v0.3.3 — Parallel Execution Hardening

Parallel Execution Hardening

Field-tested fixes from a 14-story parallel execution run (ql/image-mode feature). Addresses all critical issues discovered during real-world parallel agent orchestration.

Fixed

• File-conflict-aware DAG scheduling — filter_file_conflicts() prevents spawning parallel agents that share files. Wired into both dispatch sites in quantum-loop.sh.
• Worktree nesting prevention — _resolve_repo_root() resolves nested paths to top-level repo, preventing .ql-wt/ inside .ql-wt/ path explosion.
• Windows long-path fallback — Paths > 200 chars fall back to repo-namespaced temp directory. Emergency /tmp last resort.
• Editable install race — PYTHONPATH injection replaces pip install -e . in parallel worktrees (src-layout + flat-layout).
• Worktree cleanup retry — 3 attempts with 2s delay for Windows file locks; rm -rf only as fallback.
• Cross-wave conflict detection — filter_file_conflicts seeds with in_progress stories' files.

Added

• Inline review gate (Step 3B.4) — Spec compliance + code quality after each parallel merge.
• Full-feature code review (Step 4B) — Holistic review of entire branch diff for cross-story consistency, architecture coherence, and security.
• 14 new tests — 79 total (35 DAG + 15 worktree + 29 spawn), all passing.

Upgrade

/plugin marketplace update quantum-loop

Or pull latest and restart Claude Code.

Full changelog: CHANGELOG.md

v0.3.2 — Enforce Worktree Isolation

Fixed

• Mandatory worktree isolation — isolation: "worktree" is now documented as MANDATORY for parallel execution, with specific failure modes listed (bash contention, file conflicts, quantum.json races)
• Correct tool naming — orchestrator now references "Agent tool" (not "Task tool") with exact parameter names
• Atomic quantum.json updates — new Step 3B.1 batches all in_progress status writes into a single atomic update before spawning agents
• Monitor loop — changed from polling to waiting for Claude Code completion notifications
• State management discipline — only the orchestrator writes quantum.json; Edit tool banned (use Python/jq); multi-story updates batched into one write
• Implementer parallel mode — implementer agents in worktrees no longer edit quantum.json (stale copy); report via output message instead
• Anti-rationalization guards — 2 new entries blocking "skip worktree" and "worktrees won't work on this OS" excuses

Full changelog: https://github.com/andyzengmath/quantum-loop/blob/master/CHANGELOG.md#032---2026-03-10

v0.3.1 — Post-Mortem Fixes

What's New

Closes all 7 issues from the Math Research Agent post-mortem — the first real-world 34-story parallel execution. Delivered in 3 layers: schema, agent protocols, and runtime enforcement.

New Schema Fields

• contracts — cross-story agreements for shared values (secret keys, env vars, types). Implementers enforce exact values; propose-and-wait on disagreements.
• wiring_verification — grep-based mechanical check on tasks that create modules. Spec-reviewer verifies: missing string = fail, present = pass.
• consumedBy — declares cross-story consumption. Implementers import existing components instead of inlining duplicates.
• coverageThreshold — configurable test coverage gate. Quality-reviewer runs coverage tool and fails stories below threshold.
• staleThresholdMinutes + startedAt — stale story detection with configurable timeout (default 20min, CLI-overridable).

Agent Protocol Improvements

• Coding standards enforcement — quality-reviewer reads CLAUDE.md, .claude/rules/, and codebasePatterns; violations are CRITICAL severity
• Coverage gate — quality-reviewer detects coverage tool (c8/nyc/pytest-cov/go test/JaCoCo) and enforces threshold
• Contract reading — implementer reads contracts before implementing; propose-and-wait if disagreeing
• Lifecycle awareness — ql-brainstorm question #7 (LIFECYCLE) and ql-spec Pre-Save lifecycle checklist
• testFirst mandate — ql-plan defaults testFirst: true for all tasks; exempt tasks require justification

Runtime Enforcement

• Stale story detection in orchestrator, quantum-loop.sh, quantum-loop.ps1, and templates
• Final verification sweep — full test suite before COMPLETE; failure blocks completion
• Execution observations — auto-generated post-mortem doc after every run; optional GitHub issue filing
• CLAUDE.md defensive check — warns and skips stories in_progress by other agents

Bug Fixes

• Atomic in_progress + startedAt write prevents crash window
• startedAt cleared on all exit paths (BLOCKED, timeout, merge conflict, unrecognized signal)
• Null-guard failureLog in observations jq
• startedAt cleared in crash recovery

Tests

• 4 new shell test files (14 tests): stale detection, startedAt lifecycle, final sweep, observations

Full Changelog: v0.3.0...v0.3.1

v0.3.0 - Cross-Story Integration Layer

Informed by real-world field data: 14-story implementation where 100% of bugs were cross-story integration issues, 0% caught by per-story review gates.

