Skip to content

v2: fixes from the merge-bar deep analysis - #334

Open
mtoslalibu wants to merge 4 commits into
nous-v2from
nous-v2-fixes
Open

mtoslalibu wants to merge 4 commits into
nous-v2from
nous-v2-fixes

Conversation

@mtoslalibu

Copy link
Copy Markdown
Collaborator

Fixes from the deep analysis of why Nous v2 lost the merge bar to Claude + a 4-line prompt (dbg / pd / search, transcripts compared turn by turn). Refs #330.

Machinery

  • Stop rule: claims_needed.conditions were an exact-string gate that never matched (search ran all 10 iterations, pd all 5). They are now advice: each later DESIGN message lists every unmet requirement. min_seeds / needs_ci still gate answered. Replayed on the recorded runs, they would stop at iteration 2-3 for search and 2 for pd.
  • Validator: every schema error at once (up to 10), each with its JSON path. Before this, only the first error was reported, so one file took 3-4 Stop blocks. The effect contract is clearer (value is one number, note holds prose); these two fields caused 63 of 76 real schema errors.
  • Noise: zero noise (deterministic) no longer blocks CONFIRMED/REFUTED or demands extra seeds.
  • Session: setting_sources=["project","local"], so the user's ~/.claude settings, plugins and global CLAUDE.md no longer load (about 10k tokens per turn; verified live). This matches the isolated config of the recorded L2 baseline.
  • Cost: the CLI's own total_cost_usd, cumulative per session id, is the authority, and it is kept across resume in state.json. The old meter read 10-28% low.

Method (general, not task-specific)

  • Principles carry an optional scope, the settings they were tested in, which is shown in later DESIGN messages. Rule 8: a principle holds only in its scope; test it before relying on it elsewhere.
  • Rule 9: validate against the task's own space or criterion, never a reduced version you built.

Reviewed: a bug review and a simplicity/overfit review. An auto-fix that wrote files back was dropped, as was a required scope. A double count on resume was fixed.

688 tests pass; no live LLM calls in tests. The next step is re-running the merge bar plus a task the fixes were not tuned on.

🤖 Generated with Claude Code

toslali-ibm and others added 4 commits October 26, 2026 12:09
…se, principle scope, clean session, cost)

- stop rule: conditions are advice, not a gate; DESIGN lists unmet claims_needed
- validator: fix formatting in place, report every schema error with its path
- noise rule: zero noise (deterministic) never blocks a status
- principles: add/revise carry the scope they were tested in; scope shown in DESIGN
- method: scope rule and check-against-the-original rule
- session: setting_sources project+local (no user ~/.claude leak)
- cost: CLI total_cost_usd per session is the authority

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…al scope, zero noise skips effect check

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…count)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ath from campaign.yaml)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
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.

2 participants