AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

foundry-iq

Foundry IQ knowledge bases. WHEN: make local or Blob documents searchable; create/diagnose KBs; triage unsupported connectors or multi-source KB creation/reconfiguration (even with classic-search terms); connect existing KB to agents (including multi-source KBs); create/reuse a Search service; retrieve from an existing knowledge base with citations; no brand words needed. Read its procedure before query/target questions. NOT: other KB providers, repository-file search, classic Azure AI Search index/query/app work with no Foundry IQ KB or connector request, generic agent creation.

Original source
microsoft
License
MIT
Pinned revision
3f5b461b2314e01c400564051e92469114c5b68f

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read microsoft--foundry-iq at revision 3f5b461b2314e01c400564051e92469114c5b68f, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 45

lifecycle/cleanup.md

Plan exact cleanup

When to use

Optional lifecycle: plan run-created KB/KS and Prompt cleanup. Read helper contracts.

Do not use

Not creation, repair, ownership recovery or arbitrary child deletion. Hosted teardown: hosted-cleanup-unsupported.

Inputs and discovery order

Use request/session/records, then exact Azure readback. Ambiguous names need one focused question; no subscription inventory/repeated intake. No approval for planning.

| Input | Why needed | Required? | Discovery order | Safe default | If missing or unanswered | Reconfirmation trigger | |---|---|---|---|---|---|---| | Target | Scope | Yes | Request/session/records | None | Select exactly | Scope | | Owner/records | Proof | Yes | Trusted audit | None | Block | Evidence | | Bounds | Complete reads | Yes | Explicit/default | 20 pages/500 objects | Block at limit | Bounds | | Consent | Authority | Before write | Unchanged envelope | Unapproved | No write | Plan/state |

Owner metadata is neither ownership proof nor RBAC. Validate record bindings, not authenticity: retain original approvals/results and original successful call records. Never invent provenance or treat equivalent GET as create evidence. Existing/reused/updated/foreign resources stay unowned. Readiness isn't index content/ownership. Require original child versions, not guesses.

Decisions

  • KB: delete its definition only; retain sources, generated objects and consumers. Retrieval stops; consumers are not detached.
  • KS: the existing executor deletes only the source. The service cascades to its exact generated index (File), or index, indexer, skillset and datasource (Blob/ADLS). Show those names before approval. Retain local/Storage originals, accounts, containers, models, services/grants. Uploaded File copies and indexed content are deleted.
  • Before KS cleanup, enumerate every page of KBs in that Search service. References block: separately plan/delete an owned KB first; never detach outside-plan KBs. Native source in-use rejection remains a final guard. Enumerate scoped knowledge sources, indexers and skillsets to detect shared children; compare exact child definitions/ETags with original snapshots. Unknown consumers, missing objects, denied/partial inventories, or changed incarnations block.
  • Prompt: delete the connection, optionally preceded by one selected run-created version. Preserve containers, other versions, KB, models/grants. Deletion removes that version's tool binding; it never edits tools on a retained version. Require SDK ≥2.4, unshared state and complete paged agent/toolbox inventories, all versions including drafts, with exact version GETs. Only the selected version may consume it. Hosted/workflow/external definitions, unknown tools/skills/policies block. Sharing outside this project blocks: project agent inventory cannot prove external consumer absence. isSharedToAll: false alone is not proof.

Proposed plan

python helpers/cleanup_plan.py --plan request.json

Closed request: schema_version: "1.0", nonempty owner, exact target, creation_input_file, creation_receipt_file. Relative paths resolve beside the request. Use original producer outputs, not hand-built response JSON.

Before the original creation, retain proof with:

python helpers/search_reconcile.py --input approved-kb.json --cleanup-receipt-dir "<absolute-private-directory>"

The same flag applies to approved file_source.py and prompt_connect.py invocations under helpers/. Use an existing private user-owned directory outside the skill; validate paths/permissions without ACL changes. From cleanup_receipts, select the target's verified receipt. acknowledged is diagnostic, not authority. Immutable receipts contain filtered hashes/IDs/versions, never CU keys, auth/cookies or credentialized bodies. The keyless-Blob callback's privacy restriction is unchanged. Verified proof requires native create acknowledgement plus matching original readback. Reuse/update, recovery, missing identities/ETags or failed child readback block; later GETs cannot repair proof. Post-write failures are partial; retain errors/receipts.

Original Blob checkpoints from blob_source.py --receipt-dir also work with creation_result_file binding the checkpoint digest. ACK/recovery-only evidence cannot prove original child versions. Never recreate resources for cleanup proof.

Optional inventory_limits: closed integer max_pages/max_resources, defaults 20/500, maxima 100/5000 per collection; total Prompt versions also bounded. Connection-only agent_version + agent_creation_receipt_file select one version; name comes from the same creation input. Never select automatically.

Closed target:

  • Search: type: "knowledge-base" or "knowledge-source", endpoint, name, api_version (2026-04-01 or 2026-08-01-preview).
  • Prompt: type: "prompt-agent-version", project_resource_id, project_endpoint, name, exact numeric version.
  • Connection: type: "project-connection", same project identity, name.
  • Hosted: only type: "hosted"; blocked.

Initial receipts authorize only their version; native ID/time/definition are guarded. Retain the container.

Confirmation

Save only execution_input as JSON. Review exact delete/retain/ordering, then change only approval.confirmed from false to true. Never recalculate hashes. plan.cleanup_approved: true labels a destructive proposal: it is not consent while envelope approval is false.

Mutation

Executors: helpers/search_reconcile.py --input approved-cleanup.json or helpers/prompt_cleanup.py --input approved-cleanup.json. No combined creation/cleanup approval. Planners never delete, detach, provision, change roles, invoke retrieval/models, or auto-approve.

Closed dependency_guard snapshots and targets are revalidated immediately before destruction. For connection+version, check protected state before writing, delete/ verify the version, then rescan consumers before conditional connection DELETE. Stop on failure; report writes/remaining resources. Never strip guards or handcraft unguarded legacy envelopes.

Verification

ETag, fingerprint, identity, definition, sharing, inventory or child drift needs new planning/approval. These are fresh observations, not an atomic dependency lock; avoid concurrent administration. No force-delete or retry around a native dependency rejection. After source absence, GET each exact child to verify the cascade; remaining/unreadable children produce a failure, never child DELETE.

Failure and partial completion

Denied isn't absent. A definitive target 404 returns already-absent, no execution input and no new ownership; never authorize a future same-name replacement. Missing referenced/generated objects or incomplete pagination block. Unsupported enumeration (including drafts) names the missing operation; never substitute a filtered or partial list. Execution preserves blocked/partial; never advance after failed verification or delete a reused consumer.

Cleanup

No resources created; no recursive cleanup. Hosted teardown and arbitrary external consumer discovery remain unsupported.

Return contract

Exit 0: planned/already-absent, not cleanup completion. Exit 2: no-write blocked with first failure and retained scope. planned returns exact closed-schema execution_input, plan_fingerprint, executor, and delete/retain/blocked/ordering in approval_summary.

References

Authorities: failure/conflict/uncertainty only. Source cascade/in-use, paged KB inventory, paged agents and draft versions.