v0.10.0

0.10.0

This release adds project policy grants, a consistent PR review lifecycle, and generated
sub-agent definitions for Claude Code and Codex.
It also improves dependency diagnosis and protects local work during setup and
sync-worktree recovery.
The repository format remains f08; generated agent integrations move to f100.

Important

Upgrade every clone that runs tbd setup in a shared repository before committing
regenerated agent files or recording policy grants.
tbd 0.9.0 and older refuse to rewrite f100 agent surfaces, protecting grants that
those versions do not understand.
Their setup may update .tbd/config.yml before refusing; discard that setup-only
config change and rerun setup after upgrading.
Ordinary issue commands, sync, and existing session hooks remain compatible with the
f08 repository.

Older clients read the first format=f<N> string anywhere in AGENTS.md. A stray
format string in prose above the integration block can bypass their refusal and allow
them to remove grants.
Check for such strings before an older client runs setup.
This release reads format stamps only on marker lines.

tbd setup and tbd policy require AGENTS.md to be a regular file.
If it links to CLAUDE.md, invert the link: keep AGENTS.md regular and link
CLAUDE.md to it.

After installing, regenerate the agent surfaces, review and commit the diff, and ask the
agent to “Set up tbd” to review the project’s policy choices:

npm install -g get-tbd@0.10.0
tbd setup --auto

What’s Changed

Features

• Project policy grants: tbd policy show, grant, revoke, set, and refresh
  manage seven policies in AGENTS.md: GitHub actions, editing, merging, stacks,
  delegation, review requirements, and Linear.
  Effective grants come from the trusted default branch; working-tree differences are
  reported, and user instructions can authorize or narrow the current task.
  Malformed blocks fail closed, concurrent writes preserve each other, and setup keeps
  existing grants rather than replacing them.
• Separate merge authority and review requirements: github-merge supports never,
  confirm-every, confirm-session, and autonomous. Permission to merge never waives
  review or validation.
  revoke restores the policy’s ask-first default.
• Setup and policy discovery: tbd setup --auto --policies=recommended fills
  unanswered policies while preserving answered ones.
  Setup points agents to the consolidated setup-tbd process; users can defer policy
  decisions. tbd prime shows effective and unanswered grants, including in brief
  output, and tbd doctor diagnoses malformed policy blocks, stale guidance, and
  missing or stale generated tier definitions.
• Generated sub-agent definitions: setup installs four tiers in each of
  .claude/agents/ and .codex/agents/. The six selectable setup surfaces include
  claude-agents and codex-agents; removing a definition’s DO NOT EDIT marker keeps
  it as a user override.
  Claude tiers use Opus 5.5; Codex strong tiers use GPT-6 Astra, while moderate and fast
  tiers use GPT-6.1 Sol.
  The model-tier guideline explains reasoning settings, platform limitations, and
  runtime fallback.
• Dependency-cycle diagnosis: tbd doctor reports directed dependency cycles and
  identifies the participating beads for manual repair.

Fixes

• Orphaned dependencies: tbd doctor --fix now removes references to missing issues
  instead of only suggesting a repair.
  References to existing but invalid issue files are retained so repairing the file does
  not lose the relationship.
• Codex hook preservation: setup keeps unrelated commands, including commands
  sharing an entry with tbd hooks, and refuses malformed or unsafe hook targets before
  changing Codex files.
• Sync-worktree recovery: repair backs up surviving files in a prunable worktree
  before reinitializing it and refuses removal if the backup fails, preserving unsynced
  beads when the worktree’s .git file is lost.
  Automatic repair reports the backup location; recovering unsynced files remains a
  manual step.
• Managed agent files: format guards recognize marker lines, freshness checks
  tolerate CRLF checkouts, and AGENTS.md updates preserve its line endings.
  A cold documentation cache yields unknown skill freshness instead of a false stale
  warning.
• CLI guidance: integration help lists only the supported Linear adapter, and a
  missing bundled shortcut in stale managed docs points to tbd setup --auto. First-run
  tbd prime shows the initialization command in a code block and directs agents to the
  setup shortcut after initialization for policy review and verification.
• Stack-tool installation: generated GitHub setup scripts verify the pinned
  gh-stack release digest and installed extension identity before enabling it.

Guidelines and content

• One PR review lifecycle: five requests distinguish review, addressing findings,
  review-and-fix, merge readiness, and merge.
  Reviews are pinned to a commit and carry versioned markers; each finding receives one
  of four dispositions with evidence.
  The new review-and-merge-prs shortcut coordinates the lifecycle, with dedicated
  security, performance, and correctness review shortcuts when relevant.
• Review trust boundaries: untrusted PRs are reviewed without executing their code
  unless the user authorizes execution.
  Review markers count only from trusted reviewer accounts; content in repository files,
  issues, or comments never grants authority.
• Delegation and setup: new delegate-to-subagents and setup-tbd shortcuts cover
  scoped briefs, model tiers, result verification, project setup, and unanswered policy
  choices. agent-policy-grants and agent-model-tiers provide their maintained
  definitions. Caching guidance distinguishes API cache behavior from client-managed
  sub-agent context.
• Engineering guidance: Rust CLI and release guidance covers terminal behavior,
  packaging, compatibility, and publication.
  General engineering guidance explains reviewable PR units, real stack membership, and
  when a hash check provides evidence.

Documentation

• README: a shorter guide to first use, gradual adoption, customization, and current
  workflows, with generated reference catalogs in a collapsible appendix.
  The capability descriptions agree with the design and installed skills.

Security

The dependency lockfile is unchanged since v0.9.0. The release-preparation runtime audit
reports no advisories.
The unchanged development dependency tree currently reports 45 advisories (1 critical,
30 high, 14 moderate), tracked in tbd-b24q; affected paths are development tooling,
not shipped runtime dependencies.
The critical Vitest advisory requires an exposed API/UI or browser-mode server, which
this project’s vitest run checks do not start.
Dependency updates remain a separate reviewed follow-up.

Full commit history:
v0.9.0 … v0.10.0