socket pickup under an enforced harness. Antigravity's routing (SESSION_CONFIG) is enforced, so its session array gets the pickup with no change here; checked locally, its credential mask leaves the agent socket reachable and keeps run/ and .local_secret hidden, and only kirocrew-core receives a pickup. Whether the Antigravity runtime hands that element's env to the stdio server unchanged has not been measured (the signed-in capture did not test it).

f065d86 differs from 12f22c1 in two things. The watch bound answers the fork AI review's blocking finding on 12f22c1 (a-bound-bounds-every-field-it-retains): it changes client_permissions.py and the Antigravity cell of the host contract, and adds 11 cases to test_acp_antigravity_backend.py. The host_auth.py comment now records the maintainer's acceptance of the own-token exclusion. The new cases cover a must-ask call with an over-long id (stated, asked first, MCP-marked, relabelled from read), an id of exactly 512 characters, an over-long id that can never trip, a 1 MiB kind, and the overflow row beside an over-long id. Six of the 11 fail on 12f22c1's code; the other five pin behaviour the bound must not change.

a95e158 differs from 2d323b3 only in where the unasked-call tripwire lives. UNASKED_WATCH_LIMIT, _watched_kind, _watch_tool_call and _tripwire_unasked_call move unchanged from acp/client_permissions.py into a new owner module, acp/client_unasked_watch.py, which AcpClient binds like its other owners and _CLIENT_OWNERS lists. client_permissions.py is now 1,469 lines (1,456 at base), under the 1,600-line working target that check_file_line_cap.py reports (#30770). The binding inventory in test_client_bindings.py records the new owner of the two methods, the acp-client.md owner table and the Antigravity cell of the host contract name the new file, the watch-limit tests import the limit from it, and the id-log scan in test_deny_bounded_write.py and the redact-before-bound scan, which covered this code in client_permissions.py, now list it too.

Since a95e158, the PR also bounds Antigravity's MCP identity. This answers the fork AI review's blocking finding on a95e158 (a-bound-bounds-every-field-it-retains, at the _meta.mcp row in acp/_dispatch_parts/tool_call_identity.py). AcpClient cached a tool_call frame's _meta.mcp.server and _meta.mcp.tool per open call as sent, so 256 open calls naming a 1 MiB tool held 256 MiB. That row of _MCP_IDENTITY_META_CHANNELS is now bounded like the KAS row, by the 512-character bound of Crew's own MCP tool surface (mcp_identity_name). A half over the bound is never cut: it is retained as empty and reported unreadable. For Antigravity, AcpClient records that mark per call (a later readable tool_call under the same id clears it, as it rewrites the identity caches), and _build_permission_event sets mcp_identity_unreadable on that call's permission event, which the gate already refuses on every path (gate_tiers, _resolve_permission and the auto-approve floor). The client sets the flag itself rather than handing the mark to the shared parser as its unreadable store, because that store also turns on the KAS built-in id fold. A new requires_server field keeps the KAS row's server-present rule on that row alone, so a frame naming only an oversized tool is bounded too, and classify_tool_call no longer reads the codex rawInput pair for a frame a bounded channel marked unreadable: Antigravity stamps is_mcp_tool_call beside _meta.mcp, and its rawInput holds the tool's own arguments. No other client records the mark, and the kiro-cli, goose and codex channels and KAS's behaviour are unchanged. test_acp_antigravity_backend.py gains 12 cases (an oversized tool, server, both or tool-only half, each with and without the MCP marker; 512-character names kept; the mark cleared by a later readable frame; no mark on goose or opencode), and the channel pin in test_tool_call_classification_kas.py now lists the Antigravity row as bounded. Each of the five parts of the fix was reverted in turn, and each revert fails at least one new case. client_permissions.py is 1,477 lines.

On 2d323b3, targeted tests were run on a Windows laptop (-n 8) over 200 files: the same selection as on e89e2d3 (the PR's test files plus the test files of #30086, #30108, #30237, #30616, #30670, #30674, #30241 and #29882, test_strands_mirror, test_tools_settings_notice and test_slash_command_no_channel; every test file main changed between bf9b611 and a104344; every test that reads backend_install or the KAS bundle route; and the host-contract parity, file line cap, host auth, tool gate, harness contract, crewmate harness and KAS disabled-tools tests), plus the test files of #30308 and #31016, and without test_slack_code_channel_workspace.py, whose git-subprocess tests timed out under the parallel load on that machine; neither this PR nor main has changed that file since a104344. 11,879 passed, 220 skipped, 1 xfailed, 35 failed and 8 errors. Each failure, and the collection error behind the 8 errors, was rerun alone at base 1e990f8 and at 2d323b3. 30 fail at both: five Copilot tool-gate tests and two test_pid_lifecycle reclaim tests that fail on that host, five tests that need symlink privilege (WinError 1314), 15 test_decisions_sagemaker_provider tests, two test_app_update_swap_leftovers cases, and test_sandbox_argv_launcher_plan, which calls os.geteuid and does not collect on Windows. The other six (test_memory_no_regression, test_dashboard_approval_modes, test_browser_harness_supervisor and test_dashboard_server_composition_contract_reach) pass alone at both commits; they failed on file-lock and timing errors under the load. None fails only at 2d323b3. On a95e158, the tests for the move (21 files, among them the client binding and facade tests, test_acp_antigravity_backend, the tool gate, harness contract, harness parity, host-contract parity and file line cap tests, the log, redaction and security posture scans, and the Copilot, DeepSeek, pi, goose and Claude permission tests) gave 2,657 passed, 7 skipped and 5 failed, the same five Copilot tool-gate tests that fail on that host at base. Before the rebase, with the MCP identity bound applied on a95e158, the tests for the bound (93 files: the 21 files for the move, every test file that reads the MCP identity channels, tool-call classification or permission-event code, the ACP frame tests, and the hooks and MCP tool surface tests) gave 7,126 passed, 424 skipped, 6 xfailed and 9 failed, and all 9 also fail alone at a95e158: the five Copilot tool-gate tests, a symlink-privilege test (WinError 1314), the protected-process case of test_platform_compat_windows_tree, the priming case of test_acp_client_stale_turn and the KAS run_command case of test_side_readonly_kas. On 3041371, after the rebase onto bab2387, those 93 files plus the PR's test files and the 145 test files main changed since 1e990f8 that read the ACP client, backend, host auth, harness parity or mirror code (235 files) gave 16,079 passed, 837 skipped, 9 xfailed and 13 failed. All 13 also fail alone at bab2387: the same nine, two more Copilot cases in test_acp_launch_goldens, and the two test_pid_lifecycle reclaim tests that fail on that host. 2d323b3 also merges without conflicts onto main 665a0e4; on that local merge, the PR's test files plus the host-contract parity, file line cap, sandbox governance mask, harness contract, tool gate, host auth, agent-socket and harness parity tests (28 files) gave 2,948 passed, 310 skipped and 7 failed, the same five Copilot and two test_pid_lifecycle host failures. On e89e2d3 the 197-file selection gave 11,915 passed, 197 skipped and 31 failed, all 31 also failing at base a104344, and on f065d86 the 45-file part gave 3,132 passed, 42 skipped and 7 failed, all 7 also failing at base. black, isort and flake8 pass on the 65 Python files the PR changes at 3041371, and mypy on tool_call_identity.py, client_tool_events.py, client_permissions.py, client_unasked_watch.py, client_class.py and client.py. On 650551b, check_harness_parity.py (--test and plain), docs_lint.py (--test and plain), gen-pseudolocale.mjs --check and five targeted test files (test_acp_harness_contract.py, test_acp_tool_gate.py, test_acp_antigravity_backend.py, test_agent_sdk_host_auth.py and test_spec_hook_kiro_cli_payload.py, -n 4) were run on a Windows laptop. The checks pass; the tests gave 805 passed and 2 failed, both Copilot cases refused because that laptop has a .mcp.json above pytest's temp folder (copilot_repo_config_issue checks every parent folder), which this PR does not touch. 616cb8b's patch differs from 650551b's only in the two #33215 rows, the #34367 paragraph, the #34668 row, the #34555 and #34475 rows and the #35278, #29221, #34696, #35268, #34736, #35485, #31950, #35260, #33578, #35508, #35756, #24732 and #36462, #36829, #36955 and #37105 changes and the COMPACTION_SEEN row above. 650551b's patch differs from 3041371's only in agent-host-contract.md, the two lines in acp/harness/antigravity.py that wire up adapter_expose_files, the Antigravity sign-in texts in agent_sdk/host_auth.py, and the find_file/search_directory rows in acp/harness_tool_names.py with their assertion in test_acp_antigravity_backend.py. CI on b4e6fb9 finished backend shards 1 to 4 on Linux and Windows: shards 2 and 3 passed, and shards 1 and 4 each failed only the one test named above; shards 5 to 8 were still running when 650551b replaced it.

Full run on 6167f32 (an earlier head, base 4b820ae; its diff against its base differs from 3041371's only by the #30616, #30874, #30308, #31812 and #31571 resolutions, the watch bound, the tripwire's move to its own module, the MCP identity bound and the host_auth.py comment above, while main itself moved from 4b820ae to bab2387), on Linux (x86_64, Python 3.12.16, Node 24.22.1) with the CI commands:

Not run locally: Windows and macOS CI lanes, the sandbox Sensitive Change Review, and a second signed-in Antigravity run.

