Instruction file imported from softwaresalt/autoharness (.github/instructions/escalation-protocol.instructions.md). Copyright stays with the author.

Escalation Protocol Instructions

This instruction is installed whenever _stage.agent.md or _ship.agent.md is present in the workspace (either-agent condition — distinct from the two-agent-only role-enforcement.instructions.md gate). It defines the escalation-payload contract that a halting agent compiles when a consecutive-failure threshold is crossed, and the single canonical definition of ESCALATION_DEGRADED referenced by both agent templates so the term is defined once and never drifts between them.

Status: Protocol Contract Active Now; Runtime Telemetry Substrate External (P-013.6)

This document defines the auto-escalation protocol contract — the payload shape, the routing resolution rule, and the degraded-state fallback. The agent-directed steps are active now: when an installed _stage.agent.md or _ship.agent.md crosses its own already-existing, agent-observed consecutive-failure/iteration threshold (each pipeline agent's own Stop Conditions table — no new runtime component is required for the agent itself to follow this directive), the halting agent compiles the escalation payload, resolves the escalation route, applies the same-route guard, hands the payload off for analysis, and halts without executing the failed operation again. This is a real, present-tense behavior change to each agent's own stop-condition handling, not a future one.

What remains external and dormant (routed OUT, per the 106-F plan) is narrower than "the whole protocol": specifically, (a) a standing, independent runtime telemetry event emitter/sink/queryable store that autonomously records events outside the acting agent's own reasoning loop, and (b) an automated, non-agent threshold-evaluation engine that would watch that telemetry substrate and fire escalation without any acting agent participating in the decision. Until that substrate exists, evidence_path / artifact_refs in the payload contract below are best-effort references (or absent, when telemetry is disabled) rather than guaranteed pointers into a live event store; the protocol does not depend on that substrate to function today. Operators must not be misled into believing a machine independently monitors and fires this protocol — it is always the acting agent, following its own already-tracked failure counters, that compiles the payload and resolves the route.

Authority-Preservation Invariant (NON-NEGOTIABLE)

Escalation under this protocol is a reasoning escalation — a request for analysis of the failed unit of work by a stronger/deeper reasoning model — and is never an authority escalation. Compiling an escalation payload, resolving an escalation route, or handing off to engram MUST NOT, by itself:

• authorize a shipment claim or task claim,
• authorize a merge or admin-fallback merge,
• authorize a source, template, schema, or backlog mutation beyond what the invoking agent's own Role Boundary already permits,
• bypass P-001 (release-unit sequencing), P-009 (merge-commit-only), P-014 (local review readiness), P-017 (dark factory contract), or P-020 (post-merge compaction).

The terminal state of an escalation is a halt + engram handoff for operator/asynchronous review. The handoff is for asynchronous or operator review, not a fourth attempt. It is never a silent, unsupervised expansion of what the halting agent is allowed to do.

Escalation-Payload Contract

When a halting agent crosses its consecutive-failure threshold (see the invoking agent's own Stop Conditions table), it MUST compile an escalation payload containing the following fields before handing off and halting:

Escalation Route Resolution (P-013.6 / F02FD596 nested per-role hierarchy)

The escalation route resolves via a role-scoped precedence:

1. Nested per-role override: model_routing.<role>.escalation (stage.escalation or ship.escalation, matching the acting agent's own role). When declared, the acting role resolves its escalation route from this nested block. A nested override that declares only some fields falls back per field to model_routing.tier3 for the missing fields — never to the legacy flat route below; an explicit nested override never silently defers to the shared legacy route.
2. Legacy flat fallback (DEPRECATED): when the acting role has no nested escalation override, model_routing.escalation (the pre-F02FD596 flat, role-agnostic route) resolves instead, with a deprecation notice. This workspace declares no nested stage.escalation or ship.escalation override, so this is this workspace's currently-resolved escalation route: model_family = gpt-5.6-sol, model_provider = openai, reasoning_effort = high — a distinct vendor/family from the Stage role route (claude-opus-5/anthropic), so the handoff can provide genuinely independent analysis rather than a same-route no-op.
3. Tier3 fallback: any field still unresolved after (1)/(2) falls back per-field to model_routing.tier3 (claude-opus-5; this workspace's tier3 is a plain model-identifier string, so its provider/reasoning- effort sub-fields resolve empty via fallback).

This mirrors the P-013.5 stage/ship role-route fallback pattern: a targeted override, never a parallel tier taxonomy.

Both-present fail-closed (H2): the legacy flat model_routing.escalation and any nested <role>.escalation (for either role) MUST NOT both declare a non-empty field. When both are present, resolution is AMBIGUOUS and MUST fail closed — a schema-level constraint rejects this shape where expressible, and the loader/verification layer enforces the same rule as a backstop. Never auto-pick a winner; migrate fully to nested per-role escalation, or remove the nested override(s), to resolve the ambiguity.

ESCALATION_DEGRADED (canonical definition — defined once, referenced elsewhere)

ESCALATION_DEGRADED is declared when any of the following hold:

1. Route unavailable. The resolved escalation route cannot be dispatched by the runtime (the model/provider is not available in this environment).
2. Engram unavailable. The terminal engram handoff surface cannot be reached (no MCP tool, no file-based fallback capable of receiving the handoff).
3. Same-route no-op (role-scoped, H3). The fully-resolved escalation tuple (model_family, model_provider, reasoning_effort) for the ACTING role's own effective escalation route (nested <role>.escalation if declared, else legacy flat, else tier3 fallback) equals that SAME acting role's own already-resolved role route tuple (P-013.5) for this session — compared only against the acting role's own route, never a different role's route. For example, an agent whose role route is already pinned to tier3 receiving an unset escalation route that also falls back to tier3 would "escalate" to an identical model, which is not a real escalation.

When ESCALATION_DEGRADED is declared, the halting agent MUST fall back to its existing operator-halt path (the Stop Conditions escalation prompt already defined in its own agent template) rather than proceeding as though escalation succeeded. _stage.agent.md and _ship.agent.md both reference this single definition rather than re-declaring it, preventing drift between the two templates.

Terminal Engram Handoff

When the resolved route is available and not degraded, the halting agent records the route in the payload's resolved_escalation_route field, hands the compiled payload to engram for asynchronous or operator review, and halts. It MUST NOT re-execute the failing operation after its circuit is open. The handoff is for asynchronous or operator review, not a fourth attempt. This handoff is informational/diagnostic — it carries no authority of its own (see the Authority-Preservation Invariant above).

Relationship to Other Policies

• P-013 / P-013.5: This protocol extends the existing tier-escalation and invocation-time routing enforcement machinery; it does not replace either.
• P-001 / P-009 / P-014 / P-017 / P-020: Explicitly preserved — see the Authority-Preservation Invariant.
• P-006 (Plan Hardening): Same-route degradation and dormant-until-runtime status are documented failure modes for any plan extending this protocol.