Repository navigation
feat(cp): runtime client for Phase 0 (openab run worker, openab agent CLI) - #1593
Open
chaodu-agent wants to merge 15 commits into
Open
chaodu-agent wants to merge 15 commits into
chaodu-agent wants to merge 15 commits into
Conversation
Port of the openab-core runtime control-plane client from feat/cp-runtime-client-main 9455b50 (PR 1471) onto origin/main 6214700. Plan: .agents/tasks/phase0-cp-runtime/plan.md, Commit 1. Taken as-is: - control_plane/mod.rs, client.rs, executor.rs (new files) - config.rs: ControlPlaneConfig, CpAgentType, the [control_plane] validation, and the control_plane_* config tests. The AgentCoreConfig whitespace hunk in the same commit is not taken. - lib.rs: pub mod control_plane - crates/openab-core/Cargo.toml: runtime openab-cp dependency with default-features = false. The dev-dependencies (openab-cp server, axum) are left for commit 6, because only the cp_client test needs them. - cron.rs, directives.rs: the two format! to to_string clippy hunks Applied 3-way onto main, keeping main code: - acp/pool.rs: see deviation below - adapter.rs: PromptExecution and the stream_prompt_blocks return value - dispatch.rs: the .map(|_| ()) adapter Deviations from the plan: - acp/pool.rs: main already has discard_session returning Result<()>, which the dispatch rollback uses. The branch version returns bool and does not send session/cancel, so it is added as discard_single_use_session. Main's discard_session is unchanged. The executor is the only caller and was updated. - crates/openab-cp/src/proto.rs: MIN_RUNTIME_FRAME_BYTES (4 lines) ported from the branch. The client uses it, and main does not define it. The plan said #1588 added it, but main proto.rs has no change since the merge base. - Cargo.lock: cargo added one line, "openab-cp", under openab-core. - executor.rs: rustfmt applied. It is the only new fmt diff this commit adds. Verification (macmini, cargo and rustc 1.99.1, clone ~/repo/openab-phase0): - cargo check -p openab-core -p openab: pass - cargo clippy -p openab-core -p openab -- -D warnings: pass - cargo test -p openab-core --lib control_plane: 50 passed, 0 failed - cargo test -p openab-core --lib: 754 passed, 1 failed: secrets::tests::resolve_exec_nonzero_exit. The same test fails on a clean origin/main export (6214700), so it predates this commit. Not fixed here. - cargo fmt --check: no new diffs in files this commit touches, apart from executor.rs, which rustfmt was applied to. Main also has fmt diffs. - Linux run: not needed for this commit.
Port of the src/main.rs hunks from feat/cp-runtime-client-main 9455b50 (PR 1471) onto origin/main 6214700. Plan: .agents/tasks/phase0-cp-runtime/ plan.md, Commit 2. What it adds: - HeadlessMode and headless_run_mode: an adapter-less config picks the full runtime for [control_plane] type = worker, or for [mcp] plus [control_plane]. [mcp] alone stays facade-only. A primary alone stays an error, because its prompts come from a chat platform. - supervise_control_plane and wait_for_control_plane_exit: a client task that exits or fails stops the process after the ordered cleanup. - The control-plane client is spawned after the router and stopped before pool.shutdown, so the CP sees a clean disconnect. - Discord and no-Discord runtime waits race the client task with select!. Main code kept: the main() structure, the pool/router construction, and the existing shutdown order. The only removed lines are the facade block moved into the match, the old no-adapter bail text, and the bare shutdown_signal awaits that now sit inside select!. The bail text now also names [line], [lineworks], [feishu] and [teams], as the branch wrote it. Verification (macmini, cargo and rustc 1.99.1, clone ~/repo/openab-phase0): - cargo check -p openab-core -p openab: pass - cargo clippy -p openab-core -p openab -- -D warnings: pass - cargo test -p openab: 26 passed, 0 failed. Includes tests::nothing_configured_is_still_a_startup_error and the 8 headless mode tests. - cargo test -p openab-core --lib control_plane: 50 passed, 0 failed - cargo test -p openab-core --test cp_client: no such target yet. It is added in commit 6. - cargo fmt --check: 14 diffs in src/main.rs, the same count as main. No new diffs. - Linux run: not needed. Known failure on main, not caused by this series: secrets::tests::resolve_exec_nonzero_exit Fails on origin/main 6214700 with the same assertion. Not fixed or skipped.
