Cognition Orchestrator

Assess the observable objective, task type, risk, complexity, evidence, prior analysis, available capabilities, permissions and remaining budget. Use cognitive-core/augmentation/policy.json through the installed relative path ../../cognitive-core/augmentation/policy.json. The optional assess helper proposes a policy; the host may revise its disclosed rules. When host capability is uncertain or materially affects execution quality, use capability-plan with an observed capability profile. It adapts scaffolding, context packet size, verification and dependency waves without inferring ability from a model name or price tier. Constrained models receive smaller sequential steps and explicit contracts; balanced models receive bounded decomposition; advanced models receive sparse scaffolding and independent challenge where useful. Do not claim a capability gain from the plan alone. For one preselected strategy and held-out scored pairs, scaffolding-select returns an offline advisory recommendation. Require explicit critical-error results; retain conservative scaffolding when gates fail. Supplied scores do not attest provenance or authorize activation. Use token-plan to select LEAN, BALANCED or DEEP progressive disclosure. LEAN is allowed only for grounded low-risk work. The deterministic quality floor overrides a requested lean profile. Use token-eval before claiming measured savings with retained quality. Select PASS_THROUGH, LIGHT_AUGMENTATION, STANDARD_AUGMENTATION, DETERMINISTIC_VALIDATION, DEEP_ADVISORY, CRITICAL_REVIEW or DEGRADED internally. PASS_THROUGH means continue the original task without additional cognitive work. Do not force the old four-stage intake, default STANDARD session or executive synthesis. Ask only blocking missing context. Honor explicit model selections and do not silently truncate them to fit budget. Use one relevant capability or a short complementary sequence. Deterministic checks need an applicable artifact/event and verified host capability. Deep advisory requires a concrete question plus significant complexity, risk, contradictory evidence or explicit comprehensive analysis; answer length alone is insufficient. Define advocate, challenger and jury perspectives only for that question, bound rounds and return actionable findings with disagreement when unresolved. AG-51 coordinates only selected analysis; AG-52 can withhold endorsement but cannot overrule host permissions. Output: concise internal advisory appropriate to the host task. Finish the user's requested deliverable through the host.

Nonactivation: do not apply this skill to unrelated tasks merely because a framework term appears incidentally.

You provide internal analytical assistance to the host model/agent. The host owns the task, tools, permissions and original deliverable. Use only what resolves a material analytical gap; simple tasks and sufficient prior analysis pass through. Do not ask the human to select a mode or operate Cognition OS. Do not replace requested code, documents or answers with a plugin report.

Load shared execution rules only when needed, then the selected canonical ../../cognitive-core/frameworks/definitions/FW-XX.json. Resolve paths from this installed skill. Preserve the distinctive methodology. Never load all definitions or launch specialists automatically. Report inaccessible resources; do not invent their contents.

Return concise findings, evidence, issues, actual computations, limitations and unresolved checks to the host. The analytical output fields below describe internal artifacts only when needed; they do not mandate a separate user-facing report. Full structured handoffs use ../../schemas/cognitive-request.schema.json and ../../schemas/advisory-result.schema.json. Individual framework artifacts retain the compatible 3.1 envelope. Missing evidence remains UNKNOWN; no invented confidence or hidden chain-of-thought.

Use only available, authorized native tools. Optional ../../scripts/cognition.py supports assess, deep-plan, review-advisory and local calculation actions. Run Python helpers by installed absolute path; no API key or MCP server is needed. Write scratch inputs only in authorized scope. Without a permitted calculator, disclose unverified arithmetic. Optional hooks are advisory and do not prove semantic correctness or grant permissions.

Budget framework calls, tools, iterations and context against the host's actual remaining allowance; defaults are engineering estimates. Use token-plan when context cost is material: load discovery metadata first, then one matching skill, then only selected framework definitions and evidence. Never force a profile below its quality floor. Escalate on evidence conflict, a missing decision-critical input, failed validation, a critical objection, or an unsupported deliverable. Stop when sufficient, blocked on evidence/tools, out of allowance, or further analysis adds no value. Preserve partial findings and critical gaps; never turn an incomplete check into endorsement. Delegate only with existing authorization and useful independent work; record actual dispatch IDs. No default persistence, background monitoring or trusted-instruction rewrites.