Manual verification

The existing 1.4.4 live fixture records the signed-out initialize/session/load and MCP-startup exchange. One signed-in session (turn-live-*) confirms the shell, edit, read and MCP argument names on the wire. Compatibility with 1.4.5 is not verified.

Checked on Linux with the official 1.4.4 archive, signed out with an invalid API key: initialize reports version 1.4.4 and protocol 1; a server launched with Crew's spawn argv has a /proc/<pid>/cmdline of its absolute path and one NUL, and _is_managed_agent_process accepts that live pid; a session/new naming a missing stdio command still creates the session and nothing reports the dropped server. The binary needs an IPv6 loopback, so it ran in a network namespace that provides one.

NOT CHECKED: 1.4.5 wire behavior; an enterprise policy tenant; whether a sub-agent's own tool calls reach session/request_permission (the capture refused invoke_subagent, and Crew refuses it until one shows it). Sensitive Change Review needs a maintainer-owned in-repository branch.

Screenshots / video

None attached. The visible change is an Antigravity row in the backend picker and its backend card, which shows the installed, missing-executable and missing-helper states. Local captures used mocked probe data, not a signed-in session.

Related Issues

Supersedes #26171 (in-repo move; author LuisBrel).

N/A: no existing issue tracked this contribution. It continues #25980, which is closed and was not merged.

App repo: KiroCrewCaptain/kirocrew-app-aws-control#10

Checklist

  • At most two commits (one is the norm), with a Conventional Commits title (feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)
  • Existing tests pass and new tests added for new functionality
  • Self-review completed; code follows project style guidelines
  • Documentation updated (if applicable)
  • No secrets, credentials, or internal references in the diff

Contribution License Agreement

@bolichen97
bolichen97 requested review from a team and KiroCrewCaptain as code owners November 16, 2026 13:25
@bolichen97
bolichen97 enabled auto-merge (squash) November 16, 2026 13:25
@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Sensitive Change Review

Sensitive files changed:

  • src/kiro_crew/agent_sdk/tool_gate.py

  • @bolichen97 posted a complete reasoning template for 4641a47.

@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Nov 16, 2026
@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)

UX-level review of 4641a47a57d0a4ec70f940b8026196264bd9d388 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

UX-Verdict: PASS

cannot evaluate: missing screenshots of the new "Google Antigravity" harness choice. The PR adds a user-visible backend option, a backend card, an install hint and sign-in/version refusal messages, and the PR supplies no image. Nobody has looked at any of these surfaces.

Evidence gaps

  • "Google Antigravity" entry in the Agent Backend tab (pages.developer.agentBackendTab.antigravity, acpBackendName case 'antigravity'): no screenshot. Attach one of the backend list with Antigravity selected, so its card and badges are visible.
  • Antigravity backend card when the helpers are missing (_probe_antigravity lists agy_acp_server / localharness_external and shows antigravity_install_hint()): no screenshot. Attach one of the card in the MISSING state, with the install hint text.
  • Install hint and sign-in remedy saying only ACP 1.4.4 is verified, which the PR says is shown on the backend card: no screenshot. Attach one of the card or notice showing the refusal for an unsupported version.
  • Shared model picker with Antigravity's configOptions model ids (ACP_BACKENDS_CONFIG_MODEL_SELECT_IS_CANONICAL): no screenshot. Attach one of the open model/effort picker in an Antigravity session.
  • New-session notices (toolsSettings not applied; spec preToolUse hooks don't run): no screenshot. Attach one of each notice as it appears in a new chat.

[UX-REVIEWED] 4641a47

Description read

[DESCRIPTION-READ] b2d2165e0529431da5e5f5c1657d7c9d5f56c6e0cbe1f49b47c4c0777dfb8b7d

That sha256 names everything this verdict read: the PR description, plus the 3 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — ✅ PASS

Design-level review of 4641a47a57d0a4ec70f940b8026196264bd9d388 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

Design-Verdict: PASS

Nothing blocks: the new harness plugs into the existing per-backend seams and capability sets, and the codex-only branches become set membership instead of new identity checks.

Notes

  • The unasked-call tripwire fires only after a call completes, so an unasked non-read call has already run before the session stops. The PR states this. If a later Antigravity version runs more tools unasked, the version pin stops the harness from starting rather than making it less safe.

[DESIGN-REVIEWED] 4641a47

Description read

[DESCRIPTION-READ] 47b9686498c046aa1b2d4b3bcd7205b7901938a0e24a7b7f6dbb947af6862beb

That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 4641a47a57d0a4ec70f940b8026196264bd9d388 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

First-Principles-Verdict: PASS

Nothing blocks. Antigravity is a new agent.acp_backend choice and every existing default stays as it was. Each guard rests on a captured 1.4.4 frame, and the new model-select set follows the repo's named-set rule.

What this change ships

Inventory (10 items) — 9 justified

Intent: let users drive Google Antigravity's standalone ACP runtime from Kiro Crew like the other harnesses. This is an ADDITION.

