Add BEM pre-step gate to fleet dispatcher for multi-issue parallel dispatch

[Trecek]

Summary

Add a BEM (build-execution-map) conflict gate to the fleet dispatcher's multi-issue dispatch path, and raise the fleet semaphore's default concurrency to allow the gate to actually enable parallel execution. The work spans four areas: (1) replace stale serial-only guidance in the fleet dispatcher prompt with BEM gate instructions, (2) fix inaccurate semaphore language in the campaign prompt, (3) raise max_concurrent_dispatches default from 1 to 3 and add a max_total_issues: 12 cap to FleetConfig, and (4) add contract and config tests that enforce these prompt and config invariants going forward.

All machinery already exists: bem-wrapper is a food-truck recipe, dispatch_food_truck returns l3_payload.dispatch_plan inline, and FleetSemaphore already supports N>1. The gap is entirely in prompts and config defaults.

Requirements

REQ-BEM-001: Multi-issue conflict gate section in fleet dispatcher prompt

Add a new prompt section to _build_fleet_dispatch_prompt() in src/autoskillit/cli/_prompts_kitchen.py that instructs the fleet dispatcher to:

1. Detect when the user requests 2+ issues dispatched (regardless of whether user says "parallel" — any multi-issue dispatch needs the gate)
2. Dispatch bem-wrapper first: dispatch_food_truck(recipe="bem-wrapper", task="Build execution map for conflict analysis", ingredients={"issue_urls": "<comma-separated URLs>", "base_branch": "<target branch>"}, capture={"execution_map": "${{ result.execution_map }}"})
3. Read l3_payload.dispatch_plan from the dispatch_food_truck response — this is a JSON array of {"group": N, "parallel": bool, "issues": "..."} objects
4. Sequence subsequent dispatches per the group plan:
   • parallel: true groups → dispatch all issues in the group simultaneously
   • parallel: false groups → dispatch sequentially within the group
   • Wait for all dispatches in group N to complete before starting group N+1
5. Fall back to sequential dispatch (one at a time) if bem-wrapper fails
6. Single-issue dispatches skip BEM entirely (no pre-step needed)

REQ-BEM-002: Update stale serial dispatch guidance

Replace the stale text at _prompts_kitchen.py:121-122:

- Serial execution: dispatch one food truck at a time. fleet_lock enforces this —
do NOT attempt parallel dispatches.

With guidance describing BEM-gated parallel dispatch. The new text should describe: single-issue dispatches proceed directly; multi-issue dispatches require the BEM pre-step; group ordering and merge-wait between groups; max concurrent dispatches is 3. Reference the existing max_parallel cap that BEM already enforces on group sizes.

REQ-BEM-003: Contract test for fleet dispatcher BEM gate

Add a contract test (in tests/contracts/) asserting that the fleet dispatcher prompt returned by _build_fleet_dispatch_prompt() contains multi-issue conflict gate instructions. Similar pattern to test_sous_chef_routing.py:176 which asserts sous-chef contains BEM invocation instructions.

Suggested assertions:

• Prompt contains "bem-wrapper" reference
• Prompt contains "dispatch_plan" reference
• Prompt contains "multi-issue" or "conflict" gate language
• Prompt does NOT contain the stale "Serial execution: dispatch one food truck at a time" text

REQ-BEM-004: Include bem-wrapper in fleet dispatcher recipe table

If _food_truck_section (the recipe table injected into the fleet dispatcher prompt) does not already include bem-wrapper, ensure it appears so the dispatcher knows it's available for dispatch without calling list_recipes first.

REQ-SEM-001: Raise max_concurrent_dispatches default to 3

Update the default from 1 to 3 in both src/autoskillit/config/_config_dataclasses.py (FleetConfig.max_concurrent_dispatches) and src/autoskillit/config/defaults.yaml (fleet.max_concurrent_dispatches). This allows the semaphore to permit up to 3 food trucks running simultaneously, which BEM groups can utilize when parallel: true. Users can still override via .autoskillit/config.yaml or the AUTOSKILLIT_FLEET__MAX_CONCURRENT_DISPATCHES env var. Update FleetConfig.validate() to enforce 1 <= max_concurrent_dispatches <= 3 (hard cap to prevent runaway subprocess spawning).

Note: FleetSemaphore itself needs no changes — it already supports N>1 via asyncio.BoundedSemaphore(max_concurrent).