Added

• Cross-story integration review (ql-review Stage 3) — traces call chains across story boundaries using LSP (grep fallback). Runs after dependency chains complete and as final gate before COMPLETE.
• Final integration gate — orchestrator runs import smoke test, full test suite, and dead code scan before declaring COMPLETE
• File-touch conflict detection — ql-plan flags parallel stories modifying same file, adds reconciliation tasks, stores conflicts in quantum.json metadata (fileConflicts)
• Consumer verification pattern — wiring acceptance criteria belong on the consumer story, not the creator
• Edge case test requirements — boundary values, type variations, collision scenarios, scale tests required for all testFirst tasks
• Edge case reference doc (references/edge-cases.md) — Python, JS, Go, Rust testing gotchas. Implementer reads it at the start of every testFirst task.
• Import chain verification — ql-verify requires integration evidence for multi-story features
• Cursor marketplace manifest (.cursor-plugin/plugin.json)

Changed

• Orchestrator Step 4 split into Final Integration Gate + Completion
• Implementer always reads references/edge-cases.md for testFirst tasks
• All versions bumped to 0.3.0 (plugin.json, marketplace.json, cursor plugin.json)

v0.2.1 - Gap Fixes + Marketplace Discoverability

Bug fixes from systemic gap analysis + marketplace discoverability improvements.

Fixed (13 gaps from workflow review)

• DAG query output — lib/dag-query.sh now outputs valid JSON arrays instead of newline-delimited strings
• Spawn prompt — Clearer numbered workflow steps for worktree agents, --dangerously-skip-permissions for background TTY-less agents
• Implementer wiring check — New integration wiring section: agents must verify new code is imported and called from at least one place (prevents dead code from parallel execution)
• Implementer signals — Fixed completion flow: commit before signaling STORY_PASSED
• Orchestrator cross-ref — Fixed error handling step reference (3A.5 → 3A.7)
• Template status values — Normalized all completed → passed across templates/quantum-loop.sh
• REPO_ROOT — Fixed to use $(pwd) instead of $SCRIPT_DIR in parallel mode
• marketplace.json — Version bump to 0.2.0, agent count 3 → 4 (orchestrator)
• README — Fixed skill directory names to match ql-* convention
• .gitignore — Added .quantum-logs/

Improved

• Marketplace discoverability — All 6 skill descriptions now include "Part of the quantum-loop autonomous development pipeline (brainstorm → spec → plan → execute → review → verify)" so searching "quantum-loop" on SkillsMP surfaces all skills as a cohesive suite
• Filed SkillsMP indexing issue — Skills approved via CLI but not appearing on website; under investigation

Install

npx skills add andyzengmath/quantum-loop

v0.2.0 - Orchestrator Agent + Windows Support

Orchestrator agent for reliable execution + Windows overnight support.

Added

• Orchestrator agent (agents/orchestrator.md) — manages full execution lifecycle inside Claude Code with DAG query, sequential/parallel dispatch, two-stage review, retry logic
• Native PowerShell script (quantum-loop.ps1) — Windows overnight runs without bash/WSL
• SkillsMP compatibility — name field in all SKILL.md frontmatter
• ql-plan runner copy — copies quantum-loop.sh/ps1 into project after planning

Fixed

• Lost work in parallel mode — agents must commit before signaling; orchestrator adds safety commit before merge
• Merge failure on dirty tree — stash working tree before merge, pop after
• Stale worktree branches — delete existing branch before git worktree add -b

Changed

• Simplified skills/ql-execute/SKILL.md from ~300 lines to ~50 line dispatcher
• CLAUDE.md parallel mode: agents explicitly told to commit before signaling

Windows Support

Three options for Windows users:

1. /ql-execute — interactive, orchestrator agent with native worktree isolation (recommended)
2. quantum-loop.ps1 — autonomous overnight, native PowerShell sequential loop
3. WSL2 + quantum-loop.sh — autonomous overnight, full parallel support

v0.1.0 - Parallel Execution

Parallel execution via DAG-driven worktree agents.

Added

• 7 shell library modules (lib/) for DAG query, worktree lifecycle, agent spawning, monitoring, atomic JSON writes, crash recovery
• 7 test suites with 110 tests
• --parallel and --max-parallel flags for quantum-loop.sh
• /ql-execute parallel orchestration via Task subagents
• Crash recovery for orphaned worktrees
• CLAUDE.md parallel mode instructions

v0.0.1 - Initial Release

Initial release of Quantum-Loop.

• 6 skills: brainstorm, spec, plan, execute, verify, review
• 3 agents: implementer, spec-reviewer, quality-reviewer
• quantum-loop.sh sequential autonomous loop
• Dependency DAG execution
• Two-stage review gates (spec compliance + code quality)
• Iron Law verification
• Anti-rationalization guards