[mtoslalibu]

Summary

This PR delivers Nous v2 from issue #330 into the nous-v2 branch. It does not change main. main keeps old Nous until a separate, later decision.

Nous v2 is Claude Code plus a light scientific method plus checks enforced by code:

• One session, four phases. One Claude Code session (Agent SDK, claude-agent-sdk==0.2.172) runs BASELINE → DESIGN → EXECUTE → REPORT. Nous switches the phases.
• In-process hooks record every Bash command to runs.jsonl, run the validators when a phase ends (Stop), and write a digest before compaction. Nous never blocks or judges a command.
• JSON artifacts with schemas. A failed check goes back to the agent to repair; nothing is thrown away.
• Locked predictions. They are hashed before EXECUTE.
• Honest claims. Every claim value must match its evidence file, and claims must beat measured noise.
• Light, task-aware method. A method prompt of about 5 KB plus a checklist per task type (debug, verify, optimize, policy, explain).
• Stopping, resume and effort. The run stops when the question is answered or when it stalls (2 quiet iterations). It has a failure breaker and nous resume, and it handles effort when output hits the cap.
• Code isolation. All code changes happen in a separate git worktree at <repo>/.nous-experiments/<run_id>/ on branch nous/<run_id>, so the user's checkout is never touched.
• Cost is measured and reported, but it drives no decision.

Old Nous is deleted: the planner/executor split, the dispatchers, the shell hooks, gate summaries, the critic and complexity tiers, and the five v1 CLI flags. orchestrator/ went from about 15,900 lines to about 3,400; the v2 runtime is about 2,700.

Closes nothing yet. Refs #330.

How it was built and checked

• 21 plan items were built by a Ralph-style /goal session, with a reviewer session alongside. Each item went test-first, then /pr-review-toolkit:review-pr, then commit and push. The plan and full log are in docs/plans/nous-v2/ (SPEC, plan.json, progress.txt, LEAN_AUDIT.md).
• A full-branch review found 6 blocking bugs (P1–P6). Each was fixed with a test that fails before the fix. Details are in progress.txt.
• 669 tests pass. There are no live LLM calls; the conftest.py guard is kept and now extended to ClaudeSDKClient.connect.
• 4 live smoke runs on the BLIS debugging task all pass on mechanics. Smoke 3 and 4 run on the final code: exit 0, repro recorded, claims matched 1.0. In smoke 4 the user's checkout stayed clean, and the commits landed on nous/<run_id> in the campaign worktree.

Merge bar (R13): NOT met after 3 improvement rounds

See docs/plans/nous-v2/EVAL_REPORT.md. These are 3 oracle tasks compared against the recorded Claude + 4-line prompt (L2). Frontier-CS p5 and p225 were deferred because there is no Docker judge on the build machine.

task	accuracy vs L2	cost vs L2	trail C1
debugging	pass (2/2)	~1.3×	1.0
formula check	fail (1/2 vs 2/3)	0.8–1.0×	1.0
config search	fail	~4×	1.0

Across every run: 0 lost rounds, 0 crashes, 0 pre-registration violations. The merge bar runs found and fixed 7 real Nous bugs. This PR delivers a working, tested base. Closing the remaining accuracy and cost gaps is follow-up work in PRs into nous-v2.

Known follow-ups

• s390x (IBM Z) deployment: update to Nous v2 CLI, campaign.yaml and artifact layout #331: the s390x (IBM Z) deployment still uses v1 flags and keys. It breaks on v2 and must be fixed before nous-v2 merges into main.
• Nous v2 follow-ups from the pre-merge branch review (non-blocking) #332: non-blocking gaps from the branch review:
  • idempotent resume bookkeeping;
  • STOP and halt clearing;
  • evidence namespacing and secrets on export;
  • runs.jsonl fields;
  • the cost meter undercount (about 11%);
  • and others.
• EVAL_REPORT next steps: two of them need a design decision before anyone builds them, because they touch core choices: blocking commands, and merging DESIGN with EXECUTE.

Test plan

• .venv/bin/python -m pytest -q: 669 passed
• Live smoke on the final code: smoke/3.json and smoke/4.json, exit 0
• Docs and example checks: tests/v2/test_docs_examples.py
• Grep: no v1 terms in README, docs, prompts or skills; no removed flags in the code

🤖 Generated with Claude Code