Littlepowers 1.3.1

Littlepowers 1.3.1 is a lightweight compatibility and recovery-safety patch.

Highlights:

• automatic Hook recovery is confined to an exact Git repository or worktree ledger root, preventing a non-Git manager-directory ledger from steering unrelated child-project tasks;
• Contract guidance excludes planned write targets and generated implementation evidence from immutable parent sources;
• all Codex starter prompts fit the current 128-character host limit;
• package metadata and documentation are aligned for Codex, Claude Code, Qoder, and OpenCode.

Verification:

• 201/201 local Python tests passed;
• all 11 skills and all four package surfaces validated;
• exact commit 260005f passed Linux, macOS, Windows, and Claude Code CI: https://github.com/clsaa/littlepowers/actions/runs/31415328154
• a fresh exact-tag checkout peeled to that commit and passed Codex, Claude Code, Qoder, and OpenCode smoke validation;
• authenticated Codex and Qoder CLI Hook-routing probes returned the exact workflow ID;
• the installed Codex 1.3.1 plugin repeated the exact-workflow probe with no Littlepowers starter-prompt warning;
• local Codex, Claude Code, and Qoder installations now each report enabled Littlepowers 1.3.1; the old Codex alpha was removed and Claude Superpowers remains disabled.

Boundary: Claude Code 2.1.226 strict validation passed, but the local Fable model probe was blocked before model execution by expired OAuth, so this release does not claim an authenticated Claude implementation run.

Littlepowers 1.3.0 — Review Lease and parallel workflows

Important

Littlepowers 1.3.0 writes schema-4 ledgers. A 1.2 runtime cannot read them.
Before rolling back, cancel any open Review Gate, pause or finish the workflow,
restore its matching pre-schema4 schema-3 archive, and only then install 1.2.
Do not hot-replace Littlepowers inside an active task.

Littlepowers 1.3.0 is the stable Review Lease release for Codex, Claude Code,
Qoder, and OpenCode. It keeps planning proportional and model-neutral while
making approved outcomes, review boundaries, and parallel worktree status more
durable and deterministic.

Highlights

• Review Lease: four explicit policies cover blocking review, a fixed
  implementation mandate, timed fallback, and unattended continuation. Gates
  bind the exact artifact path, bytes, Contract sources, workflow, and revision.
• Less unnecessary waiting: a fixed Lean/Compact implementation request can
  proceed without another approval stop. An explicitly windowed Claude Code
  gate may arm one bounded same-session callback; no retry, output capture,
  permission bypass, model override, or effort override is added.
• Parallel project iterations: an optional Project Workflow Index summarizes
  up to 16 explicitly registered worktrees. Every worktree retains its own
  ledger and single writer; hooks never scan or load the index.
• Outcome integrity: Outcome Lock continues to require complete acceptance
  coverage, approved scope deltas, baseline fidelity, and aggregate completion.
  Implementation is one continuous product scope; tasks, checkpoints, small
  commits, and rollback units remain safe engineering boundaries, not staged
  product slices.
• Lightweight by default: ordinary routing adds no extra model call, agent,
  repository scan, scheduler, daemon, broad test run, telemetry, or model/effort
  selection. Direct work remains direct; Lean remains brainstorm → plan.
• Cross-host packaging: the same 11 skills and protocol are validated for
  Codex, Claude Code, Qoder, and OpenCode.

Verification

• Release commit: c095354a01ff7f09cb539261a44fc1efc9f7d88b.
• Three independent expert reviews: Critical 0, Important 0, work-unit
  compliance pass, approved-outcome fidelity pass, code quality approve.
• Local aggregate: 198/198 Python tests passed; all 11 skill validators and all
  four host package surfaces passed.
• Exact-commit CI:
  Ubuntu/Python 3.9, macOS/Python 3.13, Windows/Python 3.13 (including the Hook
  launcher probe), and Claude Code strict validation all passed.
• Exact-tag CI:
  the same four jobs passed from annotated tag v1.3.0.
• Independent local-tag and GitHub-remote-tag clones both passed the 11 skill,
  Codex package, Claude strict, Qoder, OpenCode, identity, and empty-ledger
  doctor smoke checks.

The release does not claim authenticated GPT-5.6 Sol/Ultra, Fable 5, Opus
4.8/5, Claude dynamic-workflow orchestration, or a real Windows Claude runner
end-to-end certification. The protocol is model-neutral; the runner evidence is
deterministic fake-host control and argv coverage plus cross-platform CI.

See the README
for exact-tag installation, updates, rollback, troubleshooting, security, and
host-specific limitations. The detailed integrated evidence is in
evals/results/2026-08-01-v1.3.0-release.md.

Littlepowers 1.2.0-alpha.1 — Outcome Lock

Outcome Lock moves scope integrity from prompt-only guidance into a lightweight schema-3 state boundary.

