Skip to content

Add LLP 0081: bind worker results and retries to verifiable evidence - #12

Open
bgmcmullen wants to merge 1 commit into
mainfrom
llp-0081-evidence-workers
Open

bgmcmullen wants to merge 1 commit into
mainfrom
llp-0081-evidence-workers

Conversation

@bgmcmullen

@bgmcmullen bgmcmullen commented Oct 29, 2026 •

Copy link
Copy Markdown

Summary

Implements HypAware recommendation hyprec-b61864d6c44a4c8c, "Prepare workers once and verify their results," from the Neutral token efficiency report (team logs, 2026-09-01 to 2026-09-28).

The recommendation's deliverable is a design spec, so this PR adds LLP 0081 as a Draft and appends Extended-by: forward references to LLP 0012, 0020, 0021 and 0075. No runtime code changes.

What the recordings showed

  • 160 of 162 attributed EnterWorktree attempts failed (98.77%) across 55 parent sessions, Sep 4 to 28. Workers were dispatched into the repository root and refused by the harness, then retried.
  • On Sep 28, 24 of 90 flagged errors were a missing Python interpreter across 22 child agents; each one rediscovered the gap independently.
  • Retry escalation treated these infrastructure failures as model failures and consumed coding attempts.

What LLP 0081 specifies

  • The deterministic launcher creates and verifies a detached worktree at the exact revision before a model starts; workers receive it and never call EnterWorktree.
  • A bounded runtime capability manifest (interpreters, package manager, cwd, HEAD) is recorded once per launch and passed to workers.
  • A versioned worker-result schema bound to exact revisions; a testsPass boolean alone cannot satisfy a gate.
  • Attempt accounting by controller-observed failure class. Only implementation failures consume the 5/3/2 model tier ladder; infrastructure and observation failures use bounded controller retries.

Context

The Neutral fleet was stopped on 2026-10-12 in favor of HypForge, and the recommendation was marked dismissed on 2026-10-16 for that reason. This PR records the artifact the recommendation produced against the Neutral codebase it was measured on. Whether the design carries into HypForge is a separate decision that needs HypForge evidence.

Checks

  • npm test: 306 of 307 pass. The one failure, "runner kills a timed-out process group and exposes timeout instead of success", fails identically on main at ad05e5d and is unrelated to this docs-only change.
  • LLP 0081 does not collide with any number on main; 0080 and 0082 are on other open branches by the same author.

🤖 Generated with Claude Code

Draft spec minted from HypAware recommendation hyprec-b61864d6c44a4c8c
(usage-review/2026-09-01-to-2026-09-28). Across September, 160 of 162
EnterWorktree attempts in the fleet failed at the repository root, and
workers repeatedly rediscovered the container's missing python3. Worker
setup here is prompt-driven and the implement Workflow's retry ladder
counts every non-landed task as a model failure, so infrastructure
failures escalate the model without evidence.

The spec asks for a deterministic launcher that prepares and verifies a
detached worktree before the model starts, a measured runtime capability
manifest passed to workers, a durable launch manifest, typed
revision-bound worker and review results, and attempt budgets classified
by failure class and persisted across restarts. Only implementation
failures climb the 5/3/2 ladder. Review caps, human grants, ship-risk,
and merge authority are unchanged.

Numbered 0081 because the unmerged codex/owned-worker-cancellation
branch already holds 0077 and 0078. Adds Extended-by forward-refs to
LLPs 0012, 0020, 0021, and 0075.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@bgmcmullen
bgmcmullen requested review from philcunliffe and removed request for philcunliffe October 29, 2026 22:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant