Skip to content

fix(nightly-v2): refresh the patch stack onto the current OV2 branch - #100

Open
saphid wants to merge 2 commits into
mainfrom
fix/nightly-v2-refresh-stack
Open

saphid wants to merge 2 commits into
mainfrom
fix/nightly-v2-refresh-stack

Conversation

@saphid

@saphid saphid commented Sep 30, 2026 •

Copy link
Copy Markdown
Owner

Problem

Fork Nightly Orchestrator v2 has not built since 24 Sep. One pinned commit no longer exists on GitHub, and upstream has since rewritten and advanced t3code/codex-turn-mapping, so several patches stopped applying.

Fix

The stack is replayed on the current branch tip (23309198ba, 1 Oct), and the manifest pins the new commits. They live on automation/nightly-v2-20261001-stack, a branch that exists only to keep them reachable. (automation/nightly-v2-20260930-stack holds the previous replay onto 8297ccf85d.)

Old patch Result
10 of the 16 patches carried over, same content
adapt custom updater to V2 profiles re-resolved against upstream's updater test harness
keep OV2 desktop and server releases on the fork re-resolved against the same harness
resume orphaned provider starts after recovery re-resolved around upstream's new early return; test uses upstream's namespace imports (pingdotgg#14387)
keep startup recovery running past undecodable thread projections test imports converted to upstream's namespace style (pingdotgg#14387)
recover from slow CLI health probes partly dropped. Upstream replaced the probe with a cached OpenCode 1.x/2.x generation probe (pingdotgg#14239), which needs a real version to route the instance. The server fallback that skipped the version on a CLI timeout no longer fits that design. Only the AI Enablers launcher is kept, renamed chore(ai-enablers): keep the local OpenCode launcher in source control; it caches opencode --version itself.
preserve racy timestamps in review indexes dropped: upstream has its own fix
await ACP prompt settlement before soft-steer interrupts dropped: merged upstream (pingdotgg#13201)

Two patches are new:

  • test(desktop): adapt fork fixtures to current OV2 test helpers: two fork tests no longer compiled.
  • fix(server): fail a recovered run when its final start attempt fails: the recovery patch could leave a run running forever if its last retry could not open a session. The run now fails, its provider thread returns to idle, and its original start time is kept.

This supersedes #94.

Verification (on 23309198ba + stack)

  • The real assembler (downstream-nightly.mjs resolve then apply) builds v0.0.44-nightly-v2.20261001.1790840661129402. The assembled tree matches the tested tree except for the generated build metadata and package versions.
  • vp check: 0 errors. vp run typecheck: passes.
  • Desktop: full suite, 1,400 tests pass.
  • Server: 62 tests in the six touched or related files pass, including OpenCode provider, version probe and recovery.
  • Web (133), mobile (7), contracts and shared (16), build scripts (86) and launcher (unittest) tests in the files the fork touches pass.

Not verified: no build has run in CI yet, because the workflow is disabled until #93 merges. Re-enable it with gh workflow enable 352233980 -R saphid/t3code.

Review

The previous replay had an independent Codex gpt-6-sol review (three passes); its findings are fixed. This restack has not had an independent review: the T3 orchestrator was not available in the session that made it.

Claude Opus 5.5 (1M context) in Claude Code.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AHKKXVog6PVQ1HEvdCKgxJ

The v2 stack stopped applying: one pin pointed at a commit that no longer exists, and five other patches conflicted with the upstream branch after it was rewritten.

Replay the stack on t3code/codex-turn-mapping at 8297ccf and pin the new commits, which live on automation/nightly-v2-20260930-stack. Two patches are dropped because upstream now has them (racy review index timestamps, ACP prompt settlement). Two are added: fixture updates for current test helpers, and a fix so a recovered run fails cleanly when its last start attempt fails.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@github-actions github-actions Bot added vouch:trusted PR author is trusted by repo permissions or the VOUCHED list. size:M labels Sep 30, 2026
Upstream fast-forwarded t3code/codex-turn-mapping by 36 commits after the
previous refresh. Two patches needed changes:

- "recover from slow CLI health probes" conflicted with upstream's new
  OpenCode 1.x/2.x generation probe (pingdotgg#14239). Its server fallback is dropped
  and only the AI Enablers launcher is kept.
- The startup-recovery and orphaned-start patches now use the namespace
  imports upstream adopted in pingdotgg#14387.

The commits live on automation/nightly-v2-20261001-stack.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AHKKXVog6PVQ1HEvdCKgxJ
@github-actions

github-actions Bot commented Oct 2, 2026

Copy link
Copy Markdown

Thread transfer impact

⚠️ The latest CI run did not produce a thread transfer result for e02be64.

This comment will update automatically after the next completed run.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

size:M vouch:trusted PR author is trusted by repo permissions or the VOUCHED list.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant