gilgamesh-gauge

Gauge Skill — Gilgamesh (Phase G intake)

When to use

Load during Phase G (Gauge). Runs once per mission, immediately after skills/verify-incoming/SKILL.md passes. Triage cost ≈ 1 step. Do not load during Inventory, Lock, Grind, or Attest.

Gauge is the additive-proof gate: Gilgamesh only spends cycles on missions it can plausibly win and that fit no specialist. Every misfit bounces cheaply, keeping the fallthrough seat structurally non-negative to the orchestrator's session — and keeping it from becoming a dumping ground (the identity-drift risk, R-032).

---

Decision Tree (run in order — first failure exits)

Step 1 — Validate the mission-contract

Validate the inbound payload against schemas/mission-contract.v1.json. It MUST carry {objective, scope(paths,mode), deliverables, evidence_required, stop_conditions, authority(read/write/exec/network/secrets/deploy)}.

• Valid → continue to Step 2.
• Malformed / missing fields → REFUSE{SCHEMA_INVALID}.

Step 2 — Specialist-fit check (keep-or-kick at generalist scale)

Does the mission map cleanly to a roster specialist? Gilgamesh never outranks a specialist and never enters Step-1 of dispatch (R-019) — it lives only in the fallthrough branch.

If the mission is…	REFUSE and suggest
a read-only discovery / scout	atlas
a plan / spec authoring campaign	ramza
a loop-native coding campaign (implement feature end-to-end)	vivi
a localized ≤2-file verifier-backed micro-edit	kupo
a root-cause / defect hunt	vigil
a decision record / trade-off adjudication	forge
a document / long-form artefact	idg

• Clean specialist fit → REFUSE{SPECIALIST_FIT, suggested_specialist: <name>}.
• Underspecified / ambiguous target → REFUSE{UNDERSPECIFIED} (the orchestrator emits a clarification_request).
• Actionable and fits no specialist → continue to Step 3.

Step 3 — Instantiate the capability-authority table

Instantiate schemas/capability-authority.v1.json from the contract's authority grant. Six rows × two columns (default / escalation):

Row	Default	Escalation ceiling
read	scoped repo + tool output	broaden scope.paths (approved)
write	sandbox only	none — PROPOSE-only is constitutional
exec	named oracle commands in sandbox	more oracle commands (approved, sandboxed)
network	none	proxied read via broker (approved)
secrets	none	broker-issued, single-use, mission-scoped
deploy	none	never-grantable

• If the mission requires authority above the ceiling — most importantly deploy, or a secret the table cannot broker — REFUSE{OVER_AUTHORITY}.
• Otherwise freeze the instantiated table into TaskState.authority and KEEP.

---

Output format

Emit exactly one of:

KEEP{authority: <instantiated capability-authority table>}

REFUSE{code: "<SCHEMA_INVALID|SPECIALIST_FIT|UNDERSPECIFIED|OVER_AUTHORITY|NOT_ROUTER|NO_ORACLE>", suggested_specialist: "<name|null>", note: "<one sentence for the orchestrator>"}

On KEEP, control passes to Phase I (Inventory). No other state from Gauge enters Inventory except the frozen authority table; Gauge is a one-shot gate.

---

Notes

• Triage cost ≈ 1 step. Gauge must be cheap; do not invoke atlas-aci here — gather happens in Inventory.
• Structural, not verbal. KEEP iff the contract validates, fits no specialist, and stays within the grantable ceiling. "It looks doable" is not a KEEP predicate.
• Specialist-preferring by construction. When in doubt between KEEP and SPECIALIST_FIT, prefer the specialist — the orchestrator can always re-route back.
• Deploy is a wall, not a slope. No mission, no escalation, no operator override reachable from within Gilgamesh grants deploy. Route it upward if it is truly needed.
• REFUSE is cheap and correct. A fast REFUSE with a suggested specialist is more valuable than a slow, out-of-scope grind.

---

Gauge Skill — Gilgamesh Phase G intake, refusal, and authority instantiation