[codex] follow up glm thinking recovery by terisuke · Pull Request #63 · Cor-Incorporated/opencode

terisuke · 2026-04-05T07:48:48Z

What changed

compact on RepetitionError instead of stopping the turn
ignore stale assistant messages without finish when building the next turn
return the last finished assistant when an unfinished stale assistant trails it
add regression coverage for repetition compaction and polluted unfinished-session recovery

Why

PR #56 improves GLM-5.1 thinking corruption handling during the active stream, but two gaps remained:

repetition loops still stopped the turn instead of forcing compaction
already-saved broken unfinished assistant output was still re-injected into later turns, poisoning resumed sessions even after the live-stream fix

Impact

resumed GLM sessions stop inheriting malformed unfinished thinking output
direct follow-up turns can recover from a previously polluted session without carrying stale assistant text back into model context
the follow-up stays stacked cleanly on top of feat(opencode): add memory system and GLM thinking corruption fix #56

Root cause

The saved corrupted assistant output in the affected session had no finish, but SessionPrompt.run still kept that message in MessageV2.filterCompactedEffect(sessionID) and passed it back into the next model turn. In parallel, SessionProcessor.halt only auto-compacted ContextOverflowError, not the already-detected RepetitionError path.

Validation

bun test test/session/processor-effect.test.ts
bun test test/session/prompt-effect.test.ts
bun run typecheck
rebuilt the live binary with OPENCODE_CHANNEL='live/guardrails' OPENCODE_VERSION='0.0.0-live/guardrails-202604041233' bun run build --single
verified the real affected session with opencode run -s ses_2a395c5e6ffeuPcAR7KBys7W91 "Check GitHub directly and answer in one line only: is https://github.com/Cor-Incorporated/Grift/pull/412 OPEN or MERGED?", which executed gh pr view 412 --repo Cor-Incorporated/Grift --json state --jq .state and returned MERGED

Notes

this PR is intended as a follow-up stacked on feat(opencode): add memory system and GLM thinking corruption fix #56
local live deployment had been pointing at an older worktree, so the feat(opencode): add memory system and GLM thinking corruption fix #56 behavior was not actually present in the live wrapper before this verification

ToolRegistry.layer now requires Question.Service and Todo.Service as direct dependencies after upstream refactor. Build those layers explicitly before providing them to the registry, matching the pattern in prompt-effect.test.ts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Skip parallel implementation enforcement when repository has no commits yet. This prevents the deadlock where mutations are blocked until team tool is called, but team/yardadd requires HEAD to create worktrees. - chat.message hook: detect HEAD-less repo and suspend parallel policy - yardadd(): fail early with clear error if HEAD is missing Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Empty the evals Set so all configured providers are available for standard work. Guard both eval-related gate checks with evals.size > 0 to prevent provider-eval from being blocked when no eval providers are defined. Update tests accordingly. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…de (#25) Verify that parallel implementation policy is suspended when repository has no commits, preventing the deadlock described in issue #25. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Add Notification service with zero dependencies (darwin/linux/win32) - macOS notifications attributed to Terminal.app (fixes click target) - Add terminal focus detection to avoid notifications when focused - Add 'notifications' config option to tui.json (default: true) - Listen to session.idle for completion notifications - Update session.error to also show system notifications - Fix documentation plugin example (osascript attribution) - Beautify TUI toast: full borders + variant icons (✓✗ℹ⚠)

- Add .catch() to session.idle notification to prevent crashes - Use proper escapeForOsascript function for macOS - Add more terminals to detection list (vscode, code) - Restore escapeXml for Windows branch

feat(opencode): add memory system and GLM thinking corruption fix

[codex] fix team redirection false positive for read-only bash

github-actions · 2026-04-05T07:48:58Z

Hey! Your PR title [codex] follow up glm thinking recovery doesn't follow conventional commit format.