  1. "Google Antigravity" can be picked as a harness, with its own backend card, install hint and sign-in remedy — justified
  2. Any Antigravity version other than 1.4.4, or one that reports no version, is refused when a chat starts — justified
  3. Permission mode is pinned to default, auto_edit and yolo are refused, and only /plan and /usage reach the harness — justified
  4. A tool call that finishes without asking stops the session; only the read-only built-ins are exempt — justified
  5. run_workflow and invoke_subagent calls are refused, and arguments named like paths or commands are fed to the gate — justified
  6. The model picker takes ids from the configOptions select instead of the combined model-and-effort list — justified
  7. Antigravity's token files go on the credential floor, the deny-diff list and the aws_control packaging fence — justified
  8. MCP servers go through session/new, and stdio servers that cannot start are pruned first — justified
  9. Session Activity counts Antigravity sessions and the dashboard glossary names it; provider labels move to provider_labels.py to stay under the line cap — justified
  10. The DeepSeek mirror table is corrected, with a regression test — rides along

[FIRST-PRINCIPLES-REVIEWED] 4641a47

Description read

[DESCRIPTION-READ] f3c3a8ea3fc5912a7c67252bf0b07c776bd30562e584fbb405e3a309cdc94706

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — ✅ no blocking findings

Reviewed 4641a47a57d0a4ec70f940b8026196264bd9d388 — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 4641a47

Verdict parsed from the review's SHA-scoped output markers for commit 4641a47a57d0a4ec70f940b8026196264bd9d388.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 4641a47a57d0a4ec70f940b8026196264bd9d388 false-positive: <reason, 20 to 500 characters>

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of 4641a47a57d0a4ec70f940b8026196264bd9d388 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 4641a47

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 4641a47a57d0a4ec70f940b8026196264bd9d388 false-positive: <reason, 20 to 500 characters>

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Security Scope Review -- ✅ PASS

Which legitimate operations does 4641a47a57d0a4ec70f940b8026196264bd9d388 newly refuse? Updated in place on each push. A script-confirmed regression blocks; the model proposes candidates and never decides a refusal.

Verdict basis: adjudicated, zero confirmed regressions.

Security scope review -- adjudicated candidates

  • macos-latest (platform posix): 12 classified, 0 newly refused, 12 still allowed.
  • ubuntu-latest (platform posix): 12 classified, 0 newly refused, 12 still allowed.
  • windows-latest (platform windows): 7 classified, 0 newly refused, 7 still allowed.

No candidate flipped from allowed to refused.

Scope-Verdict: PASS

Reason: The shell composite this lane can classify (sensitive-bash, exfil, deny-rules) is not modified by this diff, and no script-confirmed regression was handed to this review. The tightening that matters is in Antigravity's permission-event builder and its credential floor. The classifier cannot reach either one, so both are named under UNADJUDICATED below and do not block.

TIGHTENING: Antigravity permission requests are now judged as unverifiable shell calls when the tool is run_workflow or invoke_subagent, or when any argument name contains the word script or subagent. Disagreeing command spellings now drop command. Every argument whose name contains a path-like part is now fed to the sensitive-path check. -- The threat it refuses is a call whose effect no command or path rule can judge: "_ANTIGRAVITY_UNVERIFIABLE_TOOLS = frozenset({"run_workflow", "invoke_subagent"}) … is_shell = _antigravity_unverifiable … _resolved_raw_params = antigravity_unverifiable_params(_resolved_raw_params)" (tool_call_identity.py, permission_requests.py). The same diff also adds .gemini/antigravity-acp/acp_token.json and .gemini/antigravity-acp/acp_business_token.json to the credential floor through AGENT_AUTH_DECLARATIONS, and adds GEMINI_HOME to the home-override variables.

ADJUDICATED: 18 candidates, 0 confirmed newly-refused

PLATFORM COVERAGE: The diff makes three platform assumptions: the $GEMINI_HOME / ~/.gemini home layout, the Windows archive's .exe names (agy_acp_server.exe, localharness_external.exe), and file: URI decoding through url2pathname.

  • posix: covers the home layout. Rows read settings, list the token directory, echo and printenv GEMINI_HOME, filter env for it, create the directory, check the runtime version, and run the maintainers' tests and parity script.
  • windows: covers the same operations in Windows spelling. Rows use %USERPROFILE% and %GEMINI_HOME% with \ separators, type/dir, and the agy_acp_server.exe name.
  • any: one platform-neutral row, the forwarded git status --short.

One gap is not covered: the Windows file: URI decoding in _antigravity_file_uri_paths, with drive letters and UNC hosts. It works on tool arguments, not shell text, so no row can probe it.

