| name | deep-loop |
| description | Operator command to drill ONE topic to primary sources with current data — iterative multi-agent research that deepens each round until the topic saturates, delivered as a founder-legible decision over a fully cited report; watch mode re-runs on an interval and pings only on material change. |
| argument-hint | <topic> [depth: quick|standard|deep] [rounds: N] [fresh-since: 30d|6mo|DATE] [watch: 6h|daily|weekly] [perspectives: "a, b, c"] [saturation-k: N] [save: vault|here|off] [budget: $N] |
| disable-model-invocation | true |
| allowed-tools | Read, Grep, Glob, Bash, Agent, Task, Workflow, WebSearch, WebFetch, Skill, ToolSearch, mcp__claude_ai_Exa__web_search_exa, mcp__claude_ai_Exa__web_fetch_exa, mcp__margot__deep_research |
deep-loop — the one-topic research-depth engine
INVOCATION — read this first. disable-model-invocation: true means calling this skill
through the Skill tool errors by design (it stops the sub-agent fleet auto-firing an
expensive swarm). When the operator types /deep-loop <topic>, Read this file and run the
gates below inline — never Skill(deep-loop). Mirrors /nexus.
What this is / is not
The estate's single-topic depth engine: Exa-first freshness + honest Mixture-of-Agents
(arXiv 2406.04692) + a swarm loop that re-targets each round at the previous round's open gaps.
It inherits nexus G1/G3/G6/G7 by pointer to
~/.claude/skills/nexus/references/orchestration-playbook.md — zero duplication — and is itself
the depth engine /nexus may dispatch at its G3/G4. It is NOT a quick lookup (D0 downshifts
those), NOT storm's neutral long-form article (route out, don't nest), NOT the built-in /loop
interval runner (watch-mode composes with it), and NOT the lighter built-in deep-research
single-pass harness — use deep-loop when the topic needs iterative deepening and current data.
Invocation and flags (English flags, smart defaults)
| Flag | Values | Default |
|---|
depth: | quick (3 proposers) / standard (4) / deep (6, K=2, 5 rounds; Opus aggregators only with budget:) | standard |
rounds: | N (hard cap) | 3 (deep: 5) |
fresh-since: | 30d / 6mo / ISO date | 12 months |
watch: | 6h / daily / weekly — recurring re-run via built-in /loop | off |
perspectives: | "a, b, c" — overrides auto-discovery | auto (3–6) |
saturation-k: | consecutive dry rounds before stop | 1 (deep: 2) |
save: | vault / here / off | vault |
budget: | $N — raises TAO_MAX_COST_USD for this run only | ceiling as set |
An underspecified topic gets 2–3 tight clarifying questions before anything fires.
Live progress contract
One glanceable line per event, so the operator can act on the log alone:
R{n}.L{layer} {role}×{count} [{lane}] … {new}/{total} claims verified · ${spend}
Plus one line per stop-condition trigger naming the guard that fired.
The gates
D0 — Appetite (always). Classify with the nexus G1 classifier (by pointer). A
small/well-specified question → answer directly or run a single Exa pass and STOP — the
apparatus must not fire for a narrow ask.
- Completion: appetite named; the null case honoured.
D1 — Frame & mine (always). Ground in the vault before reaching outward:
node "$HOME/2nd Brain/2nd Brain/_system/brain.js" find "<topic>" (nexus-recall; self-heal the
index on a miss). Restate the topic as an outcome. Discover 3–6 research perspectives from the
menu in references/moa-swarm-playbook.md; the
perspectives: flag overrides.
- Completion: perspectives named; prior context mined; recall hit/miss recorded.
D2 — MoA round. Each round: L1 proposers (Sonnet, one perspective or one gap each, own
retrieval down the Exa ladder) → L2 aggregators ×2 (each receives ALL proposer outputs:
A1 corroborate/reconcile, A2 divergence/gap-miner; each must emit a delta note) → L3 synthesist
(Opus, the only seat that writes claims to the corpus; emits the gap ledger + running Executive
Read + materiality verdict). Full layer spec, perspective menu, and gap-ledger schema are in
references/moa-swarm-playbook.md; read it before opening D2.
- Completion: ≥1 round ran; every load-bearing claim tiered + corroborated or tagged unverified.
D3 — Swarm loop. Round 1 is breadth (perspectives). Round N (≥2) is depth: rank the previous
round's gap ledger by load-bearing-ness and seed a FRESH proposer swarm one-per-gap (floor 2);
the corpus carries forward read-only. Gaps are the literal task list of the next round — depth
compounds, breadth is never re-run. An empty ledger stops the loop.
- Completion: each round after the first traces to named gaps.
D4 — Exa lane + integrity. Exa is the primary freshness lane behind a mandatory, never-silent
ladder: Exa → WebSearch (whitelist-seeded) → WebFetch — per
references/exa-and-integrity.md (includes the ToolSearch
preload step for the deferred Exa tools). The research-integrity bar is nexus G3 verbatim,
by pointer: credibility ladder, corroboration bar, gap-mining, honest degradation.
- Completion: no naked single-source load-bearing claim; the lane that actually ran is reported.
D5 — Stop conditions. Five guards — HARD_STOP, TAO_MAX_COST_USD (pre-round projection),
round cap, per-round agent caps, saturation (K consecutive dry rounds; dual-condition test +
cheap pre-filter) — specified in
references/moa-swarm-playbook.md. Caps always beat
saturation. Partials persist to disk after every round.
- Completion: the stop reason is recorded in the report footer.
D6 — Adversarial verify. Route the final corpus through opus-adversary (nexus G6, by
pointer): flip-test each load-bearing claim; a claim that flips is downgraded to
"assumption (unverified)" before it can reach the Executive Read.
- Completion: load-bearing claims survived an independent adversary, or were downgraded.
D7 — Deliver. Open with the nexus G7 Executive Read (by pointer — ≤120 words, plain
language, register ban-list). Below --- detail ---: the cited report with {source URL, tier}
per claim, the unverified list, the defended divergence (or "none found"), the REMAINING gap
ledger, and the method footer (rounds, stop reason, active lane per round, spend). save:vault
writes the sanctioned Outcomes note (~/2nd Brain/2nd Brain/Outcomes/YYYY-MM-DD-<slug>.md) +
OKF regen via the generator; watch: routes 6h→built-in /loop, daily/weekly→schedule, with
edge-triggered pings. Details in
references/output-and-watch.md.
- Completion: decision-first top; cited detail below; partials labelled at the top.
Guards (mandatory whenever D2 opens)
HARD_STOP, cost projection, round/agent caps, structural recursion depth-1, and the model
ladder are specified ONCE in references/moa-swarm-playbook.md
§Guards — the ORCHESTRATOR checks them between rounds. Degradation is stated, never silent.
Mechanism
The orchestrator owns the swarm loop, guards, and per-round persistence; each MoA round executes
as ONE Workflow call (or batched parallel Agent calls where Workflow is unavailable). The
canonical orchestrator loop + round script live in
references/workflow-script.md — never inline them here.