Please update it to start with one of:

feat: or feat(scope): new feature
fix: or fix(scope): bug fix
docs: or docs(scope): documentation changes
chore: or chore(scope): maintenance tasks
refactor: or refactor(scope): code refactoring
test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.

github-actions · 2026-04-05T07:49:00Z

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

GitHub-hosted runners are not allocated for pull_request events on forked repos. Switch to pull_request_target which runs in the base repo context. Explicitly checkout PR head SHA to test the correct code. Closes#62 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…target fix(ci): use pull_request_target for fork runner compatibility

Cherry-pick of PR #63 (d33e51a) which was accidentally merged into the feature branch instead of dev. - RepetitionError triggers compaction instead of stopping the turn - Filter stale unfinished assistant messages from resumed sessions - 3 regression tests added Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Copilot

Pull request overview

This PR improves session recovery and compaction behavior (especially around repetition loops and stale unfinished assistant messages), and also introduces/updates notification and guardrails-related behavior plus CI workflow adjustments.

Changes:

Compact on repetition loops and prevent unfinished/stale assistant output from being re-injected into subsequent turns.
Add regression tests for prompt/session recovery and repetition compaction paths.
Add cross-platform system notifications in the TUI and extend guardrails profile behavior (including a new team orchestration plugin), plus CI script/workflow tweaks.

Reviewed changes

Copilot reviewed 37 out of 37 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
packages/opencode/src/session/processor.ts	Treat `RepetitionError` like context overflow to trigger compaction.
packages/opencode/src/session/prompt.ts	Sanitize message history to ignore unfinished assistants; adjust loop exit/interrupt behavior.
packages/opencode/test/session/processor-effect.test.ts	Regression test asserting repetition loops return `"compact"`.
packages/opencode/test/session/prompt-effect.test.ts	Regression tests for ignoring unfinished assistant pollution and returning last finished assistant.
packages/opencode/src/notification/index.ts	New notification utility (macOS/Linux/Windows).
packages/opencode/test/notification.test.ts	Tests for terminal-app normalization and Windows toast XML escaping.
packages/opencode/src/config/tui-schema.ts	Adds `notifications` setting to `tui.json` schema.
packages/opencode/src/cli/cmd/tui/app.tsx	Trigger OS notifications on `session.idle` / `session.error` when not focused.
packages/opencode/src/cli/cmd/tui/ui/toast.tsx	Toast UI polish (icons, border, stricter error typing).
packages/opencode/src/cli/cmd/tui/feature-plugins/home/tips-view.tsx	Updates tip to reference built-in notifications in `tui.json`.
packages/web/src/content/docs/plugins.mdx	Docs update: AppleScript line + mention built-in TUI notifications.
packages/web/src/content/docs/ar/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/bs/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/da/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/de/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/es/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/fr/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/it/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/ja/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/ko/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/nb/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/pl/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/pt-br/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/ru/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/th/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/tr/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/zh-cn/plugins.mdx	Docs update: AppleScript notification snippet.
packages/web/src/content/docs/zh-tw/plugins.mdx	Docs update: AppleScript notification snippet.
packages/opencode/test/scenario/harness.ts	Adjusts layers provided to `ToolRegistry` in scenario harness.
packages/guardrails/profile/plugins/guardrail.ts	Removes default eval-only provider set; gates eval-only logic behind non-empty set.
packages/guardrails/profile/AGENTS.md	Updates provider admission guidance to reflect new policy behavior.
packages/guardrails/profile/plugins/team.ts	New team/background orchestration tools + enforcement hooks for large requests.
packages/opencode/test/scenario/guardrails.test.ts	Updates provider-lane expectations; adds tests for team plugin redirect/headless behavior.
packages/opencode/package.json	Ensures artifact directory exists before junit output in CI.
packages/app/package.json	Ensures artifact directory exists before junit output in CI.
.github/workflows/typecheck.yml	Switches to `pull_request_target` and checks out PR head SHA.
.github/workflows/test.yml	Switches to `pull_request_target`, checks out PR head SHA, tweaks artifact upload naming/options.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-05T08:16:25Z

 on:
 push:
 branches:
 - dev
-pull_request:
+pull_request_target:
 workflow_dispatch:

 concurrency:
 # Keep every run on dev so cancelled checks do not pollute the default branch
 # commit history. PRs and other branches still share a group and cancel stale runs.
 group: ${{ case(github.ref == 'refs/heads/dev', format('{0}-{1}', github.workflow, github.run_id), format('{0}-{1}', github.workflow, github.event.pull_request.number || github.ref)) }}
 cancel-in-progress: true

 permissions:
 contents: read
 checks: write

 jobs:
 unit:
 name: unit (linux)
 runs-on: blacksmith-4vcpu-ubuntu-2404
 steps:
 - name: Checkout repository
 uses: actions/checkout@v4
 with:
+ ref: ${{ github.event.pull_request.head.sha || github.sha }}
 token: ${{ secrets.GITHUB_TOKEN }}



