CHANGELOG.md

summary	Timeline of guardrail helper changes mirrored from Sweetistics and related repos.

Changelog

Unreleased

• Corrected GitHub secret provisioning to omit --body for stdin and added a skill validation guard against the literal-dash trap.
• Taught the clawsweeper-status snapshot to report queue handoff health, the ready/admissible split, backoff and parked reasons, and shed-since-reset, so exact-review items parked on retry exhaustion are no longer invisible behind a healthy verdict.
• Added the project-structure skill: a TypeScript symbol-map generator that compresses a repository into one context-loadable file with dense/skeleton/exports tiers, plugin-boundary listings, and measured token budgets.
• Removed the obsolete scoped-commit helper and returned commit recipes to standard Git now that agent work uses isolated worktrees.
• Made the million-token Codex provider and context settings an explicit atomic invariant, with a fatal preflight diagnostic for the unrecoverable openai plus 922K/700K split configuration.
• Added a fleet audit and repair action that disables Claude commit, pull-request, and session-link attribution while preserving unrelated settings and detecting higher-precedence overrides.
• Added narrow Homebrew 6 trust handling for exact third-party formulae already declared in a fleet profile.
• Made Apple-classified outdated/unusable Xcode runtimes and unavailable simulator devices required fleet drift, with a booted-device-safe audit and repair action.
• Established the SF Mini's classic OpenSSH fleet path, documented symmetric tailnet TCP 22 policy and proof rules, and recorded the MiniClaw duplicate-daemon regression and public-SSH fallback.
• Added the separately owned SF Mac Mini and distinguished its local/Tailscale names from FoundationClaw while its trusted SSH path remains pending.
• Verified FoundationClaw's provider identity, installed Tailscale and Jump Desktop Connect v10, and documented its data-preserving credential-reset escalation before GUI activation.
• Added ClawMac provider-outage triage that escalates console, NIC-link, and switch-port inspection without repeated power cycles or data-affecting recovery.
• Restored MiniClaw's canonical Homebrew Tailscale node and removed its duplicate GUI identity from fleet guidance.
• Removed obsolete Mac identities from remote fleet discovery guidance after pruning them from Tailscale.
• Reconciled the remote-Mac topology with provider purchases, including FoundationClaw's MacStadium identity and MiniClaw's canonical Tailscale identity.
• Reserved GPT-5.6's maximum output budget in codex-huge-context and moved fleet compaction to a verified 922K input window with a 700K safety threshold.
• Made codex-first treat the Gorilla-backed Clawdex endpoint as already model-routed, preventing recursive Codex delegation after the fleet proxy migration.
• Added a secret-safe Codex direct-API preflight so million-token launches fail before an unauthenticated Responses request when a machine is missing its Keychain delivery copy.
• Generalized interactive 1Password routing to select the active approval workstation within the matching personal or work-managed environment while preserving service-account isolation and safe offline fallback.

2026-07-17 — 0.12.0

Highlights

• Turned maintainer-orchestrator into a long-running control plane for autonomous queue triage, proof-driven changes, dependency maintenance, and release proposals across Peter's repositories.
• Added fleet maintenance, safe repository synchronization, package ownership audits, and Xcode fleet management for Peter's Macs.
• Replaced the old Codex review path with isolated structured autoreview and added Claude Code-only codex-first delegation for implementation-heavy work.
• Added scripts/sync-skills so Codex and Claude share one canonical skill and instruction mirror across agent-scripts, manager, and repo-owned skills.
• Hardened 1Password, npm, and macOS release workflows around scoped service access, stable tool identities, noninteractive signing, and verified publication boundaries.

Maintainer Orchestration

• Expanded maintainer-orchestrator with one tracked Codex thread per repository, 30-lane scheduling, durable status, forgotten-work preservation, exact-head landing, dependency sweeps, VISION capture, and release-readiness proposals.
• Added a dedicated OpenClaw mode with root-owned discovery, qualified execution lanes, contributor routing, live permission checks, serialized landing, and OpenClaw-specific proof and changelog rules.
• Added autonomous GitHub queue triage with URL-first item briefs, maintainer-comment routing, author context, live proof requirements, safe spam closure, and explicit Peter decision briefs.
• Added the non-majority repository ledger, owner-maintained crawl-family overrides, and clearer root ownership for orchestration policy and worker titles.
• Made dependency updates, internal operating repositories, bounded cleanup, safe dirty fast-forwards, and candidate-scoped release blockers autonomous.
• Improved ClawSweeper status reporting for worker capacity, exact-review occupancy, workflow waiters, bounded API reads, and accurate failure and closure counts.
• Tightened shared agent policy around exact-head proof, contributor credit, screenshot safety, post-merge recaps, background-task visibility, public mutation, and release authority.

