auto-driver

One action line per branch. Dispatch. Do not narrate.

Load the hypothesis

bash "${CLAUDE_PLUGIN_ROOT}/lib/auto-detect.sh"

If the detector yields no parseable envelope (empty/non-zero env hiccup), don't stall — treat as raw.

Returns one JSON object (situation, summary, ambiguity, single_plan, multi_plan, in_flight, workspace, workspace_action, recommendation, plans, git). The detector emits deterministic facts — it has no transcript access. You have the transcript: surface summary, then decide the route from situation + ambiguity + the plans/git facts. The conversation-vs-stale-plan call is YOURS, not the detector's (U4 — see below).

situation	ambiguity null → dispatch	ambiguity non-null → AskUserQuestion
in-flight	(FRESH run) bash "${CLAUDE_PLUGIN_ROOT}/lib/auto-resume.sh" "continue <run-id>"	(STALE run) options = resume vs start-fresh; on the resume option (carries run_id) → auto-resume.sh "continue <run-id>"; on "Start fresh" (run_id null) → treat as raw (ask what to work on)
ambiguous-runs	(n/a — always ambiguous)	options = the in-flight run-ids; on answer, resume the chosen run
reviewed-plan	if single_plan.freshness is stale and this session is about other live work, prefer conversation-context (below). Else run the goal-aware pre-step; if it does not reshape, load auto-launch (the launch chooser) via Skill: it gates on driving_session_id — self-driven silent-applies, interactive confirms — then dispatches lib/auto.sh "<path> --workflow w"	(n/a — single plan unambiguous)
multi-plan	If every plans[].freshness is stale and this session is about other live work, use conversation-context (below) and do NOT offer the stale plan actions; otherwise run the goal-aware pre-step first	options = the FRESH plans (path → auto.sh "<path>"); a "Fan out all N" option (path null → auto-spawn.py fanout) appears ONLY when the set is fresh. An all-stale set offers no plan actions — it routes to conversation-context
raw	(n/a — always ambiguous)	open "what should we work on?"; on answer, route as freeform text (or conversation-context, below). Summary may include dirty-tree context (the git fact).

Conversation-context is driver-owned (U4). The detector no longer emits a conversation-context situation — it can't sense the transcript. When the facts show no in-flight run and no LIVE plan to act on (situation = raw, or a reviewed-plan/multi-plan whose plans are all freshness: stale) AND this session is clearly about live work, DON'T act on the stale/absent plan: classify state → recommend (lib/recommender.py) → author goal → dispatch the entry workflow (pre-dispatch escalate if unsure). A FRESH plan (freshness: fresh) always wins over conversation.

Goal-aware plan routing (pre-step)

For reviewed-plan / multi-plan, interactive runs only (driving_session_id null — self-driven/headless runs skip this whole step, take the row above unchanged). Recover THIS invocation's goal: the typed /auto intent or a session /goal's text (read only — never query/run/bind/clear it); else infer from the session (advisory). Weight the plans (the fuzzy judgment is yours → ordered matched paths), hand verdicts to the crisp router, run its reason (map + guardrails: driver-reference.md §17; rubric goal-plan-relevance-rubric.md):

python "${CLAUDE_PLUGIN_ROOT}/lib/goal-route.py" \
  '{"authority":"<explicit|inferred|none>","matches":[<ordered matched paths>],
    "all_plans":[<multi_plan.paths>],"interactive":<driving_session_id is null>}'

Reason markers (§17): explicit-suppress, inferred-re-rank, no-match-unchanged, no-goal-unchanged.

Argument-aware freeform: a plan path → auto-launch → auto.sh "<path> --workflow w"; else classify with lib/verb-classify.py → the auto-launch chooser — freshest plan / clear-intent-no-plan→a1@plan / /ce-plan.

Workspace / conversation-context / unknown (§17): branch workspace_action; unknown → raw.

Dispatch grammar (reference)

• Single plan: bash lib/auto.sh "<plan-path> [auto] [--review-plan] [--backend ce|native] [--goal <text>] [--workflow <name>]"
• Batch fanout: python lib/auto-spawn.py fanout <plan> [<plan>...] [--intent "<text>"] — the spawner owns worktrees/ports/sidecar + dispatch; the driver never shells out per sub-run.
• Resume: bash lib/auto-resume.sh "<continue|abort|retry|skip> [<run-id>] [<step>]"

Theory + edge cases (docs/contracts/driver-reference.md): prepare/execute, goal binding, pulse/handoff/fan-out, exit reasons, goal-aware routing (§17) — load on demand.