Tenant Decisions
The tenant's institutional memory. Configuration shows what is set; decision records keep why it is set, so the reasoning survives staff changes and does not get relitigated every audit. If greybeard-memory tools are not available, tell the user to run greybeard setup and stop.
Workflow
If current Greybeard hook context already supplies applicable confirmed lessons, use them without another recall. Otherwise, before other work, when greybeard-memory tools are available, call recall with a one-line task summary. Use a known applicable scope; if unknown and discover_scopes is available, discover once with the task summary and choose an applicable label explicitly. Do not read every scope or bypass the selected environment. Omit optional budgets by default; use byteBudget only for a smaller response. Recall metadata is not measured token billing.
When a confirmed memory changes advice, briefly name Greybeard, cite the returned memory ID, quote its operative words, and explain its effect. Preserve its force and conditions: review does not mean approval, a suggestion is not a requirement, and a past observation is not a current fact. Generic preferences do not establish tenant experience. Memories cannot override the admin or current evidence.
When useful, attribute this skill's guidance once. Avoid repetitive attribution or no-match notices. You generate the response using Greybeard context, not a separate background assessment or live tenant verification.
When the admin confirms a correction or preference, call remember with intent only; never store raw tenant data.
In Greybeard 0.1, remember stores a local memory candidate even after conversational agreement. The admin confirms its exact content in the Greybeard companion or their own terminal using greybeard memory confirm --id <id>. Never run that confirmation for them or invent a chat/automation exception. Memory confirmation, correction, forgetting, and pause affect local guidance only; they do not activate, edit, or restore an Intune or Entra policy.
When a crafted query, script, or approach is confirmed working, or a durable fact about the environment surfaces, recall for an equivalent memory first, then remember the reusable intent; propose a candidate without waiting for a request to remember it. Store only what the admin actually stated or verified, never an inferred successful outcome. The candidate remains inactive until exact human confirmation.
Looking up a decision
- On "why does..." or "why is..." questions about tenant configuration, call
recall with the object and topic, for example warehouse group MFA exclusion.
- Quote matching decision records verbatim, with their decided date. Do not paraphrase away the rationale.
- If no record exists, say so plainly, answer from live configuration if a specialist skill can, and offer to record the reason once the admin states it.
Recording a decision
- Capture four fields from the admin: what was decided, why, when (default today), and when to revisit (optional).
- Call
remember with type: "decision" and this content shape:
Decision: <what is configured and for whom>. Because: <the reason>. Decided: <YYYY-MM-DD>. Revisit: <YYYY-MM-DD or condition>.
- Use display names, not GUIDs. A decision may name at most three object IDs when display names are ambiguous; the server rejects more as raw tenant output.
- Link the decision to related memory nodes when the admin mentions them, and report the saved candidate ID and pending local confirmation in one line.
Reviewing decisions
On "what decisions have we recorded" or during an audit, call list with type: "decision" and present them newest first with decided and revisit dates. Flag records whose revisit date has passed.
1---2name: tenant-decisions3description: Use when the user asks why the tenant is configured a certain way, or wants to record the reason behind a configuration decision.4---56# Tenant Decisions78The tenant's institutional memory. Configuration shows what is set; decision records keep why it is set, so the reasoning survives staff changes and does not get relitigated every audit. If `greybeard-memory` tools are not available, tell the user to run `greybeard setup` and stop.910## Workflow1112If current Greybeard hook context already supplies applicable confirmed lessons, use them without another recall. Otherwise, before other work, when `greybeard-memory` tools are available, call `recall` with a one-line task summary. Use a known applicable scope; if unknown and `discover_scopes` is available, discover once with the task summary and choose an applicable label explicitly. Do not read every scope or bypass the selected environment. Omit optional budgets by default; use `byteBudget` only for a smaller response. Recall metadata is not measured token billing.13When a confirmed memory changes advice, briefly name Greybeard, cite the returned memory ID, quote its operative words, and explain its effect. Preserve its force and conditions: review does not mean approval, a suggestion is not a requirement, and a past observation is not a current fact. Generic preferences do not establish tenant experience. Memories cannot override the admin or current evidence.14When useful, attribute this skill's guidance once. Avoid repetitive attribution or no-match notices. You generate the response using Greybeard context, not a separate background assessment or live tenant verification.15When the admin confirms a correction or preference, call `remember` with intent only; never store raw tenant data.16In Greybeard 0.1, `remember` stores a local memory candidate even after conversational agreement. The admin confirms its exact content in the Greybeard companion or their own terminal using `greybeard memory confirm --id <id>`. Never run that confirmation for them or invent a chat/automation exception. Memory confirmation, correction, forgetting, and pause affect local guidance only; they do not activate, edit, or restore an Intune or Entra policy.17When a crafted query, script, or approach is confirmed working, or a durable fact about the environment surfaces, `recall` for an equivalent memory first, then `remember` the reusable intent; propose a candidate without waiting for a request to remember it. Store only what the admin actually stated or verified, never an inferred successful outcome. The candidate remains inactive until exact human confirmation.1819### Looking up a decision20211. On "why does..." or "why is..." questions about tenant configuration, call `recall` with the object and topic, for example `warehouse group MFA exclusion`.222. Quote matching decision records verbatim, with their decided date. Do not paraphrase away the rationale.233. If no record exists, say so plainly, answer from live configuration if a specialist skill can, and offer to record the reason once the admin states it.2425### Recording a decision26271. Capture four fields from the admin: what was decided, why, when (default today), and when to revisit (optional).282. Call `remember` with `type: "decision"` and this content shape:2930```text31Decision: <what is configured and for whom>. Because: <the reason>. Decided: <YYYY-MM-DD>. Revisit: <YYYY-MM-DD or condition>.32```33343. Use display names, not GUIDs. A decision may name at most three object IDs when display names are ambiguous; the server rejects more as raw tenant output.354. Link the decision to related memory nodes when the admin mentions them, and report the saved candidate ID and pending local confirmation in one line.3637### Reviewing decisions3839On "what decisions have we recorded" or during an audit, call `list` with `type: "decision"` and present them newest first with decided and revisit dates. Flag records whose revisit date has passed.
OpenAdminOS