accelerated-mode
The loop runs one step at a time, each gated. On Claude Code there is an optional faster gear for the slow, repetitive parts: a dynamic workflow — a small program that fans work out across many short-lived agents at once instead of walking a list one item at a time.
This skill holds the rules that are identical across every surface that offers one. The commands hold only what differs: which script, what arguments, and how to fold the results back in.
The two gates — both must hold, every time
- Runtime gate. A
Workflowtool must be available. It exists only in Claude Code. On Codex, opencode, or any copy-paste chat surface there is noWorkflowtool. Do not emulate one with extraTasksubagents — that would break the 3-agent cap and produce an unbounded fan-out on a surface that was never designed for it. - Opt-in gate. The user asked for it this run — "verify with a workflow", a
workflowargument, or anultracoderequest. Never automatic. A single run can spawn many agents (up to 16 concurrent, 1000 over a run) at real token cost, so it is always a deliberate escalation the user chose.
If either gate is false, run the serial path. Not a degraded version of it — the same path that has always been the default.
Fallback is the design, not a safety net
The serial path is the portable one: it works on every surface the loop runs on. The accelerated path can only ever be a speed-up on one of them. So every accelerated step is written to fall through cleanly, and nothing in the loop may depend on a workflow having run.
Detection is automatic: no Workflow tool means not Claude Code, which means serial.
Guardrails
- Never fan out a gate, and never fan out a build. The PLAN→BUILD approval, the verify verdict, and the deploy decision stay serial and human-facing. A workflow may read and evaluate inside an approved step; it never replaces the step's gate, and it never writes code.
- A workflow is advisory to the gate. Its output flows into the command's own report and marker. It does not deploy, does not check off
[deployed]rows, and does not decide anything the serial path wouldn't. - Every row comes back. The scripts return one result per item they were given — an item that could not be evaluated returns
failwithunverified: true, never a missing entry. Report those as failures and say how many. A missing row reads as a row that was never on the list, which is how a silent drop becomes a false pass. - Take the stricter verdict. Where a workflow result and your own differ, take the stricter one unless you can refute it with concrete evidence. Same reconciliation rule as the independent-evaluator step.
- The 3-agent cap is untouched. The cap governs
Tasksubagents. Workflow mode is a separate, named escalation — not a loosening of the cap, and not a licence to spawn subagents outside one.
Recording the run
Every marker that can be produced by an accelerated step carries two optional, additive fields at version: 1:
json"mode": "serial" | "workflow", "agents": <int>
Set mode: "workflow" and the agent count only when a workflow actually ran; otherwise mode: "serial" and agents: 0. Consumers that predate these fields ignore them — the serial path is byte-for-byte the old behavior.
Where the scripts live
Prefer the project copy at .claude/workflows/<name>.mjs; fall back to ${CLAUDE_PLUGIN_ROOT}/workflows/<name>.mjs. /bencium-init copies them into a project on Claude Code as a convenience — a project without them still gets workflow mode through the plugin-root fallback.
| Surface | Script | Shape |
|---|---|---|
/bencium-verify Step 3c | verify-acceptance.mjs | evaluate each local row, then a fresh skeptic tries to refute each PASS |
/bencium-deploy Step 5.5 | smoke-deployed.mjs | one agent per [deployed] row, real HTTP against the live URL |
/bencium-init Step 2 | brownfield-scan.mjs | multi-lens read-only sweep, then one synthesis |
Mechanical stages default to Haiku; adversarial verification and synthesis to Sonnet.
Every accelerated path is read-only
All three read, evaluate, and report. None of them writes code, and none of them is on the build path. That is the boundary, and it is deliberate.
A parallel build fan-out existed briefly and has since been removed. It bought throughput on the one part of the loop that was never the bottleneck — the bottleneck is the human reading the plan — and being the only fan-out that wrote code, it was where most of the bugs lived. If a task is really N independent mechanical edits, split it at the PLAN gate and build the parts one at a time. Slower in wall-clock, and the wall-clock was never the constraint.