UNADJUDICATED:

  • Guard: the Antigravity unverifiable-call marking and argument-name rewriting in build_permission_event.
    • Call path: an Antigravity tool_call frame caches the stated name (harness_tool_name, with Run <name>? / Running <name> unwrapped). Then session/request_permission goes to build_permission_event, which calls antigravity_call_is_unverifiable(params, stated_tool) (or antigravity_gate_params when that returns false), and from there the unverifiable-shell tier or the sensitive-path check.
    • Inputs I would have submitted:
      • A non-MCP Antigravity tool with an argument named ScriptPath, TestScript or script_name that only points at a file to view or edit. It is refused as unverifiable because the word match fires on script.
      • An edit whose argument is named Subagents but holds plain text.
      • A view or search call whose Include/Exclude glob such as *.py, or a url-named argument holding an https documentation link, is now merged into path. That should match nothing, but no verdict was available to confirm it.
      • A CommandLine plus a lowercase command that differ only in trailing whitespace. These now conflict and are refused.
    • Narrowing argument, not confirmed: the script word rule refuses on an argument's name, not on whether a program body is present.
  • Guard: the credential floor (is_sensitive_path, which reads host_auth.credential_leaves()).
    • Call path: a file tool's resolved path goes to is_sensitive_path.
    • Inputs I would have submitted: file-tool reads of ~/.gemini/antigravity-acp/settings.json and $GEMINI_HOME/antigravity-acp/settings.json. The declaration says these stay readable, and only the two token leaves are fenced. Also a read of an unrelated ~/.gemini/ file belonging to another Google CLI. That should stay readable because the leaves are exact files, not the directory.
    • Not adjudicated: the shell gate does not read paths, so these could only be submitted as the shell-text rows above.
  • Guard: _unasked_call_is_expected now returns False for any call whose frame named an MCP server.
    • Call path: an unasked tool_call goes to _unasked_call_is_expected.
    • Input I would have submitted: an MCP tool whose name matches a listed native read. It is now treated as unexpected, and no verdict is available.

[SCOPE-REVIEWED] 4641a47

@bolichen97

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 217c2a0
  • What rule changed: No gate rule changed. agent_sdk/tool_gate.py gains one _LABELS row, ACP_BACKEND_ANTIGRAVITY: "Google Antigravity", plus its import. _LABELS is read only by label_for(), which names the harness in operator-facing refusal text.
  • Before -> after (what was allowed/blocked, now): Before, a refusal for the antigravity backend would name it by its raw id. After, it says "Google Antigravity". Nothing allowed or blocked changes.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): A wrong label in a refusal message. No command, path or credential decision reads this table.
  • Evidence checked (which sandbox result proves it): grep -n _LABELS src/kiro_crew/agent_sdk/tool_gate.py shows the only reader is label_for (line 546), used in message formatting. The Antigravity credential mask is tested by test/test_acp_antigravity_backend.py (ADAPTER_OWN_CREDENTIAL_LEAVES / adapter_hidden_credential_dirs), which passes on this head.
  • No regression:
    • Tools still work (which tools/commands were run, result): pytest on the 22 test files this PR touches, test/test_harness_parity.py included: 2292 passed. scripts/check_harness_parity.py: pass. flake8, black --check, mypy on the touched files: clean.
    • Backward compatible (old config / old data / old callers still work, how checked): Adding a dict row changes no existing key; label_for falls back to the raw id for every other backend as before. Existing backend tests in the same run passed.
    • Existing tests (which suites passed on this head): the 22 touched test files above (2292 passed) and the frontend src/api/acpBackend.test.ts (27 passed); CI on this head is running.
  • How to undo: Revert the PR's squash commit, or delete the one _LABELS row.

This PR is the in-repo move of #26171 (author LuisBrel); the template is posted by the PR author account so the check can be judged in-repo.

AI-assisted, sent by Bolin Review bot

@bolichen97

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 1d4c1a0
  • What rule changed: No gate rule changed. agent_sdk/tool_gate.py gains one _LABELS row, ACP_BACKEND_ANTIGRAVITY: "Google Antigravity", plus its import. _LABELS is read only by label_for(), which names the harness in operator-facing refusal text.
  • Before -> after (what was allowed/blocked, now): Before, a refusal for the antigravity backend would name it by its raw id. After, it says "Google Antigravity". Nothing allowed or blocked changes.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): A wrong label in a refusal message. No command, path or credential decision reads this table.
  • Evidence checked (which sandbox result proves it): grep -n _LABELS src/kiro_crew/agent_sdk/tool_gate.py shows the only reader is label_for (line 546), used in message formatting. The Antigravity credential mask is tested by test/test_acp_antigravity_backend.py (ADAPTER_OWN_CREDENTIAL_LEAVES / adapter_hidden_credential_dirs), which passes on this head.
  • No regression:
    • Tools still work (which tools/commands were run, result): pytest on the 22 test files this PR touches, test/test_harness_parity.py included: 2292 passed. scripts/check_harness_parity.py: pass. flake8, black --check, mypy on the touched files: clean.
    • Backward compatible (old config / old data / old callers still work, how checked): Adding a dict row changes no existing key; label_for falls back to the raw id for every other backend as before. Existing backend tests in the same run passed.
    • Existing tests (which suites passed on this head): the 22 touched test files above (2292 passed) and the frontend src/api/acpBackend.test.ts (27 passed); rebased onto main after test: check the ssh floor's nesting bound below the gate's depth cap #37964 fixed a main-red ssh-floor test; on this head test_denied_commands_security_ssh_self.py, test_acp_antigravity_backend.py and test_harness_parity.py pass (701 passed).
  • How to undo: Revert the PR's squash commit, or delete the one _LABELS row.

