Pensyve Mention Workflow

Use this skill to translate mention-style Pensyve requests into the current Codex plugin surfaces.

Current Capability

Codex does not currently expose true @-mention dispatch for local plugin bundles. Treat a literal @pensyve in user text as explicit user intent, then route the request through the same Pensyve MCP tools used by $pensyve, /skills, and /pensyve.

Recommended explicit invocations today:

• $pensyve what do you remember about this repo?
• /pensyve recall release workflow decisions
• @pensyve recall release workflow decisions

The third form is a compatibility convention: Codex can follow it when the text reaches the model, but the platform does not provide @pensyve autocomplete, selection, or guaranteed dispatcher routing yet.

Workflow

1. Parse the text after @pensyve.

• If it is empty, ask one short question about whether the user wants recall, remember, observe,

inspect, status, review, or forget, and how they want to mention or reference the subject.

• If it contains a repo, file, feature, service, or person, use that as the primary entity.
• Otherwise fall back to the current repository name.

1. Route read requests through existing MCP tools.

• For recall or questions about memory, call

pensyve_recall(query, entity?, types?, limit?, min_confidence?).

• For memory inventory, call pensyve_inspect(entity, limit?).
• For connection health, call pensyve_status.

1. Route write requests through existing MCP tools.

• For durable facts, decisions, preferences, and constraints, call

pensyve_remember(entity, fact, confidence?).

• For session outcomes or procedures, ensure an episode exists with

pensyve_episode_start(participants), then call:

     pensyve_observe(
       episode_id: <working episode_id>,
       content: "[proactive/in-flight/tier-1] <observation>",
       source_entity: "codex",
       about_entity: "<entity>",
       content_type: "text"
     )

1. Require confirmation before destructive actions.

• If the mention asks to forget or delete memory, summarize the target entity and wait for an

explicit yes before calling pensyve_forget(entity).

Response Style

• Keep recall summaries concise and grounded in returned memories.
• Do not fabricate memories or imply that @pensyve is a native Codex composer primitive.
• If Pensyve MCP tools are unavailable, say: "Pensyve MCP is not connected for this session."

Then continue with non-memory work when possible.

• Strip secrets, API keys, passwords, tokens, private URLs, and credentials before writing memory.