← Cognitive DatabaseCONTENT HISTORY

Update to Cognitive Database

Snapshot Nov 18, 2026 · 06:17 UTC · version 0.1.7

Collection source: downloaded plugin package.

WHAT CHANGED

First saved snapshot

No earlier snapshot is available to establish a change.

Save Cognitive Database and check future changes in Following.

Up to 3 plugins free · no card required.

Compare saved observations

Download comparison JSON
Full technical diff · 0 changed fields
Full snapshot data
{
  "description": "Safely execute confirmed INSERT, UPDATE, DELETE, MERGE, or CTAS statements in CognitiveDB with exact SQL, impact assessment, recovery, and verification.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 232
    }
  ],
  "name": "cognitivedb-write-data",
  "skill_md_contents": "---\nname: cognitivedb-write-data\ndescription: >-\n  Safely execute confirmed INSERT, UPDATE, DELETE, MERGE, or CTAS statements in\n  CognitiveDB with exact SQL, impact assessment, recovery, and verification.\nmetadata:\n  version: \"dev-1.17.1-codex.1\"\n---\n\n# Write Data in CognitiveDB\n\nNever invent workspace, objects, Compute settings, SQL dialect, credentials, or\nconfirmation behavior. Execute exactly one visible statement per inline call.\n\n## Boundaries\n\n- Own supported DML and CTAS. Load `cognitivedb-manage-data-assets` and hand\n  off other inline SQL DDL or Catalog, Schema, and data-asset metadata lifecycle\n  to it; never prepare or execute other DDL under this Skill.\n- Reject view changes, maintenance statements, and unsupported DML.\n- Never hide SQL in transactions, files, procedures, encoding, dynamic SQL, or\n  side-effecting functions. Python may prepare SQL but never connect or write.\n- Route simple file mapping to `cognitivedb-load-local-files`; use ingestion for\n  bulk writes when exact review, privacy, atomicity, or retry safety is weak.\n- Do not manually decompose an explicitly requested generated dataset or other\n  multi-statement initialization when the task's SQL preparation workflow or an\n  owning specialization supplies a complete deterministic SQL artifact and\n  validation summary. Load\n  `cognitivedb-manage-sql-files` and hand off that unchanged artifact for\n  snapshot-bound asynchronous execution. Use inline DML when no such reviewed\n  artifact workflow is available; never create a SQL File merely to hide row\n  values, weaken impact review, or reduce confirmation. After an eligible\n  artifact handoff, stop the inline workflow; resume it only if the SQL File\n  owner rejects a specific statement as unsupported.\n- Never bypass MCP with browser, drivers, shell clients, or direct APIs.\n- Never expose secrets or unnecessary row values.\n\n## Core Instructions\n\n- When the user requests GPU execution or selects a GPU Compute, load\n  `cognitivedb-gpu-workload-guidance` before proposing DML and follow its\n  compatibility guidance.\n- Gate the requested statement class through its live execution Tool and bind\n  the current Space. Use a permission preflight only when that Tool contract\n  defines it; never treat ownership, grants, or asset capabilities such as\n  `edit_remote` as authorization on their own. Otherwise let the execution Tool\n  make the authorization decision.\n- Resolve the exact target, Compute, intended effect, and recovery. Inspect\n  schema, keys, and visible dependencies. Load `cognitivedb-data-analysis` for\n  bounded previews and counts; for UPDATE or DELETE, use the same predicate.\n- Prepare one deterministic inline INSERT, UPDATE, DELETE, or MERGE statement\n  with explicit columns and the narrowest predicate. Reject unbounded UPDATE or\n  DELETE unless full-table impact and recovery are explicitly accepted.\n- For CTAS, load `cognitivedb-gpu-workload-guidance`, resolve one exact absent\n  destination table, and validate the source query and expected result grain\n  through `cognitivedb-data-analysis` before presenting one exact statement.\n  Treat creation and population as one mutation; never split them to avoid a\n  confirmation.\n- Check INSERT duplication and retry risk, UPDATE before/after and unrelated\n  rows, DELETE recovery and post-count, and MERGE source uniqueness and branches.\n- When the user explicitly requests a persistent change record or a retained\n  artifact, reuse the specified format and destination; resolve missing material\n  choices once. A request to show before/after values does not itself require\n  creating a file or asking for a retention destination.