8Dionysus

Capture a meaningful architectural or workflow decision as an ADR or decision note, place it in the canonical decision surface, and verify the rationale stays discoverable. Use when structure, boundaries, tooling, or workflow expectations change and future contributors need to know why. Do not use for tiny self-evident edits or when the real problem is source-of-truth ambiguity rather than decision rationale.

Classify whether a requested action is safe to proceed, requires explicit approval, or should not be executed. Use when a task may be destructive, security-sensitive, or operationally sensitive and the authority boundary is unclear. Do not use for clearly low-risk work or when the main need is previewing a mutation rather than deciding whether it may proceed.

Detect reviewed or repeated project processes that are candidates for automation, classify whether they are candidate-ready, and route them to the right owner layer without granting schedule or mutation authority. Use when a reviewed session or repeated project slice reveals a recurring manual route and the honest question is whether it should stay manual, become a bounded skill, or become a playbook automation candidate. Do not use for one-off exploratory work, creative synthesis with no stable trigger or output, live scheduling requests, or hidden self-change.

Clarify system or domain boundaries, name contexts, and surface interfaces so changes stay semantically scoped. Use when naming is overloaded, responsibilities are mixed, or a task needs a cleaner boundary before coding. Do not use for tiny local edits or when the boundary is already clear and the real task is contract validation.

Plan and execute a bounded, non-trivial repository change with explicit verification and reporting. Use when code, config, docs, or operational surfaces change in a meaningful way and the task benefits from an explicit plan. Do not use for trivial wording fixes or when a more specific risk skill fits better.

Bridge provisional checkpoint evidence into one explicit reviewed closeout execution chain without turning notes into final harvest, progression, or quest authority. Use when a reviewed session artifact exists, checkpoint notes or closeout handoffs already carry focus hints, and the next honest move is to reread the reviewed artifact while driving donor harvest, progression lift, and quest harvest in that fixed order. Do not use for mid-session collection only, for hidden execution inside `aoa closeout run`, or when the request tries to mint final verdicts from checkpoint notes without reviewed evidence.

Turn a validated bounded diff into one intentional local commit with explicit scope review, named verification carry-forward, and a visible stop line before push or publish. Use when a bounded code, config, or docs diff is already prepared locally and the next honest move is one local commit boundary rather than more coding. Do not use when the task still needs more repair or verification, when no bounded diff exists, or when the main task is push, PR, release, or public-share.

Design or extend contract-oriented validation at module, service, or workflow boundaries. Use when two components interact across a meaningful interface, downstream assumptions matter, or a smoke path needs an explicit contract. Do not use for purely local changes or when the real need is invariant or property testing instead of boundary validation.

Use when a `.aoa` session-memory root must be created, installed, checked, repaired, or prepared for Codex hook-based session archiving.

Use when an indexed `.aoa` session needs provisional first-pass distillation into lessons, decisions, failures, and automation candidates without promoting them to reviewed truth.

Use when an agent needs evidence from prior `.aoa` sessions about how a skill, MCP, hook, tool, API, goal, eval, test, validator, script, decision, error, receipt, or other recurring operational entity was used, what happened nearby, and which raw or segment refs prove it.

Use in any Codex session when the user mentions `.aoa`, session memory, Codex transcripts, compaction, prior session rehydration, hook failures, or AoA session-memory validation.

Use when an agent needs to build or query the portable `.aoa` SQLite search index for archived Codex sessions, with evidence refs and freshness checks.

Use when the live Codex PreCompact and PostCompact hook path must be proven through `.aoa` using the Codex app-server compact probe.

Use when AoA Codex hooks must be inspected, trusted, or compared against expected `.aoa` commands in the native Codex hook registry.

Use when many indexed `.aoa` sessions need a first-wave distillation conveyor that separates safe automatic first-pass artifacts from manual review and mechanics-improvement queues.

Detect reviewed or repeated project processes that are candidates for automation, classify whether they are seed-ready, and route them to the right owner layer without granting schedule or mutation authority. Use when a reviewed session or repeated project slice reveals a recurring manual route and the honest question is whether it should stay manual, become a bounded skill, or become a playbook automation seed candidate. Do not use for one-off exploratory work, creative synthesis with no stable trigger or output, live scheduling requests, or hidden self-change.

