Repository navigation
fix(kas): backport #35244 #35262 #35272 #35278 to 0.12 - #35389
Conversation
…#35244) On KAS, a workspace whose steering tree held one file that may be `inclusion: auto` kept the whole-tree `file://.kiro/steering/**/*.md` entry, so KAS injected every steering file on every turn beside its own native copy. Send just those files, each by its own path, in the entry's place. Keep the entry whenever the paths might load less than it did: a glob character in a name, another file entry before it or another steering entry, or more than KAS's resource budget. The user docs say the same. Closes #32153 Original author: Bolin Chen (#35244) (cherry picked from commit 7eecd4a)
On the KAS bundle route the engine runs the Node its bundle ships, and `kirocrew doctor` already reports a missing node as not needed there. `kirocrew setup` still printed "node not found on PATH - install Node.js", sending a new user to install a runtime the selected backend never uses. Setup's prerequisites notice now asks the route predicate doctor asks; on the bundle route it names the bundle's own Node instead of the install advice. Off the bundle route the warning is unchanged. Fixes #34944 Original author: Bolin Chen (#35262) (cherry picked from commit dbb3160)
…5272) A runtime start records its descendant processes once the agent has answered initialize. When that scan saw no descendant it slept 0.5 s and walked again, on the path the caller was waiting on. On Windows the walk returns nothing by design (the teardown walks the tree itself through its owned handles and taskkill /T), so every start slept and then found nothing: every chat session, the background runtime, every sub-agent. walk_observes_descendants() now states once, beside the walk, whether it can see a descendant on this platform. Where it cannot, the runtime scan, the failed-start scan and the one-shot client's scan read nothing and wait for nothing. Where it can, a start whose scan sees no descendant returns at once and its one retry runs as a task of its own after the same delay: an ordinary pass, serialized and bracketed by the root's identity like every other, which the kill cancels. The guarded pass itself no longer sleeps. Nothing a teardown or a sweep reaps is recorded later or less than before, and no kill path signals differently. Closes #35252 Original author: Bolin Chen (#35272) (cherry picked from commit 70203ee) Backport note: comments, docs and test imports resolved by hand. The release branch does not have main's /proc task-children reader (_read_task_children, _CHILDREN_INTERFACE), so: - runtime_process_tree._direct_children docstring: kept release's text and took main's new last line (Returns [] where the walk cannot see a descendant); main's Linux paragraph names a reader release lacks. - acp-client.md: left out main's "Child enumeration on Linux" bullet, which describes that same reader; the bullet's new sentence about walk_observes_descendants is covered by the hunk that applied cleanly in the descendant-scan section. - test_process_tree.py: added the two imports main's file already had (types.SimpleNamespace, runtime_process_tree as rpt). No code line was hand-edited.
) A new KAS session waits for Kiro Crew's own MCP servers before its first prompt. The engine connects four servers at a time, in the order a session declares them, and it declares the running agent's servers ahead of the session-level array that carries Kiro Crew's. So the first prompt waited for every server the agent declares: about 2 s with none, about 35 s with 60. Start the session in a server-less start agent when its own agent declares servers. The set_mode every KAS start already sends then switches to the session's own agent, which queues that agent's servers behind Kiro Crew's. The agent's servers keep the bounded wait they already had (5 s) and connect in the background after it. The session starts in its own agent, as before, when that agent declares no server, mounts the engine's sub-agent tool, or the batch cannot take one more entry. session/load is unchanged. Closes #35245 Original author: Bolin Chen (#35278) (cherry picked from commit 265b3ac) Backport note: comments, docs and one test import resolved by hand; no code line was hand-edited. release/0.12.0 has neither Kiro Cloud Sessions (acp/harness/kas_cloud.py) nor the connect-queue progress tracker (acp/mcp_connect_progress.py), so: - harness/base.py session_request docstring: main's text without its sentence about KasCloudSandboxHarness. - test_kas_mode_binding.py: imports START_AGENT_ID only; main's kas_cloud import is used by no test in this file. - providers.md: dropped the conflicting block, which is main's whole "Kiro Cloud Sessions" section; this PR only edited one line of it. - acp-client.md "KAS managed MCP readiness": kept this PR's sentences on queue order and the start agent; left out main's surrounding text on the four-slot queue and McpConnectProgress, which describe code release lacks.
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: this ports four fixes already merged on Notes
[DESIGN-REVIEWED] 1dabd86 Description read[DESCRIPTION-READ] bfd2d2fc07cde3c1a43f79b3842de6643e00889cb7e12273ca3df0dc592da52d 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 |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS This PR backports four fixes already merged on What this change shipsInventory (6 items) — 6 justifiedIntent: give 0.12 KAS users the four fixes
[FIRST-PRINCIPLES-REVIEWED] 1dabd86 Description read[DESCRIPTION-READ] 4b44d9ea40b611d715b6f1f7e849860bd3c2649bb26a2dabe443c304ff84d84a That sha256 names the PR description this verdict read. Recompute it with |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
Problem / Motivation
Goal: Bring four KAS/ACP fixes merged on
mainon 2026-11-11 (#35244, #35262, #35272, #35278) ontorelease/0.12.0for the next 0.12.x, one commit each.Each fixes something a 0.12 user on KAS can hit today: the whole steering tree sent where only matching files belong (#35244), setup asking for Node on the bundled KAS route (#35262), every runtime start sleeping on an empty process-tree scan, worst on Windows (#35272), and Kiro Crew's own MCP servers connecting last behind an agent's servers (#35278).
Why it matters
These hit every KAS chat, sub-agent and setup run on the 0.12 line. Merge with Rebase and merge so each pick keeps its own commit.
Not a goal
main; nothing is changed only here.may_skip_gate_for_session,session_profiles_govern) that the fix(kas): six KAS harness and mcp.json reach fixes #33510 backport deliberately left off this branch, so porting it means porting governance code. @bolichen97, that one is yours.CHANGELOG.mdchange.What changed (motivation → approach → change)
Each main squash commit was cherry-picked with
-xontorelease/0.12.0@8f9e308c72, in main merge order. The repository's leak guard requires the public identity, so every commit is authored as Ray Xu and names its original author on anOriginal author:line.7eecd4a11eb38b501199dbb3160d65077647e1a070203ee0de1b07e24eed265b3ac9441dabd86cf3Picks 1 and 2 have the same patch-id as their main commits.
Prerequisites. None picked. The conflicts in picks 3 and 4 come from main-only features (the
/proctask-children reader, Kiro Cloud Sessions, the MCP connect-queue tracker). Those are features, not fixes, and the hunks that touch them are comments, docs and imports only.Hand-resolved hunks (@bolichen97, please check these; no code line was hand-edited):
src/kiro_crew/acp/runtime_process_tree.py_direct_childrendocstring: kept the release text and took main's new last line ("Returns[]where the walk cannot see a descendant (walk_observes_descendants)"). Main's Linux paragraph names_read_task_children, which release does not have.docs/system-specs/modules/acp-client.md: left out main's "Child enumeration on Linux" bullet, which describes that same reader. The descendant-scan hunk that nameswalk_observes_descendants()applied cleanly.test/test_process_tree.py: addedfrom types import SimpleNamespaceandfrom kiro_crew.acp import runtime_process_tree as rpt, which main's file already had.src/kiro_crew/acp/harness/base.pysession_requestdocstring: main's text without its sentence onKasCloudSandboxHarness(noharness/kas_cloud.pyhere).test/test_kas_mode_binding.py: importsSTART_AGENT_IDonly. Main'skas_cloudimport is used by no test in this file.docs/system-specs/modules/providers.md: dropped the conflicting block, which is main's whole "Kiro Cloud Sessions" section. This PR only edited one line inside it.docs/system-specs/modules/acp-client.md"KAS managed MCP readiness": kept this PR's sentences on queue order and the server-less start agent. Left out main's surrounding text on the four-slot queue andMcpConnectProgress, which describe code release does not have.Backwards compatibility
Compatible: each fix keeps its main PR's compatibility notes; nothing here changes stored data or config beyond what those PRs shipped on
main.Removes nothing: no key, record or public function is removed; the only removals are doc and docstring text that describes main-only code.
Tests
Proof: not provable -- other: a backport of main commits whose proofs ran on their own PRs
release/*, so the related set ran here, one pytest process at a time with no xdist.test_kas_*,test_acp_runtime*,test_acp_client*,test_process_tree*,test_runtime_*,test_cli_setup*,test_session_handle*andtest_steering*file (180 files), in batches of 20: 6,983 passed, 25 skipped, 0 failed. No failure, so no clean-release comparison was needed.--target-version py310) pass on the changed Python files.Manual verification
N/A - each fix was verified on its own main PR. The 0.12.x RC soak is the end-to-end check.
Related Issues
Backport to the 0.12 line of #35244, #35262, #35272 (closes #35252 on main) and #35278.
Sibling sweep
Siblings: none -- each source PR carried its own sibling sweep; this PR only carries those fixes.
Pattern harvest
Not generalizable: this PR only carries fixes already on
main. The source PRs have their own harvest.Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.