feat(agents): dispatch-implement — Codex CLI dispatch

[jinyeow]

Parent

#207

What to build

Extend dispatch-implement to work on Codex CLI, reusing the ticket-parsing and
parallel/sequential conflict judgment already built. Add a new [agents.implementer] table
to codex/config.toml, matching the existing [agents.fixer] table's shape
(description-only spawn guidance — per-role sandbox_mode/mcp_servers enforcement is
unsupported at codex-cli 0.147.0 per ai-agents/SKILL-OWNERSHIP.md, so real enforcement
stays the orchestrating session's own top-level sandbox_mode). Dispatch via
spawn_agent/agent_type: implementer, one call per ticket, calling /implement
unmodified inside each.

An explicit model/effort override in the invocation is refused with an actionable error on
Codex CLI, not silently ignored or reinterpreted: spawn_agent's only confirmed parameters
are agent_type and the task, with no per-call model/effort field — the only settable knobs
(agents.default_subagent_model / agents.default_subagent_reasoning_effort) are
session-global, and mutating them per dispatch would affect every other subagent spawned
later in the same session. This mirrors the precedent recorded in
docs/adr/reviewers-flag-unsupported-on-codex-cli.md for the --reviewers flag.

Acceptance criteria

• codex/config.toml has a new [agents.implementer] table.
• dispatch-implement dispatches via spawn_agent/agent_type: implementer on Codex
  CLI, one call per ticket, honoring the same parallel/sequential judgment as Claude
  Code.
• An explicit model/effort override passed on Codex CLI is refused with a clear,
  actionable message rather than silently dropped or applied.
• ai-agents/SKILL-OWNERSHIP.md updated to mark Codex CLI support as done.

Blocked by

• dispatch-implement: Claude Code dispatch

Likely files (optional)

• codex/config.toml
• ai-agents/skills/dispatch-implement/SKILL.md
• ai-agents/SKILL-OWNERSHIP.md