REQ-SEM-002: Add max_total_issues fleet config cap (default 12)

Add a new max_total_issues: int = 12 field to FleetConfig in _config_dataclasses.py and defaults.yaml. This caps the total number of issues a single fleet dispatch session can process (across all food trucks and all groups). The fleet dispatcher prompt should instruct the dispatcher to count total issues before dispatching bem-wrapper and refuse (or warn) if over the cap. User-overridable via config/env var. Validation: 1 <= max_total_issues.

Note: This is distinct from the existing max_parallel (default 6), which caps issues per BEM group, not total issues across the session. max_parallel is a BEM skill argument; max_total_issues is a fleet-level config field. Both are needed: max_parallel controls group sizing, max_total_issues controls session-level scope.

Rationale for 12: with max_parallel=6 (issues per group) and max_concurrent_dispatches=3, 12 issues is ~2 fully-loaded parallel groups — the upper bound of what the system should handle in a single dispatch session.

REQ-SEM-003: Fix inaccurate "calls queue" campaign prompt claim

Replace the inaccurate text at _prompts_campaign.py:53:

The fleet semaphore gates actual concurrency; calls queue when the semaphore is saturated.

With accurate language reflecting the actual at_capacity() → FLEET_PARALLEL_REFUSED fast-fail behavior. The dispatcher must wait for a running dispatch to complete before retrying, not assume queuing. Also update _prompts_campaign.py:231 which says "max_concurrent=1 for static dispatches" — this will be inaccurate once REQ-SEM-001 raises the default.

REQ-SEM-004: Update tests for new defaults

• Update any tests that hardcode max_concurrent_dispatches=1 as the expected default (search for max_concurrent_dispatches and max_concurrent=1 in tests/)
• Add validation test for the new max_total_issues field and the max_concurrent_dispatches <= 3 upper bound
• Update test_fleet_e2e.py tests that assert FLEET_PARALLEL_REFUSED at max=1 to account for the new default of 3
• Add contract test asserting the campaign prompt does NOT contain "calls queue" language

Conflict Resolution Decisions

The following files had merge conflicts that were automatically resolved.

(None provided)

Closes #2182

Implementation Plan

Plan file: /home/talon/projects/autoskillit-runs/impl-20260507-160445-669401/.autoskillit/temp/make-plan/fleet_dispatcher_bem_gate_plan_2026-05-07_160822.md

🤖 Generated with Claude Code via AutoSkillit

Token Usage Summary

Step	Model	count	uncached	output	cache_read	peak_ctx	turns	cache_write	time
plan	claude-sonnet-4-6	1	123	13.2k	495.3k	52.1k	89	46.7k	5m 41s
verify	claude-sonnet-4-6	1	236	14.8k	1.5M	68.7k	83	56.0k	4m 8s
implement*	MiniMax-M2.7-highspeed	1	1.2M	15.0k	1.3M	74.0k	121	119.9k	5m 13s
prepare_pr*	MiniMax-M2.7-highspeed	1	87.9k	4.9k	198.4k	28.8k	22	15.2k	1m 27s
compose_pr*	MiniMax-M2.7-highspeed	1	46.4k	2.9k	169.6k	28.8k	14	15.1k	55s
review_pr	claude-sonnet-4-6	1	124	36.3k	755.8k	79.4k	45	67.4k	7m 7s
resolve_review	claude-sonnet-4-6	1	237	14.4k	1.4M	65.2k	70	52.9k	7m 44s
Total			1.3M	101.5k	5.9M	79.4k		373.2k	32m 17s

* Step used a non-Anthropic provider; caching behavior may differ.

Token Efficiency

Step	LoC Changed	cache_read/LoC	cache_write/LoC	output/LoC
plan	0	—	—	—
verify	0	—	—	—
implement	223	5925.8	537.8	67.3
prepare_pr	0	—	—	—
compose_pr	0	—	—	—
review_pr	0	—	—	—
resolve_review	16	89758.8	3304.4	901.1
Total	239	24494.1	1561.4	424.8

Model Usage Breakdown

Model	steps	uncached	output	cache_read	cache_write	time
claude-sonnet-4-6	3	605	40.2k	3.2M	196.0k	21m 58s
MiniMax-M2.7-highspeed	3	1.3M	22.8k	1.7M	150.2k	7m 35s