Rough prompt
Your starting sentence. Messy is fine. This is the raw board.
You bring a rough idea.
Handover turns it into a plan you approve, then does the work and proves it.
Think flat-pack furniture. No building until you have seen all the parts, checked the diagram, and tightened the two big approval screws.
Assembly steps.
Do them in order.
Approval screws per run.
Stage 1 + Stage 2. Chained work shares one pair for the whole chain, not one pair per link.
Gaps left unowned.
Every gap gets Fixed, Queued, or Accepted.
Handover can plan on its own. It cannot start gated work until you approve Stage 1 and, where needed, Stage 2. An unlisted action pauses the build.
The person who builds a piece never grades it. A fresh verifier checks it. A fresh judge gives the final verdict. That separation never bends.
V6 is relentless by default. A first fail starts a fix cycle, not a dead-end report. Lowering the bar to “pass” is forbidden.
Your starting sentence. Messy is fine. This is the raw board.
What Handover thinks you mean, with assumptions marked clearly.
The exact list of tools, targets, limits and budget that were approved.
Owns the plan and the record. Calls specialists. Does not do all the work itself.
Builds each unit. Works in a fresh context with only its task.
Fresh eyes. Gets the spec + the artifact, never the builder's sales pitch.
Fresh context at the end. Resolves the verdict. Cannot turn a gap into a pass.
JSON list of every unit: name, output, owner, check, and passes: false → true.
One line per decision: when, what, why, evidence pointer, result. Append-only.
Every spawn is a reservation. At least 15% is held back for verification.
Map of units and links. Shows what runs apart, what waits, and where they join.
Reusable playbook for a field: router, playbooks, checks, cleanup. Proven or draft.
Reads the registry, picks the skill whose trigger matches, writes one call line before each call.
16 engines plus the umbrella. Entry 1, Research 6, Proof 3, Stuck 3, Search 4. Core 4 loop: Anomaly, Abduction, Falsification, First Principles underneath. Invention is the outcome, not a 17th skill.
Tools you provide
Handover only asks you human questions: preference, authority, risk, a non-negotiable, or a trade-off with no rule. Everything else it settles by reading, inspecting, or testing.
Classify the failure first. Then use a distinct approach: different input, source, method, or smaller size. Same wording ≠ distinct.
Up to 5 fix cycles per run. Dead end only after 3 distinct approaches for the same failure class fail, or budget is spent.
One output, one owner, one check. Starts immediately because it has no link coming in. Its own loop handles its own fails without stopping the other branches.
Done = every unit in the list has passes: true. No unit removed, no check edited, no unit renamed to hide a failure.
The coordinator must read the registry before routing, pick a skill only when its trigger fires, write one call line before each call, and record each handoff in that skill's own format. Judge verdict PASS, Nov 15, 2026, run stage2-router-wire-2026-11-15-01, all 5 items pass.
Anomaly Hunting and Abduction routed with 2 call lines and 7 of 7 handoff fields each, Nov 15, 2026. Top anomaly: 1 usable case against a predicted 40. Top explanation: the pool was consumed.
Still in the box from V5
16 engines plus the umbrella. Each skill passed a sealed test with 2 fresh blind trials and a fresh judge, Nov 12 to 14, 2026.
Registry plus rules. Reads one row per skill, picks by trigger, writes one call line before each call.
Both fresh blind trials passed both test situations, 5 of 5 repeats each, $0 of the $4 cap, Nov 14, 2026.
Stated in the skill header. Larger envelopes are approved case by case.
Up to 5 fix cycles per run, shared across units. Router and umbrella runs use 3.
Still in the box from V4
The graph now ships with a machine-readable list. Only the passes field ever changes.
Every session: read notes + logs → pick next unfinished unit → run setup → end-to-end check → fix base if broken → then start.
Proof is on the real artifact, the way a human would use it. Not just a unit test or a page fetch.
Simple: 1 agent, 3–10 calls. Compare: 2–4 agents, 10–15 each. Complex: 10+ agents, split clearly. Stated in the package.
Fresh context, only needed tools, short summary back. Stopped at limit = partial, not failed.
When the trail gets long, summarise it: keep decisions, gaps, numbers. Notes live in a file, outside context.
Low / medium / high, by read vs write, undoability, permissions, cost. High-risk pauses or goes to a human. Listed in the manifest.
Same tool fails N times in a row (N set in the manifest) → stop retrying it, escalate to a human.
Grades during work, 0.0–1.0 on: factual accuracy, citation accuracy, completeness, source quality, tool efficiency. Judge still decides at the end.
Durable state, run/action IDs, lease + fencing token. Resume never re-runs a finished tool. Mismatch = fail closed, don't guess.
Skill-text spec only: the harness part was not built in the V4 apply run.
Harness code at 4 points: PreToolUse, Stop, SubagentStop, PreCompact. See diagram below.
Skill-text spec only: the harness part was not built in the V4 apply run.
Besides manager-calls-specialists, agents can hand control to the next agent in turn. Handoffs are edges, like tool calls.
Note on hooks + resume above: V4 wrote the specification into the skill text. The harness implementation for those parts was not built or claimed as done in the apply run.
PreToolUseBefore a tool call. Can block or change it.
SubagentStopWhen a subagent finishes. Gathers its result.
PreCompactBefore summarising. Archives the full transcript first.
StopBefore exit. Saves session state.
One coordinator in the middle.
Control passes hand to hand.
Graph holds for both: agents are nodes, handoffs are edges.
| Job size | Agents | Calls | Picture it as |
|---|---|---|---|
| Simple | 1 | 3–10 | One shelf. One person. |
| Comparing things | 2–4 | 10–15 each | A few shelves, side by side. |
| Complex | 10+ | Divided clearly | Whole room. Everyone owns one wall. |
Defaults only. Each field kit sets its own numbers for its field.
Fix fits inside the approved envelope: same tools, same destination, no new cost or outward action. Do it, test it, re-check it in the same run.
Fix needs new scope, cost, risk, or an outward action. Record it as the next job, with its own approval gate. Queued is not done, and is never shown as done.
You explicitly accepted the gap, with the reason recorded. Only explicit acceptance counts.
Done means only: all gaps Fixed, or all remaining gaps explicitly Accepted by you. Never “done, but…”
Open the real artifact like a user would. Check every link in the graph points to a real file. Check the per-unit list: is every passes true? Check the final report names its field kits, with proven / draft status and a link. If any answer is no, it is not done yet.
Scope note, printed on the box: V6 is supervised only. The research brief was the first execution pilot. Since then, supervised runs have proven 17 Discovery skills plus the umbrella and the harness router. V6 makes the Discovery router a required step in the skill text. The wiring proof routed 2 of 17 skills. The wire is proven for the Research phase engines, not for the Proof, Search, or Stuck phases. Other fields can be planned and packaged, but are not operational until a kit for that field passes its own real task. Unattended, live-data, and three-level fan-out are not in this box. The loader check never ran, because no Muse skill loader command exists in the environment where the proofs ran. The router proof covers two test situations, not live use of all 17 skills.