Repository navigation
Fix the GLM run's failures; weigh cost, noise and predictions in self-edits - #60
Merged
Merged
Conversation
Oversight woke the supervisor on levels, so after one real pass it spent the rest of its 12-pass budget in 24 minutes on the same unmet-gate count and was absent for the two hours that needed it. It now wakes on what changed since the last pass that spent a call. GLM sends array arguments as JSON strings: plan took one entry as the path ["src/a.clj"] and withheld done for it until the branch ran out of turns, and split saw no parts. Such strings are decoded, and list syntax left in a plan entry is refused as not a path. A board owner could close its dispatched task with task close, which skipped review, the critic and acceptance. That close is refused and points at done. Also: the contradiction check judges a denial per sentence; an invalid JSON escape is named with its column instead of "No matching clause: 38"; the repetition scan covers replies that end in a parse error and counts a short phrase repeated far past the floor; the verify pass reads the judge's reply without its reasoning.
…e-call formats A verdict now carries tokens per turn on both sides and applies RRSI's cost rule: a gain that added more tokens than it earned is costly, a flat change that cut them is cheaper. The deadband comes from each side's own spread instead of a fixed 0.15, and a result under the best a kept change reached is below-best. An experiment can name the version it tests and the signal it expects to move; the verdict settles the prediction and the lever memory keeps both. manifest gets a versions action and the drift block shows why the latest version was made. A side call's prompt that stops asking for what its parser reads is refused (gates.edn :reply-formats), and a judge reply with no verdict line is marked unread instead of passing silently. Gates may declare :measure; every check of one is kept in gate_checks with its value, fired or not, so a threshold can be fitted from both sides of its cut.
…en promotions The repl manifest could not run as a loop and its enforcement lives in phases.edn, so it and its cells are gone. A project edit that has been in force for failed runs and no shipped one is offered to the supervisor for removal, like a dead gate. An open experiment is shown with what the reviews said while it was in force, not only its fitness number. The arena records where each run's database is kept, pairs a passed and a failed run of one task and asks what the passing one did differently (ARENA_CONTRAST), and flags a recurring edit that names its subject or drops a guard its template had, before anyone promotes it.
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.
The endless-flight run on GLM-5.3 surfaced seven bugs, all fixed here. The worst: oversight woke the supervisor on levels rather than changes, spent its whole pass budget in 24 minutes, and left the last two hours unsupervised. plan also took a JSON-encoded list as a literal path, which withheld done until the owner ran out of turns. A board owner could close its own task with task close and skip review.
The rest applies the RRSI and DSPy review to the self-edit loop:
The unrunnable repl manifest is removed.
Full suite green locally (2860 tests).