auto-driver

auto-driver

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

Load the hypothesis

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

Returns one JSON object with situation, summary, ambiguity, single_plan, multi_plan, in_flight, workspace, workspace_action, recommendation (driver-filled, U2). Surface summary; act on situation + ambiguity:

situation	ambiguity null → dispatch	ambiguity non-null → AskUserQuestion
in-flight	bash "${CLAUDE_PLUGIN_ROOT}/lib/auto-resume.sh" "continue <run-id>"	(n/a — single run unambiguous)
ambiguous-runs	(n/a — always ambiguous)	options = the in-flight run-ids; on answer, resume the chosen run
reviewed-plan	bash "${CLAUDE_PLUGIN_ROOT}/lib/auto.sh" "<path>"	(n/a — single plan unambiguous)
multi-plan	python "${CLAUDE_PLUGIN_ROOT}/lib/auto-spawn.py" fanout <plan...> then surface manifest	(n/a — confirm-only in summary)
conversation-context	classify state → recommend → author goal → dispatch entry recipe (see below)	(n/a — pre-dispatch escalate if unsure)
raw	(n/a — always ambiguous)	open "what should we work on?"; on answer, route as freeform text. Summary may include dirty-tree context.

Argument-aware freeform: before loading the hypothesis, if $ARGUMENTS is non-empty AND does NOT resolve to a plan file, invoke /ce-plan <ARGUMENTS> via Skill and end the turn (v0.3.x routing).

Workspace handling (plan 004): branch on workspace_action. create/recreate: chain in ONE Bash call so the workspace id propagates: WS=$(python lib/auto-workspace.py create <repo> [--force] --print-id) && CMUX_WORKSPACE_ID="$WS" python lib/auto-spawn.py fanout. use/none: dispatch. ambiguous: ask switch/create/one-off.

Conversation-context (v0.6.0, full detail in driver-reference.md §11/§13): classify the current transcript + a ~2-day ce-sessions lookback (NOT raw compaction) into one state → python lib/recommender.py <state> <confidence>. escalate/ambiguous → one AskUserQuestion BEFORE dispatch, no run (NOT via the gate). kind=skill (bug/what-to-improve/perf) → recommend the ce command, no wrap. kind=recipe (vague→pipeline@brainstorm — the spine, auto-advances brainstorm→plan→work; clear-intent→a1@plan; reviewed-plan→w@work; code-unreviewed→review@work) → auto-author-goal → goal doc (bind auto's OWN predicate, NEVER native /goal) → bash lib/auto.sh "<goal-doc> --recipe <name>".

Unknown situation (defensive guard): treat as raw.

Dispatch grammar (reference)

• Single plan: bash lib/auto.sh "<plan-path> [auto] [--review-plan] [--adapter ce|native] [--goal <text>] [--recipe <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>] [<unit>]"

Theory + edge cases: docs/contracts/driver-reference.md covers prepare/execute, goal binding, tick/seam/fan-out, exit reasons, and conversation-context (§11) — load only when not covered inline.