v0.2.0

Added

• permissions.deny rules in the settings template for credential files and generated
  directories, so the file tools and the Bash commands that read files are both covered.
• Three subagent definitions — gatherer, reviewer and log-compressor — carrying their
  model, effort level and tool list, so the delegation tiers hold without a retyped brief.
• A sandbox stance and skill: container or built-in sandbox posture for an unattended loop.
• A filter-output PreToolUse hook that pipes a test, build, lint or type-check run through a
  line filter, keeping failures, summaries and the tail while preserving the exit status.
• A cost stance with frugal, balanced and max variants, and a cache-hygiene rule.
• A neutralize PostToolUse hook that flags instruction-shaped text in Bash, WebFetch and
  Read output, advisory only, never blocking or rewriting a result.
• A stop-gate Stop hook that runs the fenced ## Gate block of a repository's AGENTS.md
  and refuses to end the turn while it is red, bounded by a block count and a time budget.
• Four slash commands — /research, /plan, /build and /review — each composing skills
  the harness already ships.
• A usage-log SessionEnd hook and harness usage, reporting per-session tokens and cache hit
  rate from a local file; see docs/usage.md.
• A /handoff command that writes .claude/progress.md and promotes durable learnings into a
  dated file under docs/solutions/, and a SessionStart hook that reads that progress file and
  the last five commits back as context. The progress file is added to the global git ignore.
• A builder agent carrying the standing implementation brief — a worktree of its own, tests
  with every change, the repository's gate, one local Conventional Commit and no push — so
  /build spawns it and verifies the gate itself instead of retyping the steps.
• A spec-reviewer agent that reports scope deviations only, and a two-stage /review that runs
  it before the quality reviewer, each in a context that has not seen the other's findings.
• A planner agent carrying the Review Card contract, so /plan can hand the file-writing to a
  fresh context and keep its own for the review conversation.
• A design-judge agent carrying the design loop's scored critique and its hard gates, so the
  independent judge is a fixed definition the skill names rather than a rubric pasted each round.

Fixed

• The agent sync tests restore the environment they change, so a later test in the same run is
  no longer affected by them.

Changed

• Always-loaded context (claude/CLAUDE.md, every rule, and the longest variant of each stance) is
  capped at 200 lines and harness lint fails with a per-group breakdown when it is exceeded. The
  rules keep their operative lines and point at the skill holding the reasoning; the rationale,
  examples and evidence moved verbatim into delegation-tiering, plan-authoring,
  harness-authoring, the new transcript-hygiene and api-verification skills,
  docs/how-it-works.md and docs/preferences.md. 583 lines before, 185 after.