| name | aoa-session-memory-evidence-route |
| description | 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. |
| license | Apache-2.0 |
| metadata | {"aoa_scope":"session-memory","aoa_invocation_mode":"evidence-router"} |
aoa-session-memory-evidence-route
Trigger Boundary
Use this skill for a historical behavior question whose answer must come from
prior session evidence. Do not use it to replace current repository source,
live runtime state, decision authority, eval verdicts, or durable promotion.
The route finds and qualifies evidence. It never upgrades mention, prompt
visibility, selection, loading, or a SKILL.md read into invocation or
effectiveness.
Procedure
-
Resolve the logical <aoa-root> and the entity kind. Start with the
smallest typed route:
- source identity or “does it exist”:
entity-registry --lookup;
- behavior and nearby consequence:
usage-chain <anchor> --kind <kind>;
- goal lifecycle:
goal-lifecycles;
- task-to-answer chain:
task-answer-chain;
- exact text, command, path, error, or session id:
literal-query-plan;
- relation between two entities:
graph-bridge;
- current route quality:
live-scenario-corpus check.
-
Prefer the read-only session-memory MCP equivalent when it is actually
available. If transport or registry freshness blocks it, name the CLI
fallback and run the same typed route from <aoa-root>. A missing MCP tool
is a transport/runtime issue, not permission for unbounded raw search.
-
Read applicability before score. Inspect freshness, ambiguity, truncation,
omitted counts, cost profile, and the exact next expansion before opening
another layer.
-
For a skill anchor, read skill_evidence explicitly:
prompt_visible, selected, skill_read, edited, mentioned,
cooccurrence, and deflected are candidate states only;
procedure_observed, verified, and completed require an owner receipt
or reviewed task episode;
task_episode_refs are join keys, not proof;
invocation_claim_allowed=false is a hard claim boundary;
- foreign-correlation results remain auditable context, never consequence.
-
Expand only the missing evidence class. Use raw or segment refs to verify an
important claim, exact error, or bounded interval. Keep private bodies,
hidden reasoning, and unrelated context outside the packet.
-
Return the evidence packet to the source owner, decision owner,
aoa-evals, or another named authority. Do not write that owner from
session evidence alone.
For a specialized route, open exactly one shallow reference:
Authority Boundary
Session-memory packets are generated navigation and evidence surfaces. They
are weaker than current owner source, repository decisions, canonical skills,
central eval verdicts, and reviewed promotion surfaces. Route counts and
semantic proximity do not establish correctness, causality, invocation, or
benefit.
Verification
- Name the first route and why it matched the question.
- Name MCP or explicit CLI fallback.
- Report freshness, ambiguity, truncation, and privacy posture.
- Cite at least one resolvable raw, segment, session, receipt, or owner ref for
each important historical claim.
- For skill behavior, name exact package version/fingerprint when a reviewed
receipt provides it and list alternative explanations that remain.
- For candidate-semantics regression, run the reviewed
skill_candidate_semantics_contract corpus case and label its synthetic
evidence origin.
Stop Line
Stop when the bounded packet answers the question with resolvable refs, or
when it exposes one exact next expansion and the claim remains honestly
unresolved. Never compensate for a stale provider, closed transport,
unavailable receipt, or missing authority by widening into an unbounded
transcript scan.