Using pull_request_target while checking out the PR head SHA runs untrusted PR code with the base-repo GITHUB_TOKEN (here granted checks: write). This is a known security footgun because PR code can exfiltrate tokens/secrets or write to the repo via API; prefer pull_request for tests/typecheck, or keep pull_request_target but only run trusted code (checkout base ref) / gate forked PRs and strip permissions (and consider persist-credentials: false).

Copilot · 2026-04-05T08:16:25Z

 uses: actions/upload-artifact@v4
 with:
- name: unit-linux-${{ github.run_attempt }}
+ name: unit-${{ matrix.settings.name }}-${{ github.run_attempt }}


name: unit-${{ matrix.settings.name }}-... references matrix.* but this job does not define a matrix strategy. This will fail workflow evaluation at runtime/compile time; use a static name (e.g. unit-linux-...) or introduce an actual matrix.
Suggested change
name: unit-${{ matrix.settings.name }}-${{ github.run_attempt }}
name: unit-linux-${{ github.run_attempt }}

Copilot · 2026-04-05T08:16:26Z

@@ -13,6 +13,8 @@ jobs:
 steps:
 - name: Checkout repository
 uses: actions/checkout@v4
+ with:
+ ref: ${{ github.event.pull_request.head.sha || github.sha }}



Switching to pull_request_target (and checking out the PR head SHA) runs untrusted PR code with the base-repo token context. For a typecheck workflow this is usually unnecessary risk; prefer pull_request or ensure the workflow never executes PR-provided code with elevated permissions/secrets.

Copilot · 2026-04-05T08:16:26Z

+ .default(true)
+ .optional()


z.boolean().default(true).optional() won’t apply the default because .optional() wraps the schema after the default (parsing undefined will stay undefined). If you want a default of true, use .optional().default(true) (or drop .optional() entirely).
Suggested change
.default(true)
.optional()
.optional()
.default(true)

Copilot · 2026-04-05T08:16:26Z

+ sdk.event.on("session.idle", async (evt) => {
+ if (tuiConfig.notifications === false) return
+ const focused = await Notification.terminalIsFocused().catch(() => false)
+ if (focused) return
+
+ const sessionID = evt.properties.sessionID
+ Notification.show("opencode", `${sessionID} completed`).catch(() => {})
+ })


On non-macOS platforms Notification.terminalIsFocused() always returns false, so notifications will always fire even when the TUI is frontmost. If the intent is “only notify when focus is lost”, consider gating notifications by platform (or adding focus detection for Linux/Windows) to avoid noisy always-on alerts.

Copilot · 2026-04-05T08:16:26Z

+export namespace Notification {
+ export async function terminalIsFocused(): Promise<boolean> {
+ if (platform() !== "darwin") return false
+
+ const result = await Process.text(
+ [
+ "osascript",
+ "-e",
+ 'tell application "System Events" to get name of first application process whose frontmost is true',
+ ],
+ { nothrow: true },
+ )
+ return terminal(result.text.trim())
+ }
+
+ export async function show(title: string, message: string): Promise<void> {
+ const os = platform()
+
+ if (os === "darwin") {
+ const escaped = escapeForOsascript(message)
+ const titleEscaped = escapeForOsascript(title)
+ await Process.run(
+ [
+ "osascript",
+ "-e",
+ `tell application "Terminal" to display notification "${escaped}" with title "${titleEscaped}"`,
+ ],
+ { nothrow: true },
+ )
+ return
+ }
+


This change set introduces new cross-platform notification infrastructure, but the PR title/description are focused on GLM repetition/thinking recovery. Please update the PR description to cover the notifications feature (and any UX/OS support expectations) or split it into a separate PR to keep the follow-up scoped and reviewable.

