v0.40.3

Fixed

• scope-executor goals.md test — Adapted test to match synced generic goals.md (removed feature-row criteria assertion that no longer applies).

Docs

• appetite constraint contrast — Added paragraph explaining departure from original Shape Up calendar-based model: stelow caps preparation depth, not wall-clock time.

v0.40.2

What's Changed

Changed

• cli-tools distribution — Removed 276 duplicated references/cli-tools/ files from git tracking. Single source of truth is now stelow-product-orchestrator. Files are generated at build/install time. Updated install.sh, Pi extension, and CI pipeline accordingly.

Added

• .npmignore — Overrides .gitignore to ensure gitignored cli-tools are included in npm tarball.

Fixed

• sync-cli-tools.sh — Now excludes orchestrator-only files from sub-skills, handles skill-specific extras correctly, and ignores the orchestrator itself as a target.

v0.40.1

Fixed

• diff-gate stdout parsing — plannotator review does not support --json. Added explicit stdout pattern-matching strategy and human fallback for ambiguous results.
• Selected interface artifact — Chosen interface proposal now saved to .stelow/{date}/{dir}/interfaces/selected-interface.md as permanent artifact. Execution stage reads it for UI direction.

v0.40.0

Added

• plan-gate stage — Tech plan gate (Plannotator on spec-tech.md). Between Planning and Execution. Only in Tech Review/Code Diff modes.
• diff-gate stage — Code diff review gate (Plannotator review). Between Verification and Audit. Only in Code Diff mode.
• New review mode: Product Spec + Interface + Tech Review + Code Diff — adds diff-gate for maximum oversight.

Changed

• Review mode rename for clarity:
  • Only Product Spec → Product Spec Gate
  • Product Spec + Interface Choice → Product Spec + Interface Gates
  • All Above + Scopes In/Out → Product Spec + Interface + Scopes
  • All Above + Tech Review → Product Spec + Interface + Tech Review
• PHASE_NAMES expanded from 15 to 17. All integrations, plugins, tests, and docs updated.

v0.39.3

Added

• Evidence & Limitations: added coordination overhead research (CooperBench, clawRxiv, Co-Coder), research vs code parallelism rationale, file-overlap guard documentation

v0.39.2

Fixed

• Pulse scripts now ship with the extension — previously lived only in
  .stelow/pulse/ (gitignored), so anyone installing stelow via GitHub/npm
  got a broken Pulse. The 5 pulse files now live in extensions/stelow/pulse/
  (tracked, versioned). The extension's ensurePulseScripts() auto-copies
  them to the user's .stelow/pulse/ on first /sw-pulse invocation.
• PULSE_MODEL no longer hardcoded to haiku — defaults to empty,
  uses harness-configured model. --model flag only passed when set.
• 12 broken TOC anchors in README — double-dash to single-dash for
  single-codepoint emoji headings. toSlug test corrected.

Added

• scripts/setup-pulse.sh — standalone Pulse setup, no pi required.
  Useful for CI/CD, cron setup, or pre-staging before pi is installed.
• setup.sh Step 11/11 (Pulse) — optional step, delegates to
  setup-pulse.sh. Renumbered 10 previous steps /10 → /11.
• Pulse + HITL differentiator — top-of-README entry describing Pulse
  as background cron-driven processing with review_mode=Auto, and
  the [human-in-the-loop] marker for items needing human judgement.

v0.39.1

Fixed

• Plannotator gate now works end-to-end — new plannotator tool registered
  via pi.registerTool bypasses the bash block in gate stage. Writes receipt
  automatically on approval. turn_end auto-advances on receipt detection.

Changed

• stages.yaml: gate/int-gate block only bash — edit/write
  unblocked (needed for frontmatter stamp + receipt creation).
• gates_passed removed — declared in 22+ files but never populated.
  Replaced by direct receipt file detection. No stale state.
• gate.md rewritten — uses plannotator tool, no manual receipt step.
• Reference docs updated — tool path primary, CLI fallback secondary.

Added

• sem tool docs — entity-level diff for Execution Critique.

v0.39.0

CLI-Agnostic Subagent Fallback + Supervision

execHeadless() — 4 adapters

All CLI adapters (Pi, Claude Code, OpenCode, Codex) now have execHeadless(task, cwd?) — spawns the harness CLI non-interactively using the user's default model. No --model flag.

Adapter	Command
Pi	pi --print \"\$task\"
Claude Code	claude -p \"\$task\"
OpenCode	opencode -p \"\$task\"
Codex	codex -p \"\$task\"

subagents.md — Headless CLI Fallback

New section documenting when native subagent tool is unavailable (e.g. pi without pi-subagents extension). Includes per-CLI commands, structured output patterns, and parallel headless via & \ + `wait`.

supervise.md — CLI-Agnostic Supervision

Replaced Pi-only /supervise reference with dual-approach:

1. Headless CLI checkpoint (any harness) — discrete verification every N tool calls
2. /supervise (Pi only) — continuous real-time monitoring

Both activation table (old) and approach-by-appetite table (new) preserved.

v0.38.0

Pulse — Autonomous Inbox Processing

New background system that periodically checks .stelow/inbox/items.md and creates workflows with review_mode=Auto — no gates, no questions, no Plannotator.

Includes pulse.sh (bash, cross-platform), pulse.ps1 (PowerShell), provenance logging, and setup guides for macOS (launchd), Linux (systemd/cron), and Windows (Task Scheduler).

Key features

• Inbox item markers — prefix with [human-in-the-loop] or [hitl] to exclude from Pulse
• Conflict prevention — detects active user sessions and skips automatically
• /sw-pulse commands — status, pause, resume, process, log
• /sw-start now reads inbox when called without arguments
• /sw-inbox history — provenance log of processed items

Under the hood

• CLI-agnostic tool mapping — toAgnosticName() on all adapters, fixing a latent bug where ask in stages.yaml never matched ask_user_question
• Auto mode enforcement now blocks by agnostic tool name, not Pi-specific names
• Stages guard now uses agnostic names via adapter mapping

Tests

55 new tests in tests/unit/agnostic-tools.test.ts covering adapter mapping, stages guard integration, and Auto mode enforcement — all with real adapter instances, no mocks.

v0.37.0

Changed

• cali-product-coding-standards cleaned — removed Datastar-specific content
  (SSE-First, HATEOAS sections, LoB/SoC/tie-breaker Datastar bias). Now contains
  only universal principles (KISS, DRY, LoB, SoC, Fail Fast, YAGNI, size limits).
  Datastar design principles belong solely in cali-coding-go-stack. Affected
  references in README and cali-product-tech-planning updated.
• structured-question.md → ask.md — replaced Pi-specific
  ask_user_question tool reference file with CLI-agnostic ask.md across all
  24 skills. All skill-internal references updated.
• ctx7 → @vedanth/context7 — updated npx package name in
  doc-search.md references.
• Muxy extension improvements — adapter layer refactored for consistency
  across Pi, Claude Code, Codex, and OpenCode harnesses.
• Config/install updates — AGENTS.md, setup.sh, install.sh improvements.

Tests

• Updated skill-implementation.test.ts and appetite-consistency.test.ts to
  reference ask.md instead of structured-question.md.