esat-fleet

name

esat-fleet

description

esat + a fleet council. Runs the full esat tri-model panel (Claude critic + Codex + Grok), then ALSO fans the same review across the OpenRouter OSS swarm via fleet-fuse (sensitivity-tiered, redacted, off the first-party trio). Claude remains the sole chairman — it reconciles every voice into consensus / divergence / unique-catch. Use for the highest-stakes analysis where you want frontier + open-model breadth. Triggers: /esat-fleet, esatf, eskill-analyze-fleet, eskill analyze fleet, fleet council review.

eSkill: Analyze Fleet (esat-fleet) — Frontier + OSS-Swarm Council

esat-fleet = esat + a fleet council leg. It runs everything esat does (full world-class analysis engine + the 3-frontier-model panel), then broadens the council with the OpenRouter OSS swarm dispatched through fleet-fuse — tier-gated, redacted, and off the first-party trio so it adds breadth rather than duplicating it. You (Claude) stay the sole chairman: you reconcile frontier + OSS voices into one verdict and decide.

Use esat-fleet when you want both frontier depth (critic + Codex + Grok with actual/account-default disclosure) and open-model breadth (DeepSeek / Qwen / Kimi / GLM via OpenRouter) cross-checking the same item — the widest fixed-council in the stack below the configurable frontier tier. For frontier-only, use /esat; for a single-model pass, /esa.

Shared routing contract (load first)

Before Phase 9 dispatch, read and obey:

${CLAUDE_SKILL_DIR}/../eskill-common/references/model-routing.md

That contract defines semantic role vs requested vs actual route, preference precedence, harness identity, strict pins, disclosed automatic fallback, strongest-restriction policy, metered consent, two-stage manifest, two-phase deduplication, and draft ownership. This skill owns Phase-9 composition only.

Engine + trio — reuse esat with caller-owned draft

Run esat for the engine and trio reviewers, with one ownership change:

1. ${CLAUDE_SKILL_DIR}/../esat/SKILL.md — which reuses the eskill-analyze engine (Input Params, Triage, Delegation, mental models, Phase 0 + Steps 1–8 + Phase P Premortem (when activated), output template) and defines the fixed trio panel.
2. ${CLAUDE_SKILL_DIR}/../esat/references/trio-panel.md — Phase 9 reviewers 1–3. Use the caller-owned draft lifecycle: Tier 3 supplies $DRAFT, defines $PEER_DISPATCH_INPUT separately from raw $DRAFT, the trio must not delete the draft, and Tier 3 cleans once after all consumers finish.

esat-fleet changes Phase 9 as follows:

1. Create and retain one caller-owned draft for trio and fleet; arm an owner EXIT trap immediately after temp creation (also covers $PEER_REDACTED and fleet $REVIEW).
2. Define PEER_DISPATCH_INPUT separately from raw $DRAFT for external peer egress (see sensitivity rules below and fleet-leg.md).
3. Freeze the two-stage route manifest (immutable plan + append-only results) before dispatch.
4. Emit an effective-route preview (consent, budget, sensitivity, readiness).
5. Run the trio with caller-owned lifecycle: peer trio reads $PEER_DISPATCH_INPUT only — never raw $DRAFT (no trio-side cleanup of $DRAFT).
6. Run the fleet leg per ${CLAUDE_SKILL_DIR}/references/fleet-leg.md when policy allows.
7. Append exactly one terminal result per planned lane (including blocked/skipped with observed_route: null).
8. Derive panel status only after all terminal results exist.
9. Clean owner temps (exactly once) after trio and fleet success, skip, failure, or timeout — normal path invokes cleanup and disarms EXIT; timeout still hits the trap. SIGKILL cannot be handled.

Phase 9 — add the fleet leg

After the policy-gated reviewer legs against the caller-owned draft, also run the fleet leg per ${CLAUDE_SKILL_DIR}/references/fleet-leg.md when readiness is ready, then synthesize all voices together.

On high sensitivity, schedule zero external Phase-9 routes (not only FleetFuse): run only local/first-party reviewers (normally the in-harness critic). Skip/block peer trio Codex/Grok unless the harness explicitly classifies those routes as first-party. Append terminal results with observed_route: null for every blocked external lane. Do not claim the external frontier trio ran. On medium/low (when policy allows), the critic + peer trio (or first-party equivalents) may run, then the fleet leg when ready.

Sensitivity gate (read before running)

Phase 9 is sensitivity-tiered for every external route — FleetFuse OSS and external peer Codex/Grok — not only the fleet leg:

ESAT_FLEET_SENSITIVITY	Phase-9 external-route behavior
high (sensitive/proprietary)	Zero external routes. Local/first-party reviewers only (normally the in-harness critic). Fleet leg blocked. External peer readiness forced 0. External peer trio / peer codex / peer grok skipped/blocked unless the harness explicitly classifies them first-party. Terminal results: observed_route: null, reason high-sensitivity. Do not degrade to a full external esat frontier trio and do not claim that trio ran.
medium (default)	External peer + OpenRouter OSS only with an identified redaction path. Peer: create $PEER_REDACTED via trusted $IDENTIFIED_REDACTION_PATH and set $PEER_DISPATCH_INPUT to that file; peer trio reads $PEER_DISPATCH_INPUT only — never raw $DRAFT. Fleet: fleet_scrub fail-closed. Missing/failing redactor → readiness blocked redactor-unavailable, do not call peer / do not raw-send.
low (clearly non-sensitive)	+ jimmy / low-tier pools when permitted. Raw $DRAFT as $PEER_DISPATCH_INPUT only if the already-resolved external peer readiness flag is 1. Low never bypasses readiness.
invalid	Fail closed — block external routes with invalid-sensitivity.

Default is medium. Outbound redaction is automatic and fail-closed; never pass --no-redact. State the tier you ran in the panel. Multiple policy scopes intersect with strongest restriction wins.

Metered consent and budget

• Metered FleetFuse requires per-run consent: direct current-invocation instruction or interactive current-run answer only.
• Environment variables, project/user config, inherited shell state, expired consent, and prior-run answers are rejected as consent.
• Without accepted consent, skip/block the fleet lane and do not enter the --yes-metered branch.
• --yes-metered appears only in the consented command branch.
• ESAT_FLEET_BUDGET_USD: optional positive numeric cap. Absent → disclose provider/account cap only. Invalid → fail closed (invalid-budget).

Output

Use esat's output (eskill-analyze template + the ### Tri-Model Panel block), extended to a council per references/fleet-leg.md — the panel lists the frontier trio plus the OSS-swarm verdict, records the sensitivity tier, consent/budget disclosure, actual routes, and whether the fleet leg ran. Save to .omc/plans/esat-fleet-{date}-{focus}.md.

Switches

• ESAT_FLEET=0 → skip the fleet leg (degrade to esat); still emit a terminal skip result and clean the draft once.
• ESKILL_PEER=0 → skip the Codex+Grok trio leg (inherited from esat). Both off → Claude-only + optional fleet if policy allows.
• ESAT_FLEET_SENSITIVITY → high|medium|low (default medium).
• Cost: the fleet leg may add metered OpenRouter spend only after consent; Codex+Grok remain off the Claude reserve when using peer.