Plan commit 3, from reference 3742e03 (#1529), restricted to: - new control_plane/local.rs and primary.rs, taken whole - control_plane/client.rs and mod.rs hunks, applied as a patch. The parent blobs of both files matched HEAD, so the patch applied cleanly. - src/main.rs: Agent subcommand, AgentAction, run_agent_command, parse_agent_labels, validate_deadline_secs, local API supervision and shutdown, and six unit tests from the same commit. Not taken: openab-mcp facade.rs, src/cp_tools.rs, src/acp_tunnel_source.rs, the headless_run_mode change, and the MCP facade wiring. spawn_within_bound and new_delegation_id are copied from cp_tools.rs into main.rs because run_agent_command calls them. Five rustfmt fixes apply to added lines only. Checks, run on macmini (rustc 1.99.1) in ~/repo/openab-phase0: - cargo check -p openab-core -p openab: pass, no warnings - cargo test -p openab-core --lib control_plane: 106 passed, 0 failed - cargo test -p openab: 32 passed, 0 failed - cargo run -p openab -- agent --help: lists spawn, list, status, cancel - cargo fmt -- --check: no hunks in added code; remaining hunks predate this commit - Cargo.lock unchanged
Port of reference c061158, control_plane/* hunks only (cherry-pick -x not used; hunks applied from its diff). Not taken: src/main.rs (both hunks wire the F2 FacadeMode status, which needs acp_mcp.rs; skipped with acp_mcp.rs and pool.rs), acp/pool.rs, acp_mcp.rs, docs. Checks on macmini (cargo 1.99.1): - cargo check -p openab-core -p openab: pass, no warnings - cargo test -p openab-core --lib control_plane: 119 passed, 0 failed - Cargo.lock unchanged (no new dependencies) - git diff origin/main -- crates/openab-core/src/control_plane: 0 removed lines Linux run for this commit: not run (the plan assigns the Linux fix to commit 5).
Take 91c0a89 (local.rs only) from the closed reference branch. Shutdown now drops our listener first, then unlinks the socket only if its (dev, ino) still matches the bound node AND a connect probe is refused. A successor that answers the probe is left alone, which covers ext4 reusing the freed inode number for the next create in the same directory. Reference CI (run 34434639420, head 91c0a89) reported the cap test `connections_past_the_cap_are_refused_with_a_diagnostic` failing on Linux. Root cause: NOT reproduced. The cap test passes on p1 (ext4) at this commit, so no change beyond the reference hunk was made. The CI cause remains UNVERIFIED. Commands and results: - macmini: cargo check -p openab-core -p openab: pass - macmini: cargo test -p openab-core --lib control_plane: 120 passed, 0 failed - p1 (ext4, Linux): cargo test -p openab-core --lib control_plane: 120 passed, 0 failed - p1: cargo test -p openab-core --lib control_plane::local, 25 repeated runs: 25/25 pass - p1: cap test alone under 8-way parallel load, 40 runs (--exact, full name): 40/40 "1 passed" - cp_client tests: not run, tests/ directory is added in a later commit Cherry-picked from 91c0a89 (see trailer).
Adds crates/openab-core/tests/cp_client.rs, taken from the reference branch feat/cp-runtime-client-main (9455b50 and 3742e03, the only commits that touch the file). The test boots openab-cp in-process on an ephemeral loopback port and drives the real ControlPlaneClient against it. Dev-dependencies added to crates/openab-core/Cargo.toml: openab-cp with the server feature, and axum 0.8 for axum::serve. Cargo.lock gains only the openab-core -> axum edge. Checks run on macmini (rustc 1.99.1): - cargo check -p openab-core -p openab: pass - cargo test -p openab-core --test cp_client: 11 passed, 0 failed - cargo test -p openab-core --lib control_plane: 120 passed, 0 failed Not run: Linux (not required for this commit).