Review and Agent Workflows

• Replaced codex-review with structured autoreview, adding isolated Codex, Claude, and Pi review, safe bundle validation, regression provenance, security checks, parallel tests, and bounded multi-pass review.
• Added Claude Code-only codex-first routing for implementation, fixing, exploration, rebasing, and landing mechanics, including safe use of the ChatGPT app-bundled Codex CLI. Thanks @notorious-d-e-v.
• Added current model, effort, fast-mode, liveness, deterministic resume, and self-delegation guardrails to the Codex workflow.
• Made GitHub deep review and project triage prefer current source, real behavior reproduction, exact PR heads, and factual contributor trust signals.
• Routed screenshot and login-dependent browser work through existing Chrome state, with safe attach recovery and no silent isolated-browser fallback.
• Added explicit compatibility contracts, clean bounded-refactor guidance, generated-code skepticism, and scoped opportunistic cleanup rules.

Fleet, Release, and Remote Operations

• Added fleet-maintenance for host health, package updates, repository synchronization, ownership collisions, disk cleanup, and service-impact reporting.
• Added xcode-sync for signed Xcode inventory, stable and prerelease slot management, build-identity checks, platform compatibility, and verified installation.
• Added dependency-light fleet repo audit and update helpers with batched snapshots, clean fast paths, dirty-work preservation, collision checks, and safe fast-forwards.
• Added release-mac-app and shared macOS release helpers for changelog notes, Sparkle appcasts, signing, notarization, GitHub assets, and post-release verification.
• Hardened macOS releases for passwordless isolated keychains, preloaded secrets, modern distribution validation, Bash 3.2, lightweight tags, bracketed changelog versions, and archive version parsing.
• Refreshed remote-Mac topology and network-boundary guidance, signed local app testing, locked-Mac Git fallback, and Cloudflare-only ClickClack deployment.
• Kept GitHub reads on the Octopool shim and added cache-health recovery before live GitHub fallback.

Credentials and Safety

• Unified 1Password work on one tracked tmux session with scoped service-account access first, consent-gated desktop fallback, exact-field reads, known-item routing, and TCC-safe ~/bin/op updates.
• Unified npm authentication around reusable service sessions, safe field selection, token caching, login fallback, package reservation, publication verification, and a generic authenticated command wrapper.
• Added internal-information, confidentiality, device-aware image upload, API-key storage, and approved-destination guardrails without blocking authorized private research.
• Standardized the canonical test Gmail account, OpenClaw deployment account, personal versus corporate Mac routing, and pre-approved Gmail service login behavior.

Skills and Tools

• Added scripts/sync-skills to build Codex whole-root links, Claude's flat skill mirror, shared instruction pointers, deterministic collision handling, and stale-link pruning.
• Added skills for fleet maintenance, Xcode sync, Codex delegation, Twilio SMS, Wrangler, Things, Reminders, SSH diagnosis, agent transcripts, and shared macOS releases.
• Added skill-cleaner inventory, duplicate, usage, and prompt-budget audits plus isolated --root-only scans. Thanks @its-How.
• Added browser-tools network capture with filtering and follow mode. Thanks @mvanhorn.
• Hardened browser-tools startup, profile copying, symlink handling, and console flags. Thanks @ShiroKSH.
• Fixed xurl's OpenClaw npm installer metadata. Thanks @not-stbenjam.
• Made skill validation explicitly UTF-8-safe under C locales. Thanks @chaochaoweb3.
• Added skills.sh grouping metadata for shared skills. Thanks @vyctorbrzezowski.
• Exposed shared behavior validation, session viewing, crabbox, and crawl-family skills through the canonical mirror while removing duplicated bundled copies.

2026-05-14 — Video Transcript Dependency Update

• Updated video-transcript-downloader to youtube-transcript-plus 2.0.0.

2026-05-14 — Codex Review Finding Detection

• Updated codex-review to capture review output, report elapsed time, fail on reported P0-P3 findings, and treat empty review output as non-clean.

2026-05-14 — Codex Review Full Access

• Added codex-review --full-access for nested review runs that need localhost bind/listen tests without sandbox noise.

2026-05-14 — GitHub Search Shim Guidance

• Added AGENTS guidance to prefer shimmed gh / gitcrawl gh for broad reads and avoid raw Search API POST mistakes.

2026-05-14 — Codex Review Base Caveat

• Documented that codex review --base must not include an inline prompt; use a separate follow-up pass for custom instructions.
• Clarified that committed or PR branch review must use branch/base mode, not --uncommitted / local mode.

2026-05-14 — Codex Review Loop Guidance

• Clarified that codex-review should iterate until no accepted findings remain and document intentional rejections with useful inline comments when warranted.

2026-05-14 — README Skills Overview

• Rewrote the README around agent instructions, skills, helper scripts, and sync expectations; removed stale copied-origin notes.

2026-05-14 — Codex Review Skill

• Added a codex-review skill and helper for closeout reviews, with stdout-only default output and subagent filtering guidance for noisy review output.

2026-05-13 — Checkout Discipline

• Added CLI checkout/worktree guardrails: stay in repo cwd by default, never create worktrees unless asked, and treat sibling checkouts under ~/Projects as user-managed.

2026-05-13 — Skill Metadata Guardrails

• Added generic skill-description guidance and quieter browser recovery notes to reduce noisy auth prompts and token-heavy skill metadata.

2026-05-11 — clawmac GUI Access Note

• Documented the Peekaboo through Jump Desktop workflow for clawmac GUI prompts and Chrome Safe Storage verification.
• Documented crabmac as Peter's typo/alias for clawmac.

2025-12-22 — Remove Custom rm Shim

• Dropped bin/rm and scripts/trash.ts; rely on the system trash command for recoverable deletes.

2025-12-17 — Remove Runner; Keep Guardrails

• Removed the runner wrapper and scripts/runner.ts now that modern Codex sessions handle long-running/background work directly.
• Kept the safety-critical bits as standalone shims: bin/rm (moves deletes to Trash via scripts/trash.ts).
• Dropped the find -delete interception and the bin/sleep shim.

2025-12-02 — Release Preflight Helpers

• Added shared release helpers in release/sparkle_lib.sh: clean working-tree check, Sparkle key probe, changelog finalization/notes extraction, and appcast monotonicity guard for version/build.
• Documented the helper functions in docs/RELEASING-MAC.md so Trimmy/CodexBar-style release scripts can reuse them.

2025-11-18 — Console Log Capture

• Added console command to scripts/browser-tools.ts for capturing and monitoring Chrome DevTools console output with real-time formatting, type filtering (log, error, warn, etc.), continuous follow mode, and configurable timeouts with automatic object serialization.

2025-11-22 — Search & Content Extraction

• Added search and content commands to scripts/browser-tools.ts for Google SERP scraping with optional readable markdown extraction and single-URL readability output, leveraging the existing DevTools-connected Chrome instance.
• eval now supports --pretty-print to inspect complex objects with indentation and colors.

2025-11-15 — Chrome Browser Tools

• Added scripts/browser-tools.ts, a DevTools-ready Chrome helper copied from the Oracle repo so agents can inspect, screenshot, and terminate sessions without dragging in the full CLI. The workflow is inspired by Mario Zechner’s “What if you don’t need MCP?”.
• Documented the new helper in the README so downstream repos know how to run pnpm tsx scripts/browser-tools.ts --help.

2025-11-16 — Browser Tools Pipe Detection

• Updated scripts/browser-tools.ts to enumerate and kill Chrome instances started with --remote-debugging-pipe (the default for Peekaboo/Tachikoma) in addition to the classic --remote-debugging-port. List/kill now show “debugging pipe” when no port exists and still fetch tab metadata when it does.
• README now notes the optional NODE_PATH=$(npm root -g) trick so the helper can run from bare copies of the repo without a local package.json.

2025-11-14 — Compact Runner Summaries

• The runner's completion log now defaults to a compact exit <code> in <time> format so long commands don't repeat the entire input line.
• Added the RUNNER_SUMMARY_STYLE env var with compact (default), minimal, and verbose options so agents can pick how much detail they want without editing the script.
• Timeout heuristics now understand both pnpm and bun invocations automatically, so long-running Bun scripts/tests get the same guardrails without repo-specific patches.
• sleep invocations longer than 30 seconds are clamped to the 30s ceiling instead of erroring, which keeps wait hacks working while still honoring the AGENTS.MD limit.

2025-11-08 — Sleep Guardrail & Git Shim Refresh

• Runner now rejects any sleep argument longer than 30 seconds, mirroring the AGENTS rule and preventing long blocking waits.
• Added bin/sleep so plain sleep calls automatically route through the runner and inherit the enforcement without extra flags.
• Simplified bin/git to delegate directly to the runner + system git, eliminating the bespoke policy checker while keeping consent gates identical.

2025-11-08 — Guardrail Sync & Docs Hardening

• Synced guardrail helpers with Sweetistics so downstream repos share the same runner, docs-list helper, and supporting scripts.
• Expanded README guidance around runner usage, portability, and multi-repo sync expectations.
• Tightened path ignores and refreshed misc. helper utilities (e.g., toArray) to reduce drift across repos.

2025-11-08 — Initial Toolkit Import

• Established the repo with the Sweetistics guardrail toolkit (runner, git policy enforcement, docs-list helper, etc.).
• Ported documentation from the main product repo so other projects inherit the identical safety rails and onboarding notes.