Repository navigation
fix(skillgen): let GLM 5.3 Harness runs finish - #550
Merged
Merged
Conversation
The first Harness proof on GLM 5.3 failed on its first model call: OpenCode Go answered 400 MissingSessionID, because the gateway rebuilt headers from scratch and dropped x-opencode-session. The gateway now forwards the four x-opencode routing headers when each value is a plain identifier. None of them carries a credential.
The first full GLM 5.3 proof spent its 14 minute runner budget on 32 verification steps, about 26 seconds each, and was cancelled while it wrote SKILL.md. The sandbox limit rises to 45 minutes, and the runner asks the Agent to write SKILL.md by turn 20 so a deadline still leaves a candidate. The GitHub job deadline counted time in the queue. Jobs run one at a time, so a second tag would spend its deadline waiting behind a 45 minute run. The deadline now starts when a job leaves the queue.
harlan-zw
added
the
harlan-agent-review
harlan-github-agent
Bot
added
harlan-agent-review-required
The proof client gave up after 17 minutes, before the new 45 minute sandbox limit.
Proof run 3 ended its turn at step 6 with finish reason length: GLM 5.3 reasoning counts as output, and the gateway capped each call at 8,192 tokens. The cap rises to 32,768 for OpenCode Go. The runner also moves from skilld-harness 3.3.0 to 3.6.4, which returns failed output checks to the Agent. That run ended with no output directory, a check the Agent could have fixed in one more turn.
harlan-zw
changed the base branch from
fix/skillgen-opencode-session-header
to
main
October 7, 2026 04:40
…line # Conflicts: # workers/skill-harness/src/gateway.ts
harlan-zw
changed the title
This was referenced Oct 7, 2026
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.
🔗 Linked issue
Related to #529 and #545
❓ Type of change
📚 Description
Since #529, no Skillgen run on GLM 5.3 can finish. With #545's session headers in place, three proofs on which-nuxt 0.4.1 failed for three reasons:
SKILL.md. The sandbox limit rises to 45 minutes, and the runner asks the Agent to writeSKILL.mdby turn 20, so a deadline still leaves a candidate.length, before any file was written. The OpenCode Go cap rises to 32,768.skilld-harness3.3.0. The run ended on "Harness output directory is missing", a check that 3.6.4 returns to the Agent for up to two repair turns.The GitHub job deadline also counted time in the queue. Jobs run one at a time, so a second tag would spend its deadline waiting behind a 45 minute run. The deadline is now 60 minutes from when a job leaves the queue. The
/skillgenpage and both runbooks state the new limits.A branch deploy cannot prove this: a main deploy replaced the dispatched version 15 seconds later. The proof runs after this merges.