Repository navigation
Release v1.5.2 - #177
Merged
Merged
Release v1.5.2#177
Conversation
…#160) * docs(openspec): propose the 1.5.2 cleanup, QA and guidance follow-ups * docs(openspec): tighten the 1.5.2 change after review * docs(openspec): fix a support scenario and pin the archive steps * docs(openspec): apply the second review round and keep the system prompt section stable * docs(openspec): deliver request notes in a carrier valid for the latest message
… and validate receipts (#162) * fix(companion): clear dependency blocks against current prerequisites and validate receipts before refusing The moved-back guidance in the skill, tool description and inline guidance now unblocks against the plan's current prerequisites: a prerequisite the plan no longer lists no longer holds the task, and a still-listed skipped one counts as unfinished. Coordinator refusal guidance validates pending and reserved receipts as the per-plan reader does and reports incomplete evidence instead of a closed refusal. The retry baseline and point-lookup applicability are documented. * docs: describe how malformed delegation receipts affect coordinator guidance
…163) * fix(ui): report unconfirmed cleanup results apart from skipped plans A bulk Tidy or selection run now keeps the failure class of each request. A plan the service refused, left out before sending, or never sent after Stop is skipped; an attempted request with no valid answer is reported as unconfirmed and may have been applied. The result line says so in English, Russian and Simplified Chinese. Undo still restores validated receipts only and Try again keeps its behaviour. The cleanup fixture's service calls wait 30 seconds while its readiness probe stays short, and the selection driver asserts that every plan was archived before it uses Undo. * fix(ui): agree Russian unconfirmed wording and share the refusal check Pick the Russian pronoun and participle by the exact count of one, so 21, 31 and 101 read as plural. Export cleanupRefused from the refusal codes cleanupError already uses, and use it in the run engine. Reword the Select bullet about a withdrawn store and the fixture probe comment. * docs(openspec): mark the fixture check done; the failing drivers predate this change
* fix(ui): bind cleanup busy state to its request, keep capability from the latest health * fix(ui): drop stale health on announced account change and guard retired Try again
…sts (#165) * fix(companion): keep the system prompt section identical across requests The companion's system section carried the one-request hint, the sub-agent resume capsule (with its plan revision) and the delegation refusal note, so it changed between requests and defeated provider prompt caching. The section is now the fixed activation guidance only, byte-identical on every request where it is present. The hint, capsule and refusal note are added for one request after the content of its latest message, on a copy so the stored conversation never carries them: a text part on a user message, the output of the last tool result on a tool message (not the Code Mode wrapper result), and nothing on any other role, where the text waits for the next request with a valid carrier. Providers that drop plugin system sections now receive these notes too. Fixes #161 * docs(openspec): keep delivered context notes in place on later requests * docs(openspec): pin re-delivery, cleared line, keys and memory for context notes
…t it cannot compare (#166) * test(browser): make the large-library gate count by request and report what it cannot compare Bound each timed window by log offsets read in the evaluations that start and stop its timer, and take every count the gate compares from the request log inside them. Count run reads once per window as an informational row, keep health reads out of per-refresh traffic, record the attribution method and both inbox forms, compare the list-carried inbox form per list start, and report a budget without a comparable baseline as unverified with a non-zero exit. Document how a baseline is recorded for the current artifact format. * test(browser): pin the one-uncomparable-candidate rule and align the gate docs * test(browser): drop measurement data from the baseline docs and tighten the gate's reasons
…ts (#167) * fix(companion): keep delivered context notes in place on later requests A context note present on one request and absent from the next still edits the history: the provider writes its cache at the end of the latest message, where the note sits, and never reads that entry again. Standing notes (an assigned child's capsule, an outstanding refusal) went out on every request, so those sessions got no conversation-history cache, and extended thinking bound to the prefix was lost too. The companion now remembers each delivered note per session with its carrier (tool result: call id, position and a digest of its original output; user message: its id, else position and a digest of its content) and adds the same bytes in the same place on every later request that still holds the carrier, so each request's messages begin with the previous request's. Notes due together are one note; a capsule or refusal is delivered again only when its text changed or its carrier is gone; a refusal that stops applying gets one fixed cleared line after a complete read. The memory is bounded (128 sessions, 64 carriers and 64 KiB per session) and every path that removes the section removes the notes. Refs #161 * fix(companion): tighten context note memory after review - An incomplete empty refusal read still delivers the omission text but no longer opens the cleared line; only a real refusal does. - A note too large for its carrier leaves an armed capsule armed. - The cleared line is delivered again when a revert removes it but keeps the refusal it clears. - Notes stay on the message they were delivered on when user messages share an identifier. - A session has memory only once it holds a note, and the memory is recorded only after the notes were added to the request. - The array is restored again just before keys are computed, so overlapping requests on a reused event never hash their own copies. - Raise the per-session bound to 256 carriers and 128 KiB, so long sessions keep the cache of the start of the conversation. - Hash binary parts by their raw bytes and take the latest carrier once.
…rson (#168) * fix(ui): keep cleanup repaints and focus moves from disturbing the person A count that finishes by itself, a Tidy run's progress and the end of a run now repaint as background data, so they wait for a held control or an open picker instead of replacing it. A refresh no longer expires a result that has not been painted yet. A deferred focus move is decided when it is applied and is dropped once focus is anywhere other than the control it was queued from. * fix(ui): judge a deferred focus move from the control that had focus before the paint The recheck read the focused element after the paint had replaced the root, so a rebuilt control such as Undo looked like no focus and lost it to Tidy or Cancel. Emptiness and identity are now taken before the root is replaced. forgetPlans keeps its dev signature and no longer bumps a version; each caller picks the repaint lane. The harness replaces a focused control on every paint, and the rendered driver checks that focus on Undo survives the return of Tidy.
* fix(ui): keep Tidy and selection counts exact while plans change A Tidy or Select all count taken from a list read in more than one page now says some plans may have been missed, on both steps, without adding a read. Search freshness replies certify only the result page they were asked about and only when they covered every plan it shows, are retired by a cleanup in the frame, and report plans they show gone to the selection whichever page is shown. A page opened by Load more starts unchecked. A Select all departure carries the cleanup generation that took the plan out, so a plan restored elsewhere is counted again. * fix(ui): keep selection rows, counts and scans in agreement on departed plans A departed plan known at a newer cleanup generation is now treated the same by the row, the selection count, the snapshot plans and a scan. PlanSelection.forget takes only a map of plan to generation. A selection that throws while following plans a freshness reply showed gone no longer leaves the search half updated. Document the qualified zero line, clarify the multi-page flag, and run the two slow search scenarios on one smaller backend.
Run the port-bound backend group in its own lane beside the typechecks, the parallel backend group and the browser and automation tests, so the Linux check step finishes sooner. The serial group stays one --test-concurrency=1 invocation; a listen probe found no other lane binding Setup ports, and the built-service cleanup test joins the serial group when CP_BUILT_SERVICE=1. - Node compile cache under node_modules/.cache, emptied at each run start and turned off with a note if it cannot be created. - The Linux jobs skip the zip/unzip install when the tools are present. - The Windows job runs its two native files concurrently, each waited on as a whole process tree. - New bun run test:fast and test:all share one plan with CI. - The lane runner stops every process group it starts on every exit path, retires groups once empty, never signals an unsafe or reused id, handles SIGINT, SIGTERM and SIGHUP, and exits 128 plus the signal when interrupted. Stopping on Windows is best effort and was not run there. No tests were removed and the counts are unchanged.
Archive the 1.5.0, 1.5.1 and 1.5.2 OpenSpec changes in that order, bump the version to 1.5.2, and add the 1.5.2 release notes and changelog entry.
#175) On Windows the companion started PowerShell synchronously for every ownership and ACL check, on load, on each tool hook and on each status inspection, which blocked OpenCode's event loop for 28 to 43 s and left the model list unavailable. Windows ACL queries are removed from the companion, service and setup, and Windows state relies on the user profile's default permissions. No product code starts a child process any more. Symbolic links, junctions, hard links, canonical paths, markers and schema checks are kept, existing installs with private ACLs keep working, and POSIX checks are unchanged. A fixed 2.5 s time headroom before an intent and before exclusive creations replaces the spawn-count reservations. Docs and the windows-default-permissions OpenSpec change describe the new behaviour. Fixes #172
Add the Windows companion fix to the 1.5.2 release notes and changelog, credit the reporter of #172, and archive the windows-default-permissions OpenSpec change.
Code Review SummaryStatus: 2 Issues Found | Recommendation: Address before merge Overview
Issue Details (click to expand)CRITICAL
Files Reviewed (177 files)
Removing synchronous PowerShell calls addresses the reported stalls. Both safeguards can be restored without bringing those calls back. This was a static review; no tests or builds were run. Fix these issues in Kilo Cloud Reviewed by gpt-6.2-sol · Input: 45 · Output: 9.6K · Cached: 988.7K |
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
Promotes
devatd212ef5tomain. Merge with Create a merge commit, not squash, so the source ancestry stays. That push makes the release workflow run its checks and then publish thev1.5.2tag, the curated release Agent Progress v1.5.2: A stable system prompt, and steadier Tidy, and the extension and companion ZIPs automatically.This supersedes #174, which promoted
devbefore the Windows fix below and should be closed unmerged.Closes #161
Fixes #172
What's in 1.5.2
/api/provideranswering 503, sessions timing out; 1.5.1 Windows: the companion blocks the OpenCode event loop for 28–43 s on load (models.dev fetch times out, /api/provider 503, no model list, sessions time out) #172). The companion now starts no process for permission checks on any platform. Agent Progress no longer checks Windows folder permissions and relies on the profile's default per-user permissions, so a store or custom folder that other users can write to is now accepted; link, path, marker and schema checks remain, Linux and macOS checks are unchanged, and existing installs keep working (fix(windows): stop blocking OpenCode with PowerShell permission checks #175). The release notes and changelog carry the fix, and thewindows-default-permissionsOpenSpec change is archived (docs(release): add the Windows fix to 1.5.2 notes #176).Protocol and schema stay at 2, and the minimum host remains OpenChamber 2.7.1.
Verification
main13f4123thendevd212ef5; the tree is exactlydevplus the nine generated outputs; the promotion command's two builds matched; every version stamp reads 1.5.2.mainfrom13f4123: eligible, version 1.5.2, release name Agent Progress v1.5.2: A stable system prompt, and steadier Tidy and body matchingrelease-notes/1.5.2.md, including the Windows item and its reporter credit; the extension and companion ZIPs and checksum inventories are accepted.devplus the nine outputsproduction-runtime,plugin-failure-runtime,progress-activation-runtime,subagent-activity-runtime,subagent-upgrade-runtime)v1.5.1, 13 merged pull requests, no outside contributorsThe 20 skipped tests are 18 Windows-native tests that skip off Windows, the released-1.1.1 Windows migration journey (needs the full-Windows opt-in that PRs into
mainset in CI), and one opt-in built-service restart test. The Windows suite runs in this PR's CI.Checklist