Highlights:

• explicit parent-source binding and drift detection
• deterministic Outcome coverage, scope-delta, baseline/fidelity, and completion gates
• legacy schema-1/schema-2 reconciliation with recoverable pre-schema3 archives
• one continuous implementation stream; tasks and rollback units are not staged product slices
• shared behavior across Codex, Claude Code, Qoder, and OpenCode
• no additional model call, background scan, automatic test run, or model/effort override

Verification: 128 Python tests, 11 skill validators, isolated self-host migration, and all four host package surfaces passed. See evals/results/2026-07-26-v1.2-outcome-lock.md for exact evidence and limitations.

Littlepowers v1.1.0-alpha.1

Highlights

• Adds a lean route for bounded changes: brainstorm → plan → execute, without separate specification or design artifacts.
• Preserves approved parent requirements under one definition of done and blocks agent-created product, technical, platform, MVP, or phase slices.
• Requires an explicit Added / Changed / Deferred / Removed scope delta, or No scope delta.
• Distinguishes approved visual baselines from implementation-generated regression snapshots.
• Separates work-unit compliance, approved-outcome fidelity, and code-quality review verdicts.
• Binds recovery context to the canonical workspace root so unrelated ancestor ledgers do not drive nested projects.

Compatibility and performance

The update remains shared across Codex, Claude Code, Qoder, and OpenCode. It adds no state-schema migration, extra model call, reviewer, background scan, telemetry, or automatic test run.

Verification

• 75 Python tests passed on Ubuntu/Python 3.9, macOS/Python 3.13, and Windows/Python 3.13.
• Python compilation and the Windows Hook launcher probe passed.
• Claude Code strict plugin validation passed in GitHub Actions.
• Local release validation passed for all 11 skills and Codex, Claude Code, Qoder, and OpenCode plugin surfaces.

See the changelog and verification report for details.

Littlepowers 1.0.0

The first stable release of Littlepowers: a proportional planning, recovery, and engineering-discipline protocol for Codex, Claude Code, Qoder, and OpenCode.

Highlights

• Four supported hosts: first-class Qoder CLI/Qoder IDE and OpenCode support alongside Codex and Claude Code, sharing one skill, hook, and state implementation.
• Phase review gates: full-route artifacts (brainstorm → spec → design → plan, plus compact shape) are presented for explicit approval before the next phase; unattended end-to-end execution requires explicit authorization.
• Host plan-surface mirroring: tracked checklists render in the host interface through Codex update_plan and OpenCode's todo tool, while the Markdown plan remains the durable source of truth.
• Expert-review hardening: fixes from a seven-expert review across gate mechanics, hook root resolution, OpenCode spawn handling, and Codex post-replacement recovery — itemized in CHANGELOG.md §1.0.0.

Validation

71 unit tests pass on Linux, macOS, and Windows CI, plus compileall, node --check, qodercli plugins validate, and claude plugin validate --strict. See CHANGELOG.md for the full record.

Known limits: the Qoder IDE fires only a subset of hook events, and authenticated end-to-end runs on Qoder and OpenCode have not been recorded yet (see docs/model-compatibility.md).

Littlepowers v0.4.0-alpha.1

Adds systematic debugging, proportional evidence-gated verification, and read-only two-verdict change review for Codex and Claude Code. CI passed 50 tests on Linux, macOS, and Windows; all 11 skills and both plugin packages validated. See CHANGELOG.md and evals/results/2026-07-17-v0.4-alpha.1.md for evidence and model-compatibility limits.

Littlepowers v0.3.0-alpha.1

Littlepowers v0.3.0-alpha.1 is an expert-reviewed prerelease for Codex and Claude Code.

Highlights

• Proportional direct, tracked-direct, compact, and full planning routes.
• Prompt-boundary recovery through SessionStart, UserPromptSubmit, and SubagentStart.
• Revisioned schema 2 ledger with compare-and-swap, advisory locking, bounded context, safe artifact reads, and replacement archives.
• Coordinator-only ledger writes for Ultra and Claude dynamic workflows.
• English and Simplified Chinese onboarding plus security, compatibility, evaluation, and contributor documentation.

Validation

• 43 tests pass on Ubuntu, macOS, and Windows; the Windows Git Bash launcher probe passes.
• All eight skills and both plugin formats pass local validation.
• Claude Code 2.1.212 strict validation passes in CI; isolated installation identified eight skills and three hooks.
• GPT-5.6 Sol xhigh passed routing scenarios 1 through 9 in one evaluation campaign.
• GPT-5.6 Sol max found no remaining P0/P1 issue after adversarial review.
• Codex Ultra preserved root-only writes with two workers and rejected a stale update.

Limits

• Fable 5 and Opus 4.8 are compatible by design, but an authenticated end-to-end Claude model run is not yet recorded.
• The model evaluations do not yet meet the three-run reliability threshold.
• Do not enable Littlepowers and Superpowers as simultaneous default workflow routers.