This PR is the in-repo move of #26171 (author LuisBrel); the template is posted by the PR author account so the check can be judged in-repo.

AI-assisted, sent by Bolin Review bot

@bolichen97

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 0b43764
  • What rule changed: No gate rule changed. agent_sdk/tool_gate.py gains one _LABELS row, ACP_BACKEND_ANTIGRAVITY: "Google Antigravity", plus its import. _LABELS is read only by label_for(), which names the harness in operator-facing refusal text.
  • Before -> after (what was allowed/blocked, now): Before, a refusal for the antigravity backend would name it by its raw id. After, it says "Google Antigravity". Nothing allowed or blocked changes.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): A wrong label in a refusal message. No command, path or credential decision reads this table.
  • Evidence checked (which sandbox result proves it): grep -n _LABELS src/kiro_crew/agent_sdk/tool_gate.py shows the only reader is label_for (line 546), used in message formatting. The Antigravity credential mask is tested by test/test_acp_antigravity_backend.py (ADAPTER_OWN_CREDENTIAL_LEAVES / adapter_hidden_credential_dirs), which passes on this head.
  • No regression:
    • Tools still work (which tools/commands were run, result): pytest on the 22 test files this PR touches, test/test_harness_parity.py included: 2292 passed. scripts/check_harness_parity.py: pass. flake8, black --check, mypy on the touched files: clean.
    • Backward compatible (old config / old data / old callers still work, how checked): Adding a dict row changes no existing key; label_for falls back to the raw id for every other backend as before. Existing backend tests in the same run passed.
    • Existing tests (which suites passed on this head): the 22 touched test files above (2292 passed) and the frontend src/api/acpBackend.test.ts (27 passed); rebased onto main after test: check the ssh floor's nesting bound below the gate's depth cap #37964 fixed a main-red ssh-floor test; on this head test_denied_commands_security_ssh_self.py, test_acp_antigravity_backend.py and test_harness_parity.py pass (701 passed). The second commit only narrows Antigravity MCP hook matcher names (harness_tool_names.py); the 14 test files that cover it pass (1154 passed).
  • How to undo: Revert the PR's squash commit, or delete the one _LABELS row.

This PR is the in-repo move of #26171 (author LuisBrel); the template is posted by the PR author account so the check can be judged in-repo.

AI-assisted, sent by Bolin Review bot

@bolichen97

Copy link
Copy Markdown
Contributor Author

@iamwhatever could you review this one? It is LuisBrel's Antigravity harness from #26171, moved to an in-repo branch so Sensitive Change Review can run (it touches agent_sdk/tool_gate.py). Our account pushed last, so our approval does not count here.

  • Commit 1 is LuisBrel's original, rebased; the only conflict was one paragraph in acp-client.md.
  • Commit 2 fixes the one finding GPT 6.1 raised: an MCP server and tool name that join into a native tool name (@view/file -> view_file) no longer trigger the file-read hook, with a test.
  • The open item from our earlier change request (a signed-in captured turn) is done: the live fixtures cover shell, edit, read and MCP, and the synthesized fixture is gone.
  • Sensitive Change Review and the AI reviews pass; no reds, backend shards still running. Auto merge is on.

AI-assisted, sent by Bolin Review bot

@github-actions github-actions Bot added merge conflict Branch has merge conflicts with its base — author must resolve before merge readiness: passed Eligible automated validation passed for the current revision and removed readiness: checking Automated validation is still running labels Nov 16, 2026
@LuisBrel

Copy link
Copy Markdown
Contributor

Thanks @bolichen97. Two macOS shards are red (test_options_marker_label_closers.py and test_subagent_injection_channel_timeout_resend.py). The files they test are outside this patch, so they look like flakes.

The branch also conflicts with main now, in acp-client.md and agent-host-contract.md only. I resolved both on my fork: 04b3785f, branch resolve/antigravity-37974-docs, one commit on main 028c39db with the same patch, and it merges cleanly with the current main. In acp-client.md the one change is the permissions row, where main reworded the tripwire sentence beside this PR's Antigravity phrase. In agent-host-contract.md main's rows are kept and the Antigravity cell is appended to each.