Copilot · 2026-04-05T08:16:27Z

+const kids = new Set<string>()
+const need = new Map<string, Need>()
+const live = new Map<string, Run>()
+


The module-level kids/need/live collections grow without any eviction/cleanup. In a long-running process this can become an unbounded memory leak across sessions; consider pruning entries (e.g., on session completion/abort events, by TTL, or by max-size LRU) and/or persisting minimal state to disk instead of keeping everything in memory forever.
Suggested change
constkids=newSet<string>()
constneed=newMap<string,Need>()
constlive=newMap<string,Run>()
constcacheTTL=1000*60*60
constcacheMax=1000
classPrunableSet<T>extendsSet<T>{
privatereadonlyseen=newMap<T,number>()
privateprune(){
constnow=Date.now()
for(const[value,at]ofthis.seen){
if(now-at>cacheTTL){
this.seen.delete(value)
super.delete(value)
}
}
while(this.size>cacheMax){
constoldest=this.values().next()
if(oldest.done)break
this.seen.delete(oldest.value)
super.delete(oldest.value)
}
}
overrideadd(value: T){
this.prune()
this.seen.set(value,Date.now())
constresult=super.add(value)
this.prune()
returnresult
}
overridehas(value: T){
this.prune()
returnsuper.has(value)
}
overridedelete(value: T){
this.seen.delete(value)
returnsuper.delete(value)
}
overrideclear(){
this.seen.clear()
returnsuper.clear()
}
}
classPrunableMap<K,V>extendsMap<K,V>{
privatereadonlyseen=newMap<K,number>()
privateprune(){
constnow=Date.now()
for(const[key,at]ofthis.seen){
if(now-at>cacheTTL){
this.seen.delete(key)
super.delete(key)
}
}
while(this.size>cacheMax){
constoldest=this.keys().next()
if(oldest.done)break
this.seen.delete(oldest.value)
super.delete(oldest.value)
}
}
overrideset(key: K,value: V){
this.prune()
if(super.has(key)){
super.delete(key)
}
this.seen.set(key,Date.now())
constresult=super.set(key,value)
this.prune()
returnresult
}
overrideget(key: K){
this.prune()
returnsuper.get(key)
}
overridehas(key: K){
this.prune()
returnsuper.has(key)
}
overridedelete(key: K){
this.seen.delete(key)
returnsuper.delete(key)
}
overrideclear(){
this.seen.clear()
returnsuper.clear()
}
}
constkids=newPrunableSet<string>()
constneed=newPrunableMap<string,Need>()
constlive=newPrunableMap<string,Run>()

terisukeand others added 12 commits April 5, 2026 16:20

fix(ci): create JUnit output dirs before tests (anomalyco#20959)
552e30c

fix(tui): improve notification error handling and escaping
3d09150
- Add .catch() to session.idle notification to prevent crashes - Use proper escapeForOsascript function for macOS - Add more terminals to detection list (vscode, code) - Restore escapeXml for Windows branch

fix: use correct session.idle event type (sessionID not info.id)
38607a2

fix(tui): harden notification delivery paths
667cd1c

fix(team): allow fd-only bash redirection
e28ebc0

Merge pull request#56
bf1c1b9
feat(opencode): add memory system and GLM thinking corruption fix

Merge pull request#59
482833f
[codex] fix team redirection false positive for read-only bash

github-actionsBot added the needs:title label Apr 5, 2026

github-actionsBot added the needs:compliance label Apr 5, 2026

terisukeand others added 3 commits April 5, 2026 16:55

Merge pull request#64from Cor-Incorporated/fix/ci-use-pull-request-…
b00d3d2
…target fix(ci): use pull_request_target for fork runner compatibility

follow up glm thinking recovery
d33e51a

terisuke force-pushed the codex/fix-glm-thinking-followup branch from 7ad609f to d33e51aCompare April 5, 2026 08:11

terisuke marked this pull request as ready for review April 5, 2026 08:11

CopilotAI review requested due to automatic review settings April 5, 2026 08:11

terisuke merged commit 51e4372 into feat/glm-thinking-fix-and-memory-systemApr 5, 2026
2 of 4 checks passed

Copilotstarted reviewing on behalf of terisuke April 5, 2026 08:11 View session

terisuke mentioned this pull request Apr 5, 2026
fix(opencode): cherry-pick GLM thinking recovery follow-up #65
Merged

CopilotAI reviewed Apr 5, 2026
View reviewed changes

Navigation Menu

Uh oh!