Add two in-process tests to crates/openab-core/tests/cp_client.rs, using the real CP server and the real ControlPlaneClient with a scripted runner: - phase0_parallel_spawns_land_on_different_workers: two workers with max_delegated_sessions = 1 and the same label. Two cp/delegate frames are written before either ack is read. Asserts the two acks name worker-1 and worker-2, and that each worker runner started exactly one delegation. - phase0_worker_identity_cannot_initiate: a raw worker-type connection sends cp/delegate. Asserts error code -32003 (POLICY_DENIED) and no in-flight entry. R9 (least-loaded tie-break): no change needed. The first test passed without modification. Router::delegate runs select and capacity reservation under the admission guard, so the first spawn is counted before the second selection. Commands run on macmini (cargo 1.99.1): - cargo test -p openab-core --test cp_client phase0_parallel_spawns_land_on_different_workers: ok - cargo test -p openab-core --test cp_client phase0_worker_identity_cannot_initiate: ok - cargo test -p openab-core --test cp_client: 13 passed, 0 failed - cargo test -p openab-core control_plane: 120 passed, 0 failed - cargo check -p openab-core -p openab: ok - rustfmt --check on cp_client.rs: ok (after rustfmt applied)
Adds the [control_plane] section to docs/config-reference.md, taken from the reference branch feat/cp-runtime-client-main (91c0a89) and placed before the [cron] section. Only the [control_plane] section is included; no MCP-tool docs. Edits to the reference text, so the docs match HEAD: - Headless table: a primary with no adapter and no [mcp] is a startup error on HEAD (headless_run_mode returns HeadlessMode::None), not "Full runtime" as the reference docs say. The reference code at 91c0a89 differs; reported separately. - Dropped the MCP-tool paragraphs and rows (facade tools, spawn_agent). - Delegation deadlines bullet now covers only `openab agent spawn`. - auth_key: removed the [secrets.refs] mention; no code path for it was found. - Kept the "One runtime per socket path" bullet from c061158 (F6), which documents behavior shipped in commit 4 of this branch. Verification: - cargo check -p openab-core -p openab (macmini, rustc 1.99.1): Finished, exit 0. Cached result: this commit changes no Rust source. - CLI flags named in the docs (spawn --deadline-secs, --label, 1..=1800 validation) checked against src/main.rs: match. - Code claims (0700/0600 socket perms, 10 s connect timeout, allow_insecure_transport gate, deadline + 5 bound) checked against HEAD: match. - git diff origin/main -- docs/config-reference.md: additions only, no removed lines. - No automated docs test exists.
This comment has been minimized.
This comment has been minimized.
…s M1, L1-L3, L5, L9
M1: wait_for_control_plane_exit and wait_for_local_api_exit awaited a oneshot
receiver inside tokio::select!. Once it yielded, the shutdown path awaited the
same receiver again, which tokio panics on ("called after complete"), so
pool.shutdown() and pre_shutdown were skipped. Each wait now drops its
supervision after the receiver yields, and shutdown skips it.
Test: tests::a_supervised_exit_is_not_awaited_again_at_shutdown.
L1: openab agent --socket no longer evaluates default_socket_path() when an
explicit socket is given, so HOME unset no longer fails an explicit path.
L2: the headless-primary comment and the no-adapter error now state that a
control-plane primary with no adapter also needs [mcp]. Boot behaviour is
unchanged.
L3: MIN_RUNTIME_FRAME_BYTES doc says the server does not enforce the floor.
L5: client.rs comment matches behaviour: spawn and cancel are refused locally,
list_agents is forwarded for either role.
L9: spawn and delegation-id doc comments no longer claim the MCP facade calls
them. No facade caller exists at this commit.
Checks on the tree at this commit (macmini, cargo via ~/.cargo/bin):
- cargo clippy -p openab-core -p openab -- -D warnings: pass (exit 0)
- cargo test -p openab-core --lib control_plane: 120 passed, 0 failed
- cargo test -p openab-core --test cp_client: 14 passed, 0 failed
- cargo test -p openab: 33 passed, 0 failed (includes the M1 test)
- cargo test --workspace: 895 passed, 1 failed:
secrets::tests::resolve_exec_nonzero_exit (pre-existing; secrets.rs is not
in this PR; the test expects the "exited with" text on this macOS host)
- cargo fmt --all -- --check: fails across most of the repo at the base commit
(pre-existing). The lines this commit adds are wrapped to rustfmt style.
Commits in this pass were not built one by one. The checks above cover the
final tree of the pass.
Adds a_worker_runtime_refuses_spawn_locally. The worker's own client answers a Spawn with the primary-only error and sends nothing to the CP. Until now only the CP-side refusal (-32003) was tested. Checks (macmini): cargo test -p openab-core --test cp_client: 14 passed, 0 failed, including a_worker_runtime_refuses_spawn_locally and the two primary_local_api_* tests that cover primary-side spawn, list, await and cancel over the real CP (the AC4 evidence on a primary). cargo clippy -D warnings: pass on the same tree.
A cp/delegate from a primary enters the worker's normal agent turn with the agent's configured tool permissions. Nothing narrows tools per delegation. The config reference now says so, so the maintainer can accept it knowingly. Docs only. Code checks are the same as the preceding fix commit (clippy -D warnings pass, cp_client 14 passed, control_plane 120 passed, openab 33 passed on macmini).
This comment has been minimized.
This comment has been minimized.
Phase 0 criterion 4: stopping one worker fails its in-flight delegation as target_disconnected at the initiator, within the lease. No test covered it. The new test delegates to a hanging worker, stops that worker, and asserts the initiator's cp/delegate_result carries target_disconnected within 10s. The lease defaults to 30s and the delegation deadline is 300s, so a 10s bound rules out both as the source of the failure. Checks run on macmini at this commit: - cargo test -p openab-core --test cp_client a_stopped_worker: 1 passed, 0 failed - cargo test -p openab-core --test cp_client: 15 passed, 0 failed (after rustfmt layout fix) - rustfmt --check crates/openab-core/tests/cp_client.rs: clean (exit 0)
…(L2, L14) L2: the comment on a_control_plane_primary_alone_does_not_unlock_a_headless_boot still gave the chat-only rationale. It now matches HeadlessMode: the primary's initiating surface is the local socket and `openab agent`, which the headless boot serves only with [mcp] or a chat adapter. The production wording already said this. L14: reflowed the new_delegation_id doc paragraph and removed the orphaned line break after "while". Comments only; no behaviour change. Checks run on macmini at this commit: - cargo clippy -- -D warnings: pass - cargo clippy --workspace -- -D warnings: pass - cargo test -p openab: 33 passed, 0 failed - cargo test -p openab-core --lib control_plane: 120 passed, 0 failed
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
thepagent
marked this pull request as ready for review
October 31, 2026 18:18
Collaborator
Author
|
Important CHANGES REQUESTED What This PR DoesPorts the control-plane runtime client from closed PR #1528 onto current How It Works
Findings
Finding Details🟡 F1: Await/Check during backoff contradicts preserved history
Requested change: answer 🟡 F2: Register ack sanity
Requested change: log a 🟡 F3: Bounded but slow shutdown closeAfter drain, Requested change: use a separate short close timeout (2-5s) for the close handshake. 🟡 F4: Failure-path test gaps
🟡 F5: Timeout hierarchyThe CLI wraps a spawn in Requested change: document the hierarchy in one place (the module doc or 🟡 F6: Decide the initiating-surface owner before the facade lands
Requested change: record in the PR body or the follow-up which surface owns spawn/list/cancel validation, so the facade PR reuses one path and does not fork it. 🟡 F7:
|
| Reviewer | Verdict | Key findings |
|---|---|---|
| Reviewer A | CHANGES REQUESTED | F1-F8, F10 |
| Coordinator baseline | CHANGES REQUESTED | F9, verification of all claims at 18955be3 |
Three claims raised during this round were checked and dropped as incorrect: missing startup warning or docs for allow_insecure_transport (both exist), a CI workflow change (none in this diff), and a lease-death race from the 30s write timeout (lease default is 45s, and the lease outcome is server-side).
5️⃣ Three Reasons We Might Not Need This PR
- Size versus Phase 0 goal - about 6.7k lines across four modules prove a hypothesis that a smaller outbound-only worker could test. The primary tracker and local socket mainly pay off in later phases.
- No live validation yet - AC4 and the manual half of AC6 are unverified, and nothing has run against a real host. Merging does not prove the Phase 0 claim.
- Surface may churn - the facade port (F6) and the open headless-primary decision (L2) can reshape the CLI and socket API soon after this lands.
…und close; warn on zero budget - F1: during reconnect backoff, Await and Check read the process-lifetime PrimaryState instead of answering NotConnected, so a delegation that completed before a blip stays readable. fail_all_live() now also runs on backoff entry, so an early Err out of connect_and_serve cannot leave a parked awaiter behind. - F2: warn when the register ack grants effective_max_delegated_sessions = 0. - F3: the end-of-session close handshake runs under one 3s CLOSE_TIMEOUT instead of two sequential 30s WRITE_TIMEOUT bounds. - F10: the Err arm applies the same stability-based backoff reset; an invalid cp/delegate_result logs its serde error; the ListAgents comment states that the CP serves it to every role.
… across reconnect - a_rejected_registration_backs_off_instead_of_spinning - unknown_methods_and_malformed_frames_do_not_end_the_session - a_completed_delegation_is_readable_during_reconnect_backoff - a_worker_runtime_refuses_spawn_locally now also pins that the CP serves cp/list_agents to a worker
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This brings the control-plane runtime client that was on the closed PR #1528 onto current main, in 8 commits, scoped to Phase 0 of openabdev/openab-pty#77.
What is NOT included
acp_mcp.rs,pool.rsfacade changes,cp_tools.rs. Commit 4 skippedmain.rs,acp/pool.rs,acp_mcp.rsand docs because the F2 fix needs the facade.openab-cpcrate changes. These are already on main via refactor(cp): make openab-cp proto usable without the server #1588. The one exception is a 4-line port ofMIN_RUNTIME_FRAME_BYTESinopenab-cp/src/proto.rs, made in commit 1.openab-ptychanges.Commits
Test results
The full workspace suite ran on macmini at
b1f3dc15: 895 passed, 1 failed (secrets::tests::resolve_exec_nonzero_exit, pre-existing). CI run 41216928821 at18955be3rancargo test --workspaceand concluded success.cargo fmt --all -- --checkfails across most of the repo at the base commit (pre-existing, perb1f3dc15)..github/workflows/ci.ymlhas no fmt step, so that failure does not gate this PR. Other commits ran only the targeted commands below.Per-commit results (macmini = macOS arm64, p1 = Linux x86_64):
openab runwith [control_plane]openab agentspawn/list/status/cancel over the local socketphase0_tests passphase0_2/2; cap test passTest selection note: plain
cargo test phase0_runs zero Phase 0 tests because they live in theopenab-corecp_clienttarget. The correct form iscargo test -p openab-core --test cp_client phase0_. That form was not run standalone. Its selection ran as part of commit 7 and on p1.Not verified:
openab agentwas run against a live socket, and no Kiro login was done. Acceptance criteria 1, 2 and 4 are manual.openab agent --helpwas not run.Known failing test on main, not caused by this PR
secrets::tests::resolve_exec_nonzero_exitfails on origin/main6214700awith the same assertion. This PR does not change it.Deviations from the plan
switch -c, so it tracksorigin/main. It was pushed withgit push -u origin feat/cp-runtime-phase0./tmp/phase0.bundlewas created on macmini and copied to p1 to sync the checkout.MIN_RUNTIME_FRAME_BYTESand its test are absent from origin/main. Commit 1 adds a 4-line port inopenab-cp/src/proto.rs.[line],[lineworks],[feishu]and[teams]. Author isTautui <tautui.dev@gmail.com>, not amended.headless_run_modechange was not taken. Helpers were copied fromcp_tools.rsintomain.rs,serve_localis wired in, and six unit tests are included.main.rs,acp/pool.rs,acp_mcp.rsand docs, because F2 needs the facade.--exact) and was rerun with the full name.axum = "0.8"dev-dependency on openab-core. The plan named onlyopenab-cp.connect_rawhelper. Two spawns run sequentially on one socket.[secrets.refs]text were dropped.Open decision, unresolved: a headless
primarywith no adapter and no[mcp]still refuses to boot on HEAD. The options are (a) take the referenceheadless_run_modechange, which alters main's boot behavior, or (b) add[mcp]to the macmini primary config. This blocks plan Host setup step 6.Pending human input: tailnet HTTPS, p1 and mini host changes (kiro-cli install, login, linger), and the Part 2 re-land check (R13).
Review Contract
Goal
Bring the control-plane runtime client from closed PR #1528 onto current main, scoped to Phase 0 of openabdev/openab-pty#77. The result: an
openab runwith[control_plane]can run headless as a worker, and theopenab agent spawn/list/status/cancelCLI works over the local socket. Both behaviors are covered by tests on macmini (macOS) and p1 (Linux, partially).Non-goals
acp_mcp.rs,pool.rsfacade changes,cp_tools.rs). Needed for F2, deferred.openab-cpcrate beyond the 4-lineMIN_RUNTIME_FRAME_BYTESport. The crate already landed on main via refactor(cp): make openab-cp proto usable without the server #1588.openab-pty.b1f3dc15ran it on macmini, and CI runscargo test --workspace(see Test results).openab agent spawnandcancelagainst a worker's local socket. The runtime refuses both on a non-primary (crates/openab-core/src/control_plane/client.rs:611for Spawn, which sendscp/delegate, and:658for Cancel). Testa_worker_runtime_refuses_spawn_locally(crates/openab-core/tests/cp_client.rs:1064) pins the spawn refusal. No test covers the cancel refusal. This is by design for Phase 0, which validates primary-side spawn/list/status/cancel and worker-side delegation only. The worker-side socket command surface is out of scope for this PR and must be a follow-up.Accepted Residual Risks
secrets::tests::resolve_exec_nonzero_exitfails on6214700atoo. Accepted as pre-existing and out of scope. Mitigation: none in this PR. Its cause is not diagnosed. Recovery: none needed from this PR.openab agentran against a live socket, and no Kiro login was done. Acceptance criteria 1, 2 and 4 are manual. Mitigation: manual check before any host deployment. Recovery: revert the affected commit.[mcp]refuses to boot on HEAD. The fix choice is still open, and this PR does not decide it. Mitigation: the operator config must include[mcp], or the reference change is taken after a decision. Recovery: revert the decision commit.resolve_exec_nonzero_exitare not diagnosed.[agent]'s tool permissions, as the OS user that runsopenab run, including shell and file access on the worker host. The code does not narrow tools per delegation.RouterPromptRunner::run(crates/openab-core/src/control_plane/executor.rs:659) callsstream_prompt_blockson a fresh pool session with the configured defaultworking_dir(line 629). The executor sets no tool restriction. The ACP connection answers every agentsession/request_permissionwithout a human prompt, choosing an allow option withallow_alwaysfirst (auto-respond atcrates/openab-core/src/acp/connection.rs:276-286,build_permission_responsefrom line 49,pick_best_optionlines 17-46). Delegated prompts use the same connection. The shipped Dockerfiles (Dockerfile:76,Dockerfile.unified:129), the chart (charts/openab/values.yaml:274), the setup wizard (crates/openab-core/src/setup/config.rs:88) and the docs (docs/config-reference.md:360) pass--trust-all-tools. The code default is empty:default_agent_args()(crates/openab-core/src/config.rs:2042-2050) returns no args unlessOPENAB_AGENT_COMMANDis set. Channel and user allowlists run in the chat message path beforehandle_message(AGENTS.md message flow), not on this delegation path. The only mitigation is operator-side:docs/config-reference.md:834-838says to give a worker's key only to primaries you trust. Mitigation in code: none. Recovery: none verified in this PR. Removing--trust-all-toolsfrom the worker's[agent].argsmay change which tool requests kiro-cli makes. It does not narrow what OpenAB grants, because the connection auto-approves each permission request, and no config option inacp/orconfig.rsnarrows those grants. A real narrowing lever needs a code change not in this PR. Not verified: this PR did not test what kiro-cli does with tool requests once that flag is removed. Maintainer acceptance required.Acceptance Criteria
cargo clippy -- -D warningspasses on macmini at HEAD18955be3. Status:cargo clippy --workspace -- -D warningspassed at18955be3(recorded in its commit message).cargo checkis not recorded at18955be3; it passed (cached) at9585c377. CI run 41216928821 at18955be3concluded success.cargo test -p openab-core --test cp_client phase0_passes. Status: bothphase0_tests passed at commit 90a0100 and on p1 at 9585c37 (p1 was not run at HEAD18955be3).cargo test -p openab-core control_planepasses at HEAD. Status: 120/120 on macmini at18955be3(run ascargo test -p openab-core --lib control_plane, recorded in its commit message). Earlier: 120/120 on p1 at9585c377, and on macmini at 5222892, f2378ad and 90a0100.openab runheadless as a worker starts.openab agent spawn/list/status/cancelworks from a primary's live socket. Status: NOT verified.secrets::tests::resolve_exec_nonzero_exit, still fails identically on6214700aand is not introduced by this PR. Status: confirmed by the report.target_disconnectedwithin the lease window (45 s). Automated:a_stopped_worker_fails_its_inflight_delegation_as_target_disconnected(crates/openab-core/tests/cp_client.rs:639) stops a hanging worker and requires the initiator to receivetarget_disconnectedwithin 10 s, inside the 45 s lease default (crates/openab-cp/src/config.rs:164). Commit8354dc8arecords it passing on macmini. Manual half (stop a real host): UNVERIFIED. Steps: start a primary and a worker on separate hosts; delegate to a worker whose agent hangs; stop the worker host (shut the machine down or cut its network); confirm the primary reportstarget_disconnectedwithin 45 s. The lease-expiry path (crashed or partitioned host) is not covered end-to-end.Follow-ups
headless_run_modechange, or require[mcp]in the primary config.resolve_exec_nonzero_exit.openab agent spawnandcancelagainst a worker's local socket) as a follow-up PR.