Turn a reviewed session artifact into a bounded HARVEST_PACKET, route each reusable unit to the right AoA owner layer, and hand off to the next honest post-session skill when needed. Use when the source is a reviewed session transcript, compaction note, or recap and the real question is what should survive this session rather than merely what happened. Do not use for active or unreviewed sessions, raw session capture or indexing work, or when the only remaining object is already one repeated quest unit that just needs `aoa-quest-harvest` for the final promotion verdict.

Delegate one bounded child route through quest-passport law, local coding-agent execution defaults, hard gates for progression, self-agent, and stress posture, and governed return into reviewed closeout. Use when a parent route already has a real anchor and named outputs and the honest next move is a narrower child reviewer, evaluator, verifier, or leaf helper rather than a larger orchestration layer. Do not use when the route lacks an anchor, the outputs are unnamed, the work is `d3+` and still needs a split, or remote delegation is being used to bypass proof, approval, or closeout posture.

Make bounded infrastructure, configuration, service, or operational changes with explicit risk framing, proportional verification, and rollback thinking. Use when a change has runtime or deployment implications and needs stronger discipline than a normal code edit. Do not use for purely local code changes, or when the main need is approval classification or preview-first execution.

Classify drift, friction, proof gaps, ownership confusion, and repeated failure patterns from a reviewed session into a bounded diagnosis packet without mutating anything yet. Use when the next honest move is diagnosis before repair and the reviewed material points to repeated contradictions, blockers, or boundary blur. Do not use for live sessions, for issues that are already fully diagnosed, or when the real task is final quest-promotion triage.

Capture a meaningful architectural or workflow decision as an ADR or decision note, place it in the canonical decision surface, and verify the rationale stays discoverable. Use when structure, boundaries, tooling, or workflow expectations change and future contributors need to know why. Do not use for tiny self-evident edits or when the real problem is source-of-truth ambiguity rather than decision rationale.

Bridge provisional checkpoint evidence into one explicit reviewed closeout execution chain without turning notes into final harvest, progression, or quest authority. Use when a reviewed session artifact exists, checkpoint notes or closeout handoffs already carry focus hints, and the next honest move is to reread the reviewed artifact while driving donor harvest, progression lift, and quest harvest in that fixed order. Do not use for mid-session collection only, for hidden execution inside `aoa closeout run`, or when the request tries to mint final verdicts from checkpoint notes without reviewed evidence.

Design or extend contract-oriented validation at module, service, or workflow boundaries. Use when two components interact across a meaningful interface, downstream assumptions matter, or a smoke path needs an explicit contract. Do not use for purely local changes or when the real need is invariant or property testing instead of boundary validation.

Find, record, correct, or supersede a durable repository decision and its derived indexes. Use for prior rationale, a meaningful accepted choice, or decision-source/index drift. Do not use for ordinary documentation, an open choice, current source-authority conflicts, or runtime/generated authority mapping.

Select or apply an owner-local or cross-repository eval, route a confirmed no-fit to intake/design, or classify supplied session hits as eval candidates. Use aoa-evals for central proof bundles, named results or verdicts, reports, and lifecycle. Do not use for ordinary tests, undefined invariants, or generic session harvesting.

Carry provisional checkpoint hints into a reviewed session closeout: harvest, progression, quest-promotion, and branch handoff. Use to collect bounded hints or execute after the current session artifact is closed and reviewed. Do not use mid-session for final judgment, without reviewed evidence, as a hidden playbook runner, or to write owner truth.

Use only before any memo artifact exists to decide whether live-session, closeout, landed-work, PR, diff, or review evidence yields one guarded memo candidate, export, or no-write result. If a concrete memo candidate, export, object, corpus, lifecycle, or read-model target already exists, use aoa-memo. Do not use for eval mining, raw-session maintenance, generic harvest, recall, durable memory, or proof.

Use only when a literal ref to a closed/reviewed session packet is supplied, to extract, classify, propose promotion or branching, or identify an automation opportunity. Missing, live, or unreviewed evidence must fail closed without inventing a ref. Natural closeout preservation belongs to aoa-memo-writeback. Do not use for generic summaries, progression, direct owner writes, scheduling, mutation, or telemetry authority.

Diagnose a reviewed recurring session/workflow failure, or carry one bounded owner repair through checkpoint, rollback, and real health verification. Use for reviewed contamination, drift, repeated route/tool failure, or an established diagnosis ready for repair. Do not use on live evidence, for vague self-improvement, or to call a proposed or merely executed change verified.

Sanitize private technical material into a public-safe owner-bounded derivative, or resolve authority among authored, generated, runtime, and installed sources. Use for governed incidents, logs, configs, diagnostics, or conflicting source roles. Do not use for ordinary editing, memo/session work, direct publication, or durable-memory authority.

Shape one software responsibility boundary through bounded-context, core-boundary, or port-adapter mode. Use when responsibilities, reusable rules, orchestration, or concrete dependencies are entangled and a small owner-aware boundary is needed before implementation. Do not use for source-authority lookup, ordinary edits, test design, or a boundary that is already clear.

Titan bridge event normalization or visible replay; for queue or gate decisions use titan-console; never live steering.

Titan approval queue or record, console witness, unsent app-server plan, or receipt; aoa-sdk witnesses only.

Titan receipt ingest, candidate recall, or aoa-memo retention handoff; never generic or session memory.

Find, record, or correct durable repository decisions through one decision family.

Turn closed, reviewed session evidence into a bounded multi-axis progression candidate with baseline and attribution limits for aoa-agents review. Do not use on live impressions, without a baseline, to assign one universal score, mutate a role, grant authority, or set routing policy.

Delegate one bounded child-agent route from an anchored parent task, including host launch, named outputs, return validation, and parent closeout. Use only when the user explicitly asks to delegate or summon a narrower helper. An incomplete summon request must block, not receive an allowed decision. Do not use without a parent anchor, with unresolved branching or unnamed outputs, for unsplit deep work, or to bypass approval, owner, or proof boundaries.

AoA/Abyss durable memory and owner orientation: use when ongoing work may depend on reviewed prior decisions, provenance, lifecycle/currentness, or an existing memo artifact, even when none is named. Also use to recall, review, or evolve a candidate, export, quarantine packet, object, corpus identity, lifecycle target, or read model. First tool turn reads only this SKILL.md. Use aoa-memo-writeback only before any memo artifact exists. Do not use for raw-session retrieval, proof, routing authority, roles, workflows, KAG substrate, runtime storage, or generic notes.

Capture, diagnose, review, or prepare a bounded diagnosis/repair handoff for one concrete Abyss runtime target through the abyss-stack diagnostic spine. Use for a named preset, profile, truth goal, or diagnostic_session_v1. Review current or supplied owner evidence with last-good, drift, or freshness classification. Do not use for generic health questions, artifact trust/source disputes, actual repair/restart/approval, session mining, or tasks needing no runtime evidence.

Central AoA proof or lifecycle: first tool turn reads only this SKILL.md. Use to select, review, or evolve source bundles, named eval results or verdicts, source-linked reports, Eval Forge owner review, or source/reader drift. Use aoa-eval for local selection, application, intake/design, or session-hit classification. Candidates, readiness checks, receipts, generated readers or outputs, MCP outputs, and green commands are not proof.

Diagnose a missing or drifting AoA workspace capability across owner source, projection, host scope, prompt visibility, configuration, transport, and owner service. Use when a fresh session lacks an expected skill, MCP, or tool, or when installed, configured, and live states disagree. Stop at the earliest evidenced break and hand repair to its owner. Do not use for general cross-repo orientation, narrow owner work, or the repair itself.

Use read-only `aoa_kag` MCP to locate canonical AoA owner sources/full capability contracts; check known-capability applicability/live binding; ground evidence-bearing relations, compatibility, or cross-repo impact; or investigate/compare a bounded multi-owner question needing navigation. Return owner handles. Not for raw sessions, web research, proof/memory, role/routing/playbook decisions, mutations, standalone runtime diagnosis, or an exact authoritative source needing no navigation.

Route one OS Abyss artifact through owner controls, evidence/drift review, registry selection, and fail-closed consumer admission. Use before producing or consuming a concrete bundle, package, container, model/runtime, generated export, extension, or media artifact, or when provenance, SBOM/signature, source-ref, registry, or trust-gate evidence is stale or missing. Do not use for ordinary source edits/tests, generic supply-chain explanations, raw sessions, or source-authority disputes.