LuisBrel and others added 2 commits November 16, 2026 16:08
Register the standalone Antigravity ACP runtime as a public harness, pin the verified release, route permission mode through session config, and declare its auth, MCP, model, effort, hook, and install contracts.

Add live handshake, session, load, and MCP-startup evidence and one signed-in 1.4.4 turn corpus (shell, read, edit, refused calls and a stdio MCP call), plus focused behavior and parity tests.

Refuse run_workflow, which runs the Python in its Script argument, and invoke_subagent, whose sub-agents' own tool calls no capture has shown asking, as shell calls whose command cannot be verified.
An Antigravity MCP call is named from its trusted _meta.mcp pair, so a server and tool that fuse into a native name (@view/file into view_file) must not meet the native read hook. Return only the MCP matcher forms, as the envelope-named harnesses already do.
@KiroCrewCaptain

KiroCrewCaptain commented Nov 16, 2026 •

Copy link
Copy Markdown
Collaborator

Code change approver recommendation: A person should decide

Why: the decider estimates a 12% chance that this change introduces a defect, within the 12% limit for approving on its own, but it changes 6049 lines, more than the 400 it may approve on its own.

Adds Google Antigravity (agy_acp_server 1.4.4) as an opt-in ACP backend using the existing per-harness seams. Verified in code: the version pin is enforced before the mode=default write; auto_edit and yolo are refused; prompt commands are limited to plan and usage. Main's unasked-call tripwire is armed at spawn and excludes shell and MCP calls from the expected unasked reads, which are limited to Antigravity's read, grep and glob built-ins. run_workflow, invoke_subagent and arguments named script or subagent are refused as unverifiable shell calls, on Antigravity only and not for MCP calls. Shared code changes keep existing behaviour: requires_server preserves the KAS row; the hook vocabulary is the same for existing backends plus Antigravity; the extra harness_backend argument is inside a KAS-only branch. The own-token exclusion on the credential mask follows the accepted precedent for codex, OpenCode, pi, goose and Copilot. Default backend unchanged, every check green, extensive tests added. No blocking issue found.

An automated recommendation, not a review: a maintainer still decides. It approves on its own only when the estimated chance of a defect is at most 12% and the change has at most 400 changed lines. Decider v3c (217fe607) · commit 4641a47a57 · 6049 changed lines. If this recommendation is wrong, reply here: replies are how the approver is evaluated.

@bolichen97
bolichen97 force-pushed the crew/CommunityCrew/antigravity-harness-26171 branch from 0b43764 to 4641a47 Compare November 16, 2026 16:14
@bolichen97

Copy link
Copy Markdown
Contributor Author

Sensitive change review

  • Head commit: 4641a47
  • What rule changed: No gate rule changed. agent_sdk/tool_gate.py gains one _LABELS row, ACP_BACKEND_ANTIGRAVITY: "Google Antigravity", plus its import. _LABELS is read only by label_for(), which names the harness in operator-facing refusal text.
  • Before -> after (what was allowed/blocked, now): Before, a refusal for the antigravity backend would name it by its raw id. After, it says "Google Antigravity". Nothing allowed or blocked changes.
  • Worst case if wrong (breaks commands / leaks secret / locks user out): A wrong label in a refusal message. No command, path or credential decision reads this table.
  • Evidence checked (which sandbox result proves it): grep -n _LABELS src/kiro_crew/agent_sdk/tool_gate.py shows the only reader is label_for (line 546), used in message formatting. The Antigravity credential mask is tested by test/test_acp_antigravity_backend.py (ADAPTER_OWN_CREDENTIAL_LEAVES / adapter_hidden_credential_dirs), which passes on this head.
  • No regression:
    • Tools still work (which tools/commands were run, result): pytest on the 22 test files this PR touches, test/test_harness_parity.py included: 2292 passed. scripts/check_harness_parity.py: pass. flake8, black --check, mypy on the touched files: clean.
    • Backward compatible (old config / old data / old callers still work, how checked): Adding a dict row changes no existing key; label_for falls back to the raw id for every other backend as before. Existing backend tests in the same run passed.
    • Existing tests (which suites passed on this head): the 22 touched test files above (2292 passed) and the frontend src/api/acpBackend.test.ts (27 passed); rebased onto main after test: check the ssh floor's nesting bound below the gate's depth cap #37964 fixed a main-red ssh-floor test; on this head test_denied_commands_security_ssh_self.py, test_acp_antigravity_backend.py and test_harness_parity.py pass (701 passed). The second commit only narrows Antigravity MCP hook matcher names (harness_tool_names.py); the 14 test files that cover it pass (1154 passed). After the rebase onto main 63d5b45 (doc-only conflicts in acp-client.md and agent-host-contract.md), 34 test files pass on this head (3105 passed), plus mypy, harness parity and docs lint.
  • How to undo: Revert the PR's squash commit, or delete the one _LABELS row.