[codex] follow up glm thinking recovery - #63

[codex] follow up glm thinking recovery#63
terisuke merged 15 commits into
feat/glm-thinking-fix-and-memory-systemfrom
codex/fix-glm-thinking-followup

terisuke commented Apr 5, 2026

Uh oh!

github-actionsBot commented Apr 5, 2026

Uh oh!

github-actionsBot commented Apr 5, 2026

Uh oh!

Uh oh!

CopilotAI left a comment

Uh oh!

CopilotAIApr 5, 2026

Uh oh!

CopilotAIApr 5, 2026

Uh oh!

CopilotAIApr 5, 2026

Uh oh!

CopilotAIApr 5, 2026

Uh oh!

CopilotAIApr 5, 2026

Uh oh!

CopilotAIApr 5, 2026

Uh oh!

CopilotAIApr 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	name: unit-${{ matrix.settings.name }}-${{ github.run_attempt }}
	name: unit-linux-${{ github.run_attempt }}

-constkids=newSet<string>()
-constneed=newMap<string,Need>()
-constlive=newMap<string,Run>()
+constcacheTTL=1000*60*60
+constcacheMax=1000
+classPrunableSet<T>extendsSet<T>{
+privatereadonlyseen=newMap<T,number>()
+privateprune(){
+constnow=Date.now()
+for(const[value,at]ofthis.seen){
+if(now-at>cacheTTL){
+this.seen.delete(value)
+super.delete(value)
+}
+}
+while(this.size>cacheMax){
+constoldest=this.values().next()
+if(oldest.done)break
+this.seen.delete(oldest.value)
+super.delete(oldest.value)
+}
+}
+overrideadd(value: T){
+this.prune()
+this.seen.set(value,Date.now())
+constresult=super.add(value)
+this.prune()
+returnresult
+}
+overridehas(value: T){
+this.prune()
+returnsuper.has(value)
+}
+overridedelete(value: T){
+this.seen.delete(value)
+returnsuper.delete(value)
+}
+overrideclear(){
+this.seen.clear()
+returnsuper.clear()
+}
+}
+classPrunableMap<K,V>extendsMap<K,V>{
+privatereadonlyseen=newMap<K,number>()
+privateprune(){
+constnow=Date.now()
+for(const[key,at]ofthis.seen){
+if(now-at>cacheTTL){
+this.seen.delete(key)
+super.delete(key)
+}
+}
+while(this.size>cacheMax){
+constoldest=this.keys().next()
+if(oldest.done)break
+this.seen.delete(oldest.value)
+super.delete(oldest.value)
+}
+}
+overrideset(key: K,value: V){
+this.prune()
+if(super.has(key)){
+super.delete(key)
+}
+this.seen.set(key,Date.now())
+constresult=super.set(key,value)
+this.prune()
+returnresult
+}
+overrideget(key: K){
+this.prune()
+returnsuper.get(key)
+}
+overridehas(key: K){
+this.prune()
+returnsuper.has(key)
+}
+overridedelete(key: K){
+this.seen.delete(key)
+returnsuper.delete(key)
+}
+overrideclear(){
+this.seen.clear()
+returnsuper.clear()
+}
+}
+constkids=newPrunableSet<string>()
+constneed=newPrunableMap<string,Need>()
+constlive=newPrunableMap<string,Run>()

Uh oh!

Conversation

terisuke commented Apr 5, 2026

What changed

Why

Impact

Root cause

Validation

Notes

Uh oh!

github-actionsBot commented Apr 5, 2026

Uh oh!

github-actionsBot commented Apr 5, 2026

Uh oh!

Uh oh!

CopilotAI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

CopilotAIApr 5, 2026

Choose a reason for hiding this comment

Uh oh!

CopilotAIApr 5, 2026

Choose a reason for hiding this comment

Uh oh!

CopilotAIApr 5, 2026

Choose a reason for hiding this comment

Uh oh!

CopilotAIApr 5, 2026

Choose a reason for hiding this comment

Uh oh!

CopilotAIApr 5, 2026

Choose a reason for hiding this comment

Uh oh!

CopilotAIApr 5, 2026

Choose a reason for hiding this comment

Uh oh!

CopilotAIApr 5, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants