Make CLI and MCP ergonomics agent-first

[galligan]

Why

Dispatch is used by humans and agents. The current op-derived surfaces work, but the user-facing grammar still feels too raw in places (goal-get, repeated --lane @docs, roster, unclear brief). The next pass should preserve the contract-first foundation while making the surface nicer and more stable for real operator loops.

There is no public compatibility burden yet, so prefer the right surface over keeping awkward top-level aliases.

Scope

Improve CLI/MCP ergonomics without reintroducing drift. Prefer derived command groups, resolvers, and surface projections over hand-written per-op behavior.

Acceptance criteria

• Design and implement a consistent command grammar for related operations:
  • Prefer grouped forms such as dispatch goal see|set|clear over top-level goal-get|goal-set|goal-clear.
  • Do not keep old command aliases solely for backwards compatibility; no one depends on them yet.
  • Revisit trigger commands in the same style (dispatch trigger add|list|rm|pause|resume) or explicitly defer to the trigger issue.
• Add a universal lane/thread resolver for lane arguments:
  • Resolve current handles like @docs.
  • Resolve dispatch lane ids.
  • Resolve full Codex thread ids.
  • Resolve unique truncated thread/lane ids with conflict detection.
  • Preserve and resolve handle/name history so renamed sessions remain findable.
  • On ambiguity, return actionable candidates instead of guessing.
  • Choose and document a stable short-id display format.
• Clarify lane vs thread language:
  • thread is the Codex/App Server identity.
  • lane is dispatch's managed/control-plane view of a thread.
  • Update docs/skills/help text so this distinction is clear without being noisy.
• Revisit message/context operation names:
  • Decide whether brief should remain or be replaced by a clearer send/context verb.
  • Explore collapsing turn-starting send and non-turn-starting context injection under one verb with an explicit mode/flag, while keeping the App Server semantics clear.
  • Clarify send, steer, and any replacement for brief in help text and MCP tool guidance.
• Revisit overview/listing commands:
  • Decide whether roster should become dispatch list or be replaced by a more operational overview command.
  • Decide whether discover should remain separate from managed-lane listing, or whether one command can show both managed lanes and attachable Codex threads with clear state/source fields.
  • Consider a more operationally minded command such as sitrep for overseeing lanes: current known state, recency, active turn/goal, and last N actions.
  • Confirm dispatch status remains daemon health, or split lane/list/status semantics if needed.
• Make CLI output more agent-friendly:
  • Stable JSON schemas for key commands.
  • Consistent field names across CLI/MCP.
  • jq-friendly examples in docs.
• Reshape MCP tools around ergonomic grouped tools rather than tool sprawl where that improves agent behavior, while keeping read/write/destructive hints truthful.
• Add tests proving command grammar, resolver behavior, ambiguity handling, MCP grouping, and docs examples stay in parity.

Notes

This should not become a hand-written CLI fork. The important design constraint is still: author once, derive what can be derived, and override only where the user experience is genuinely better.