v4.3.0

[4.3.0] - 2026-07-29

The agents join the system they were supposed to be part of. Before this
release an agent was a fourth front-end without parity: it carried its own
copy of the rules, re-derived structure the plugin had already computed, and
nothing it produced fed the learning loop. Each fix below was reproduced
before it was fixed and carries a test that fails if it regresses.

Fixed

• The orchestrator told the model to run skills the model cannot run.
  Sixteen of twenty-two skills carry disable-model-invocation: true, which
  means only the user can start them by typing the slash command first in a
  prompt. /craftsman:workflow nonetheless announced "Invoking
  /craftsman:design..." for four of its seven steps, and five agents listed
  locked skills in skills: frontmatter. Every one of those references was
  unreachable: the Skill tool answers cannot be used with Skill tool due to disable-model-invocation and the pipeline stops there. The workflow now
  hands off, printing the exact command to paste and waiting, and the nine dead
  agent references are gone.

• /craftsman:team could not be started by anything except a bare prompt.
  It was classified as a heavy review in ADR-0017 and given
  disable-model-invocation: true, so no orchestrator could reach it. The flag
  is dropped and the skill is model-invocable. It stays out of a fork on
  purpose: it asks the user which template to use and calls TeamCreate to
  spawn teammates into the session, and a fork would strand both.

• An agent could be told to fork into itself. agents/architect.md
  declared craftsman:challenge, which is bound to agent: craftsman:architect.
  Same shape in agents/team-lead.md for craftsman:team.

• The subagent quality gate validated nothing. Its header said "validates
  code produced by subagents against craftsman rules"; its body logged an
  agent_type and a timestamp. It now reads the subagent's transcript, runs
  every file the subagent wrote through the same pack validators as the hooks
  and CI, records each finding in the metrics DB tagged
  subagent:<agent_type>, and surfaces the findings to the main loop as
  additionalContext. Silence remains the pass signal, and the gate stays
  non-blocking.

• The agent context injector never reached an agent. The old
  agent-structure-analyzer.sh asked each agent to re-scan the repository,
  from a hook on InstructionsLoaded - a side-effects-only event whose output
  the platform ignores, aimed at an event that does not fire on agent dispatch
  anyway. Removed. Agents now start by running
  hooks/lib/dispatch-context.sh, which returns the resolved doctrine (same
  rules engine as hooks and CI), the HEAD-cached codemap, the top hotspots and
  the correction trends: one deterministic turn instead of a self-guided scan.

• A review could silently run a tier below what the docs promise.
  challenge declares model: opus and forked into an architect declaring
  model: sonnet; the platform does not document which wins. architect,
  legacy-surgeon and team-lead now declare opus, matching the reference
  table, and the invocation-policy test fails any forking skill whose bound
  agent disagrees on the model.

• subagent-quality-gate.sh printed a bare counter onto its JSON channel.
  session_state.py increment echoes the new value; the hook never silenced
  it, so every SubagentStop emitted a stray 1 on stdout.

Added

• hooks/lib/dispatch-context.sh: single source of start-of-dispatch context
  for all 12 agents (resolved doctrine, cached codemap, hotspots, correction
  trends). The agent is a front-end of the rules engine like hooks and CI:
  it no longer carries its own copy of the rules.
• Per-agent violation metrics: the source column now distinguishes
  subagent:<agent_type> rows, so /craftsman:metrics can show which agent
  keeps producing which violation and agent prompts can be tuned from data.
• Agent contracts: every craftsman agent ends with an Output Contract (test
  evidence required before "done"), and every agent carries a one-line Memory
  Contract saying exactly what it may persist; reviewers persist rejected
  findings so the same false positive is not raised twice.
• /craftsman:team degraded mode: when CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS
  is absent the skill announces it once and runs the same composition through
  parallel subagent dispatches instead of failing; the healthcheck reports
  which mode the environment provides. The team-lead delegation matrix also
  gained the missing legacy row (legacy-surgeon / legacy-takeover template).
• tests/core/test-invocation-policy.sh crosses every agent skills: entry and
  every **Invokes:** claim with its target's invocation policy, and rejects
  self-forking bindings. Nothing checked this before, which is why sixteen
  locked skills and nine dead references coexisted with a green suite.