v2.39.0

Fixed

• The 97 agents/genius/*.md reasoning-pattern agents were auto-registered as
  invocable Agent-tool subagents with zero measured usage. An audit of every
  past Claude Code session transcript (~768MB across 31 projects) found 97 of
  98 genius agents were never invoked via the Agent tool, and the one that was
  (lamport, once) was never run in parallel with anything, no isolation or
  parallelism benefit was ever exploited, and tool restriction was already
  ruled out (every genius agent shares the orchestrator's own toolset).
  /genius-invoke, /genius-compose, and /genius-route already cover this
  ground more cheaply: all three read agents/genius/<name>.md directly and
  apply the pattern inline, no subagent spawn. plugin.json now declares an
  explicit agents array listing only the 23 team agents, so Claude Code no
  longer auto-registers the genius agents as a second, redundant invocation
  path. The genius agent files themselves are untouched, still fully
  readable by the three commands above, still 97 in number, still never a
  removal candidate under this repo's own standard. Also re-synced
  plugin.json's embedded hooks copy against hooks/hooks.json (authoritative
  since the Stop-hook dedup fix, 4bd2f1d0), which had drifted and still
  carried the duplicate registration that fix already resolved in the
  canonical file.

Added

• rules/agent-vs-skill-classification.md: five-test rule deciding whether a
  team agent may be inlined as a skill (#120). Tests are tool-grant
  guardrail, isolation need, model-tier delta from the sonnet baseline
  (tools/skill-runner.sh:31,35), packaged assets, and authority transfer.
  Applied as a worked table over all 23 team agents with per-agent file:line
  evidence: 9 subagent-only, 1 dual-surface, 13 skill-frontable. A verdict is
  a candidacy, never an authorization to remove an agent, and the rule is
  scoped to agents/*.md only, never to the 97 genius agents. Two premises
  the plan asked the table to assume were checked and corrected against
  origin/main: agents/ux-designer.md:8 no longer carries the narrow
  read-only tool grant its dual-surface staging was justified by (widened in
  #116, commit 4006565, as a frontmatter-vs-body consistency fix), and
  refactorer is named by no skill's agents: field. Recorded follow-up:
  #119's escalation banner keys on model tier only, so it stays silent for
  the four subagent-only agents that sit at the sonnet baseline
  (code-reviewer, git-historian, research-scientist, reviewer-academic).

• tools/skill-runner.sh model-tier escalation banner (#119). A skill
  whose agents: frontmatter names an agent above the sonnet baseline
  (opus, fable) now gets a !!! MODEL-TIER ESCALATION REQUIRED !!! banner
  before the procedure body, naming the agent and directing the caller to
  spawn it as a real subagent instead of inlining its procedure at the
  caller's own tier. commands/skill/run.md is updated to act on it.
  scripts/generate-skill-agent-model-matrix.py computes the full
  skill-to-agent-to-model matrix from the tree (docs/skill-agent-model-matrix.md,
  drift-checked): 65 skill files carry a non-empty agents: field, and 63
  of them name at least one escalation-tier agent.

• Delegation-contract schema + fail-closed validator (#116). schemas/delegation-contract.schema.yaml
  declares ownership, worktree policy, push authority, handback artifacts, an
  external acceptance oracle, model, tool grant, and checkpoint policy for
  any scripts/spawn-agent.sh delegation. tools/delegation_contract.py
  (stdlib-only, no new dependency) validates it and is now called BEFORE
  git worktree add: a missing, malformed, or ownership-overlapping
  contract is denied with a non-zero exit and creates nothing.

• Strict, full-tree frontmatter validator (#116). tools/frontmatter_validator.py
  parses every declared frontmatter block under skills/, agents/
  (incl. agents/genius/), commands/, and the packaged
  plugins/*/skills/ surface with PyYAML's strict safe_load, reporting
  stable {file, line, rule} records. Found and fixed 7 skill files whose
  unquoted input:/output: values contained a colon-space, parsed as a
  nested mapping key.

Fixed

• scripts/spawn-agent.sh created a git worktree before any precondition
  check (#116). No ownership, scope, push-authority, handback, or
  acceptance-oracle validation happened before the mutation. It now calls
  tools/delegation_contract.py first and refuses to create anything on an
  invalid or missing contract. A follow-up review found that three of its
  VAR="$(cmd)" assignments were bare statements under set -e, which
  triggers on the assignment's own exit status before a subsequent
  VAR_RC=$? check runs: a rejected contract exited 1 with the intended
  "deny: ..." reason silently swallowed. Moved each into an if ! condition
  (exempt from set -e by design) so the reason reaches stderr.

• Push authority was stated three different, contradictory ways (#116).
  The generated zetetic-spine text (scripts/generate-spine.py) said "push,
  and hand back immediately" unconditionally; ~116 agents' own inline
  <worktree> sections said "do NOT push"; worktree-protocol.md already
  said push is delegation-controlled. All three now defer to the delegation
  contract's push_authority field, surfaced at runtime as
  DELEGATION_PUSH_AUTHORITY. Also fixed in the same pass: ux-designer.md
  declared tools omitted Bash/Edit/Write despite its required
  procedures using them, and orchestrator.md's frontmatter selected
  model: fable while its token-budget prose described Opus's budget.