This PR is the in-repo move of #26171 (author LuisBrel); the template is posted by the PR author account so the check can be judged in-repo.

AI-assisted, sent by Bolin Review bot

@bolichen97

Copy link
Copy Markdown
Contributor Author

@iamwhatever thanks for reviewing. What changed since your review request:

  • Rebased onto current main to clear the conflict. LuisBrel's commit keeps its author and code. The only edits were two doc conflicts: the client_permissions.py row in acp-client.md, and the Antigravity column added to main's newer rows in agent-host-contract.md.
  • The two red macOS shards were test_subagent_injection_channel_timeout_resend (a resend race) and test_options_marker_label_closers::TestLinearity (a CPU-timing budget). This PR touches neither area. Both pass five times in a row locally, and the same workflow is green on other branches.
  • The second commit (from GPT review) stays: an Antigravity MCP call whose server and tool fuse into a native name (@view/file -> view_file) now meets only MCP hook matchers.

New head: 4641a47. Locally, 34 touched test files pass (3105 tests), plus mypy, harness parity and docs lint.

AI-assisted, sent by Bolin Review bot

@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: passed Eligible automated validation passed for the current revision and removed readiness: passed Eligible automated validation passed for the current revision merge conflict Branch has merge conflicts with its base — author must resolve before merge readiness: checking Automated validation is still running labels Nov 16, 2026

@KiroCrewCaptain KiroCrewCaptain left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approving. On head 4641a47 the review read the full non-test diff and traced the permission event builder, the tool-call identity cache, the unasked-call tripwire, mode routing, the version pin, the credential floor, the sandbox preflight and model selection. It ran test/test_acp_antigravity_backend.py (208 passed) and the 12 test files this PR touches (1525 passed, apart from 4 that fail only because the review checkout did not include the parity script). It also ran 14 mutations of the guard lines, and 13 of them turned a test red. Every harness identity check is positive (== ACP_BACKEND_ANTIGRAVITY or a named set), and the Kiro path is unchanged.

Security: this PR changes the permission gate, the credential floor and the tripwire, and every change makes them tighter. Antigravity runs only in default mode, and auto_edit and yolo are refused whether they arrive at open or mid-session. run_workflow, invoke_subagent, and any argument whose name has the whole word script or subagent are judged as shell calls with no command, so the deny-by-default tier refuses them. When the command spellings in a request disagree, command is dropped. Path-named arguments only add targets to the sensitive-path check. The child starts behind the sandbox preflight and the credential mask, and the two Antigravity token leaves join the floor for every other harness. Unasked file reads are disclosed in the PR. They are kept out of the spec-hook and zero-tool sets, and crew-home secrets stay behind the OS mask.

Done well: each safety choice rests on a signed-in 1.4.4 wire capture, and the replay test runs those sessions through the real client to show that the tripwire does not fire on the harness's real unasked reads.

Optional:

  • The UX lane could not evaluate this head because the PR has no screenshots. Please attach the mocked-probe captures you already have: the picker row, the card in its MISSING state with the install hint, and the version-refusal text.
  • In tool_call_identity.py:157, if a frame ever carries is_mcp_tool_call without a readable _meta.mcp, the codex fallback reads rawInput.server/tool. On Antigravity those fields hold the model's own flattened arguments. One fix is to skip that branch when _meta.mcp is a dict, with a pinning test.
  • The PR body still describes one commit at head 2eeb923. It does not mention 4641a47, which keeps Antigravity MCP calls out of the fused native hook aliases. It also lists the shared tripwire early return, which tightens MiMo Code too, under no Reader line. Editing the body fixes both, with no push needed.

@bolichen97
bolichen97 merged commit bb6e096 into main Nov 16, 2026
110 of 111 checks passed
@bolichen97
bolichen97 deleted the crew/CommunityCrew/antigravity-harness-26171 branch November 16, 2026 18:39
@github-actions github-actions Bot removed the readiness: passed Eligible automated validation passed for the current revision label Nov 16, 2026
LuisBrel added a commit to LuisBrel/kirocrew-app-aws-control that referenced this pull request Nov 16, 2026
…he synthetic-credential fixtures

The pin moves from 9b32941be6 to bb6e096df4 (the merge of kirodotdev/KiroCrew#37974, the in-repo
move of #26171), whose shared validator refuses the two Antigravity token files the local fence
names. That host no longer treats the AWS documentation example key or an AWS key ID as a
secret, so the redaction tests take #31863 and the matching test changes of #37408 (tests
only): their credentials come from kiro_crew.testing.synthetic_credentials. Recorded in
PORTING.md.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants