Repository navigation
fix(release): backport four KAS, Claude-backend and sandbox fixes to 0.12.0 line - #31328
Conversation
…thout a session key (#30643) - Bridge a remote server only when KAS itself would accept its url (https, or http to a loopback host, normalised as KAS's URL parser does). A plain http url to another host is named in the session notice instead of being bridged with its headers in cleartext. - A plain-http (loopback) bridge ignores environment proxies, so its headers never reach an http_proxy in cleartext. - In MCP registry mode, the notice names each server Crew withholds, and Crew's own servers, which KAS drops when the Kiro profile is registry-governed. - A session batch with no session key yet (background or pooled runtime) names its launch snapshots with a random key, so stdio env and streamable-HTTP headers still reach the server. - Move two config reads out of kas_agents.py for its line cap. Fixes #30630 Co-authored-by: Joe Guo <12429733+iamwhatever@users.noreply.github.com> (cherry picked from commit 50ebe42)
… out (#31064) #30643 stopped bridging an MCP server's headers to a plain http:// url on another host. That setup connected through Crew's bridge before, so it is restored by default, and the refusal moves behind agent.kas_refuse_plain_http_mcp_headers (off). The bridge bypasses environment proxies only for a loopback host, so a remote http server keeps its proxy route. The loopback host reading moves to the standard-library bridge, without a regex, so both sides share it. (cherry picked from commit f0d6dc0)
…op (#30982) When Claude Code fails to start, the claude adapter answers with a -32603 frame saying "Claude Code process exited with code 1". The chat runner matched "process exited" in that text and treated it as the agent process dying, so every turn re-queued behind "Connection lost - retrying..." until "Session stuck", and the adapter's own error never reached the user. A process that answered a request is alive. acp_error_is_pipe_death now reads only the client's own death wording on an AcpError with no JSON-RPC code, and the dashboard turn uses it in both places it decides pipe death. The answered error now ends on the error card with its text. (cherry picked from commit 3126556)
Dev Fleet runs worktree code outside the agent sandbox (a pod). The gate read apps/dev-fleet/installed.json, under the agent-writable, renamable apps/dev-fleet directory, so a sandboxed shell could forge enabled:true and start a pod outside the jail. Authorize pods on a gateway-written marker under a top-level sealed, pre-created, strict-no-follow host-exec-grants/ directory the sandbox cannot rename or symlink around. The owner-authenticated enable_app writes the marker for the app it enables (fresh and already-enabled paths); it is never derived from the enablement record. Every disable path revokes it and fails closed if it survives. The helpers are generic (no app id in core); only the pod gate reads dev-fleet's. Co-authored-by: Joe Guo <12429733+iamwhatever@users.noreply.github.com> (cherry picked from commit 2406ca9)
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: these are four clean cherry-picks of fixes already merged on Notes
[DESIGN-REVIEWED] f528cf3 Description read[DESCRIPTION-READ] c004c0376489dfea518259158d8c7731e396ffcebd3e36d47d91f6466c0c7b3f 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 Nothing blocks: this backports four fixes already merged on What this change shipsInventory (7 items) — 7 justifiedIntent: ship four KAS, Claude-backend and sandbox fixes already merged on
[FIRST-PRINCIPLES-REVIEWED] f528cf3 Description read[DESCRIPTION-READ] a97abe37a5c5f3501afa31a855fc7c55cc0f9ebd3b4d0d919bbf8a6b3526ee1c 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: |
Sensitive change review
|
Problem / Motivation
Goal: Bring four KAS, Claude-backend and sandbox fixes that merged on
mainafterv0.12.0ontorelease/0.12.0, one commit per fix, so the 0.12.1 patch ships them and each one can be reverted alone.Why it matters
envandheadersdo not reach it when the session has no key yet.http://MCP server on another host withheadersstops connecting through Crew's bridge. With it, it connects as before, and refusing it is an opt-in setting.Not a goal
main; nothing is changed only here.CHANGELOG.mdchange. Cuttingv0.12.1-rc.Nand the bare tag follow docs/build/release.md and stay with the maintainers.What changed (motivation → approach → change)
Each main squash commit was cherry-picked with
-xontorelease/0.12.0@982f542467, in this order, one commit each. All four applied clean. #30643 and #31064 travel together: #31064 is #30643's follow-up and restores plain-httpheader relay that #30643 alone would refuse.50ebe4221b6eb5ea1351f0d6dc0fc889c0c1a4b1312655623c1f5a637e182406ca9013f528cf34c0Two
Co-authored-bytrailers named an internal address; they carry the same person's public noreply address here. No tree changed.#30816 changes the Dev Fleet pod API's gate only: a pod starts when a gateway-written grant under a sealed directory exists, instead of when an agent-writable file says the app is enabled. #31064 changes
kas_http_bridge.py's outbound opener and host check. Neither adds a listener, bind, route or auth path.Backwards compatibility
Compatible: each change is the one already merged on
main, and each PR's own compatibility note applies unchanged. #30816's own note applies: a Dev Fleet enabled before this version is re-enabled once so its grant is recorded, and the pod tool's refusal says so.Removes nothing: no pick hides, deletes or renames anything 0.12.0 accepts beyond what its own main PR already reviewed.
Tests
Proof: not provable -- other: a backport of main commits whose proofs ran on their own PRs
release/*, so the related set was run here: the test files this diff adds or changes, the test files that import a source module these picks touch, and for the widely importedchat_runner,config.sections,llm_helpersandapps.managermodules the tests that name the changed functions, plus the sandbox governance-mask and protected-name tests. 80 files, sequential batches of up to 20, one pytest process each,-n 0. 5128 passed. 2 tests fail, and both fail the same way on cleanrelease/0.12.0@982f542467, so neither comes from this diff:test_acp_harness_contract.py::test_a_missing_self_served_binary_names_its_install_command[copilot]andtest_sandbox_governance_mask_model_loaders.py::TestASandboxedChildCannotMintAnAliasOfASealedLoaderFile::test_the_backend_denies_both_destinations.scripts/check_black_formatting.pypasses, and mypy passes on the 21 changed source files.Manual verification
N/A - every pick was verified on its own main PR. The 0.12.1 RC soak is the end-to-end check.
Related Issues
Backport of #30643, #31064, #30982 and #30816 to the 0.12 line for 0.12.1. Source issues: #30630, #30444, #30811.
Sibling sweep
Siblings: none -- each source PR carried its own sibling sweep; this PR only carries those fixes unchanged.
Pattern harvest
Not generalizable: this PR only carries four fixes already on
main. Each source PR has its 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.