Conversation
aviggiano
added a commit
that referenced
this pull request
Sep 28, 2026
…d a recovered producer The resume hydration patches restored durable `skipped` rows as terminal. A skip is a verdict on prerequisites, and a reset can overturn it: after `resume --retry-failed` (timetravel --no-deps) re-ran a failed agent, its verifier and every descendant stayed skipped, so the recovered output was never verified or consumed (#1141). Restore only `finished` rows and let the workflow re-derive skips. That alone (the resume half of #1154) regresses a plain resume: the first render of a resumed session runs before hydration, so every skip predicate sees an empty state map and a still-failed agent's verifier and descendants get dispatched. The scheduler therefore re-renders once hydration has run, before its first decision. It also re-renders when a decision pass re-enters after skipping nodes. Upstream recurses there with the previous render's predicates, so the direct consumer of a failed agent, and each level below it, ran its preparation into a failure instead of being skipped: the "retried per descendant" cascade #1141 also describes. Both re-renders are one patch at the top of the scheduler's decide(). Resetting the skipped descendants from `--retry-failed` instead is not possible with Smithers' timetravel: it needs an attempt row for its target and selects dependents by attempt start time, and a skipped node has no attempts. The new integration test drives the pinned Smithers 0.35.0 engine with the scheduler and hydration patches applied to a private copy of the Smithers packages: a failed agent's skip cascade executes nothing, a plain resume executes nothing, and a --retry-failed style reset reopens the verifier and descendants without re-running finished work. Both tests fail on main. The string test that pinned the old "restore finished and skipped" rule is deleted. Co-Authored-By: Claude Opus 5.5 <[email protected]>
aviggiano
added a commit
that referenced
this pull request
Sep 29, 2026
…d a recovered producer The resume hydration patches restored durable `skipped` rows as terminal. A skip is a verdict on prerequisites, and a reset can overturn it: after `resume --retry-failed` (timetravel --no-deps) re-ran a failed agent, its verifier and every descendant stayed skipped, so the recovered output was never verified or consumed (#1141). Restore only `finished` rows and let the workflow re-derive skips. That alone (the resume half of #1154) regresses a plain resume: the first render of a resumed session runs before hydration, so every skip predicate sees an empty state map and a still-failed agent's verifier and descendants get dispatched. The scheduler therefore re-renders once hydration has run, before its first decision. It also re-renders when a decision pass re-enters after skipping nodes. Upstream recurses there with the previous render's predicates, so the direct consumer of a failed agent, and each level below it, ran its preparation into a failure instead of being skipped: the "retried per descendant" cascade #1141 also describes. Both re-renders are one patch at the top of the scheduler's decide(). Resetting the skipped descendants from `--retry-failed` instead is not possible with Smithers' timetravel: it needs an attempt row for its target and selects dependents by attempt start time, and a skipped node has no attempts. The new integration test drives the pinned Smithers 0.35.0 engine with the scheduler and hydration patches applied to a private copy of the Smithers packages: a failed agent's skip cascade executes nothing, a plain resume executes nothing, and a --retry-failed style reset reopens the verifier and descendants without re-running finished work. Both tests fail on main. The string test that pinned the old "restore finished and skipped" rule is deleted. Co-Authored-By: Claude Opus 5.5 <[email protected]>
aviggiano
added a commit
that referenced
this pull request
Sep 29, 2026
…d a recovered producer (#1169) * fix(runtime): resume --retry-failed reopens descendants skipped behind a recovered producer The resume hydration patches restored durable `skipped` rows as terminal. A skip is a verdict on prerequisites, and a reset can overturn it: after `resume --retry-failed` (timetravel --no-deps) re-ran a failed agent, its verifier and every descendant stayed skipped, so the recovered output was never verified or consumed (#1141). Restore only `finished` rows and let the workflow re-derive skips. That alone (the resume half of #1154) regresses a plain resume: the first render of a resumed session runs before hydration, so every skip predicate sees an empty state map and a still-failed agent's verifier and descendants get dispatched. The scheduler therefore re-renders once hydration has run, before its first decision. It also re-renders when a decision pass re-enters after skipping nodes. Upstream recurses there with the previous render's predicates, so the direct consumer of a failed agent, and each level below it, ran its preparation into a failure instead of being skipped: the "retried per descendant" cascade #1141 also describes. Both re-renders are one patch at the top of the scheduler's decide(). Resetting the skipped descendants from `--retry-failed` instead is not possible with Smithers' timetravel: it needs an attempt row for its target and selects dependents by attempt start time, and a skipped node has no attempts. The new integration test drives the pinned Smithers 0.35.0 engine with the scheduler and hydration patches applied to a private copy of the Smithers packages: a failed agent's skip cascade executes nothing, a plain resume executes nothing, and a --retry-failed style reset reopens the verifier and descendants without re-running finished work. Both tests fail on main. The string test that pinned the old "restore finished and skipped" rule is deleted. Co-Authored-By: Claude Opus 5.5 <[email protected]> * fix(runtime): re-render after every skip, not only after a pass that only skipped The skip re-render fired on `depth > 0`, that is only when a decision pass had changed node states without dispatching anything. When a skip shared its pass with dispatched work, decide() returned Execute, and the next decision made without a render (after a retryable failure, or at a retry deadline) scheduled the skip's dependents from predicates rendered before the skip. A failed agent's consumer then still ran its preparation into a failure: on a fresh run, and on a plain resume with the producer still failed. Both #1169 reviews reproduced this on the pinned engine. Mark the predicates stale where the skip happens instead. Upstream's skipIf branch now sets the same flag resume hydration sets, and the top of decide() re-renders whenever it is set; the `depth > 0` test is gone. The cost is at most one redundant render per pass that both skipped and dispatched, when a completion's own re-render already saw the skip. The flag is renamed `skipPredicatesStale`, the new anchor is registered as `skip_marks_predicates_stale`, and `skip_predicate_rerender` no longer claims to retire with `terminal_state_restore`, because it now serves both flag setters. A third integration test adds a task that becomes runnable in the pass that skips the failed agent's verifier and fails once retryably. It fails on main and on the previous patch (`prepare:consumer` ends `failed`) and passes with this one. Co-Authored-By: Claude Opus 5.5 <[email protected]>
Collaborator
Author
|
Superseded. Two independent reviews (both reproduced with the real pinned @smthrs 0.35.0 scheduler) found the #1141 part correct but incomplete — restoring only |
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
1 participant
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.
Recovered runs now reevaluate dependency-blocked descendants, keep immutable verified artifact generations addressable across replacement attempts, and fence pause/cancel publication by the current runtime owner. This gives continuation a safer path without weakening artifact verification.
What changed
Validation
pnpm --filter @ultrafuzz/artifacts testpnpm -r --workspace-concurrency=1 typecheckMitigates #1141
Mitigates #1142
Mitigates #1153