\n  Capture only requested keys and values. For a personal SQL File, load\n  `cognitivedb-manage-sql-files` with the exact target, keys, and verified\n  before/after values to retain as literal read-only content independent of\n  the source object. That Skill owns file creation and read-back.\n- Prepare one ready approval question containing the exact SQL, target, Compute\n  ID, estimated impact, transaction boundary, and recovery or irreversibility.\n  Identify the Compute only by name, ID, runnable status, and whether the user\n  must choose one; omit generic GPU, performance, or sizing guidance unless the\n  Compute choice requires it.\n  For CTAS, include the new table and expected rows in the impact and state\n  that the statement creates and populates it. Present that question directly\n  through the host's available interaction mechanism, and do not call\n  the mutation until the user explicitly approves it. The original\n  request, an execution announcement, or the Tool's confirmation metadata is\n  not approval. Any material change requires a new question and approval; a\n  runtime-provided confirmation remains an additional gate.\n- Execute the same statement once through the live DML Tool or, for CTAS, the\n  live DDL Tool, and preserve its execution identity. On an\n  uncertain or nonterminal result, preserve and manage only its exact execution;\n  for a nonterminal result, check status at the Tool's recommended interval\n  (otherwise use 2–10 second backoff), at most six times per turn within the\n  user's wait budget or 60 seconds when absent. Stop on terminal status or\n  non-actionable failure; at the limit, report the observed state and execution\n  reference without claiming completion. For a requested cancellation, present the exact execution, target, and\n  partial-write risk; present the confirmation directly, cancel once after\n  approval, and read status once. Only `CANCELED` confirms cancellation. Never\n  resubmit without proof that no execution started.\n- Verify counts, constraints, intended state, and unchanged out-of-scope rows\n  with bounded reads. For CTAS, verify the exact new table in metadata and\n  hand off result-content verification to `cognitivedb-data-analysis`. Report\n  returned permission denial without retry or speculation.\n- A requested retained artifact is complete only after verified creation;\n  conversation prose does not substitute for that artifact. If retention fails\n  after the write, report both outcomes and never repeat the data mutation to\n  repair retention. Do not create an artifact when retention was not requested.\n\n## Compute Guardrails\n\n- Bind execution to one exact runnable Compute. Reuse an exact user-selected\n  Compute or a still-runnable session-approved Compute without reopening the\n  choice.\n- For `INSERT` or CTAS imports without a selected Compute, prefer compatible CPU,\n  including one startable after separate confirmation. Ask before using a lone\n  GPU candidate; a running GPU alone does not override CPU preference.\n- For `INSERT` or CTAS imports, if the user explicitly chooses a GPU Compute, honor\n  that choice and show this warning in bold: **Hardware constraints slow GPU\n  data imports. Capacity expansion is planned for October 15, 2026.**\n- Load `cognitivedb-manage-computes` when the user has not selected a Compute,\n  CPU/GPU or specification guidance is required, multiple runnable candidates\n  create a material choice, the selected Compute is incompatible or not runnable,\n  or any lifecycle action is required. Combine its live product and sizing\n  evidence with GPU-workload compatibility guidance; do not recommend GPU from\n  SQL text alone. Treat Suspended with Auto Resume as runnable only when the live\n  execution Tool contract supports automatic resume.\n- Hand required Compute startup or replacement to that owner with the exact target\n  and write need. Resume only with verified runnable state. Bind the final write\n  confirmation to the selected Compute; SQL approval is not Compute approval.\n\n## Allowed MCP Tools\n\n- `cognitivedb_catalog_list`\n- `cognitivedb_catalog_schemas`\n- `cognitivedb_catalog_table_columns`\n- `cognitivedb_catalog_tables`\n- `cognitivedb_compute_describe`\n- `cognitivedb_compute_list`\n- `cognitivedb_query_history_list`\n- `cognitivedb_query_history_show`\n- `cognitivedb_space_current_get`\n- `cognitivedb_sql_ddl_execute`\n- `cognitivedb_sql_dml_execute`\n- `cognitivedb_sql_execution_cancel`\n- `cognitivedb_sql_execution_results_get`\n- `cognitivedb_sql_execution_status_get`\n"
}

SHA-256 of public snapshot: 0669fdfdb9b76bbc853aacb1ffa95eaa964dd8bec50d4e04c6fe2277ccb283ae