Replace CLI orchestration loops with agent workflow prompts

[alanzabihi]

Summary

• Replace the multi-step state machines in contribute.rs (443 lines) and lead.rs (367 lines) with agent workflow prompts that call CLI subcommands as shell tools
• Rewrite experiment.md from 9 lines to 43 lines with iteration support, harness-only metric rules, and explicit leave-changes-in-place instructions
• Add contribute-workflow.md (75 lines) and lead-workflow.md (63 lines) as the agent outer loops
• Add spawn_workflow_agent to agent.rs for spawning workflow agents
• Net: 272 insertions, 627 deletions (-355 lines)

Architecture change

The CLI no longer contains orchestration loops. polyresearch contribute and polyresearch lead now:

1. Setup (clone, init, preflight)
2. Spawn a workflow agent with the appropriate prompt
3. The agent calls CLI subcommands (polyresearch claim, polyresearch submit, etc.) as tools

All 15+ CLI subcommands stay exactly as-is. The agent composes them into workflows adaptively.

Lead/contribute independence

Lead and contribute are now fully independent agent sessions:

• Lead stays on the repo root (default branch). Does sync, decide, policy-check, generate.
• Contributor creates worktrees for each thesis. Does claim, experiment, attempt, submit, release.
• Each runs git operations sequentially. No concurrent async loops sharing git state.
• Strict role separation enforced in prompts: contributor never runs lead commands, lead never runs contributor commands.

This eliminates the race condition class (bugs 11, 20, 28, 29, 35, 37, 38) by design.

Motivation

See agent-vs-cli-balance.md for the full analysis. In short: agents fail at precision, CLIs fail at adaptation. The CLI keeps the precision-critical protocol primitives. The agent handles the adaptive multi-step workflows where the combinatorial state space caused 42 bugs in the rigid Rust loops.

Test plan

• 175 unit tests pass (cargo test --lib)
• Clean build, zero warnings
• Manual test: polyresearch contribute on a real project
• Manual test: polyresearch lead on a real project
• Manual test: both running simultaneously from the same directory

---

Note

Medium Risk
Large behavioral refactor of the CLI’s core lead/contribute execution path plus new git/GitHub automation (sync retries, worktree management, staging/commit). Failures could block or mis-sequence protocol actions, so end-to-end manual runs on real repos are important.

Overview
Shifts polyresearch lead and polyresearch contribute from Rust orchestration loops to prompt-driven workflow agents. Adds new contribute-workflow.md and lead-workflow.md prompts (plus a rewritten experiment.md) and introduces agent::spawn_workflow_agent plus prompt helpers to dynamically inject --once, sleep, max-parallel, and capacity guidance.

Introduces new protocol primitives to support the agent-driven flow. Adds polyresearch resume to recreate/verify a thesis worktree, sync .polyresearch-node.toml, and rewrite .polyresearch/thesis.md; adds polyresearch commit to stage only editable-surface changes and block protected paths before committing; claim/batch-claim now gate on a new duties::claim_gate and seed worktrees with thesis context.

Hardens repo hygiene and coordination. Bootstrap now tracks untracked files created by the setup agent, normalizes PROGRAM/PREPARE line endings, ensures .gitignore ignores .polyresearch-node.toml, and can force-add agent-created helper files. Sync is rewritten to pull/rebase safely and retry pushes on non-fast-forward races, with logic to discard “sync-only” local commits when needed. Thesis generation rejects duplicate titles via normalization, prune can remove worktrees for resolved/rejected theses, submit refuses PRs with no diff vs default branch, and GitHub handling adds enable_issues plus improved rate-limit retry classification/backoff.

Adds a --once cycle guard. New cycle_guard enforces “exactly one thesis cycle” runs by preventing additional claims after a release/submit marks the guard done.

^{Reviewed by Cursor Bugbot for commit 305dc9e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Lead and contribute loops exit after agent success

High Severity

In continuous mode (once=false), both the lead and contribute outer loops break when the workflow agent exits successfully. The agent prompt says "LOOP FOREVER," but agents will inevitably exit with code 0 due to context window limits. When that happens, the process silently terminates instead of restarting the agent for the next iteration. In contribute.rs, Ok(()) => break exits immediately; in lead.rs, the break at the end of the Ok arm has the same effect once post-checks pass. The Err path correctly restarts, but the Ok path does not, making continuous operation impossible once the agent's context window is exhausted.

Additional Locations (1)

• cli/src/commands/contribute.rs#L86-L87

 

^{Reviewed by Cursor Bugbot for commit 75e7ab7. Configure here.}

[cursor]

Commit command missing results.tsv in protected files

Medium Severity

The ALWAYS_PROTECTED list in the new commit.rs omits results.tsv, a critical protocol file that tracks the experiment ledger. If an agent accidentally modifies results.tsv in a worktree, polyresearch commit would include the change. Whether this is caught depends entirely on the PROGRAM.md "cannot_modify" globs being set up correctly — a fragile assumption given that bootstrap is agent-driven.

 

^{Reviewed by Cursor Bugbot for commit 6fd5553. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 7 total unresolved issues (including 6 from previous reviews).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 305dc9e. Configure here.}

[cursor]

Claim missing node config sync to worktree

High Severity

claim creates the thesis worktree and writes thesis context, but unlike resume (which calls sync_node_config_to_worktree), it never copies .polyresearch-node.toml into the new worktree. When the workflow agent later CDs into the worktree and runs polyresearch commit, polyresearch attempt, or polyresearch submit, those commands call read_node_id(&ctx.repo_root) which looks for the config file in the worktree directory — and fails because it doesn't exist.

Additional Locations (1)

• cli/src/commands/resume.rs#L52-L54

 

^{Reviewed by Cursor Bugbot for commit 305dc9e. Configure here.}