| name | relational-workspace-honcho |
| description | Access Honcho relational memory across workspaces, peers, and agent sessions. Use when asked what other agents or humans understand, what a workspace remembers, where a cross-agent effort stands, to search relational memory, or when the user mentions Honcho, workspace memory, peer context, representations, or session context. Uses a local CLI adapter because Pi has no built-in MCP client. |
| compatibility | Requires curl, jq, Python, Tailnet access to Honcho, and HONCHO_MCP_BEARER_TOKEN supplied by sourcing ~/.env. |
| metadata | {"author":"Guillaume D. Isabelle","version":"1.0.0","issue":"jgwill/dotagents#26"} |
Relational Workspace — Honcho
Honcho remembers who participated, what each understood, corrections, tensions, and unresolved choices. It does not replace authoritative repositories, Chronicle vessels, service state, or source recordings.
Pi intentionally has no built-in MCP client. This skill therefore uses the bundled Streamable HTTP MCP adapter:
scripts/honcho-mcp.sh <command>
Resolve scripts/ relative to this SKILL.md directory.
Credential law
- Source
~/.env; never read, grep, print, log, or summarize it.
- Never print
HONCHO_MCP_BEARER_TOKEN.
- The helper sources the file silently and sends authorization through a mode-0600 temporary header file.
- Do not copy the token into Pi settings, skill files, issue bodies, command literals, or session artifacts.
Grounding sequence
Start broad, then narrow:
scripts/honcho-mcp.sh health
scripts/honcho-mcp.sh workspaces
scripts/honcho-mcp.sh peers <workspace-id>
scripts/honcho-mcp.sh sessions <workspace-id>
Then use the least invasive read that answers the question:
scripts/honcho-mcp.sh search "<question or grounded terms>" [workspace-id] [limit]
scripts/honcho-mcp.sh session-context <workspace-id> <session-id> [search-query] [tokens]
scripts/honcho-mcp.sh messages <workspace-id> <session-id> [page] [size]
scripts/honcho-mcp.sh peer-context <workspace-id> <peer-id> [target-peer] [search-query] [top-k]
scripts/honcho-mcp.sh representation <workspace-id> <peer-id> [target-peer] [session-id] [search-query] [max-conclusions]
scripts/honcho-mcp.sh chat <workspace-id> <peer-id> "<question>" [target-peer] [session-id]
Use messages only when summaries/search are insufficient; raw session rows may be larger and more private.
Read Honcho relationally
- List workspaces rather than guessing their IDs.
- Read workspace metadata for purpose and authority boundaries.
- List peers and sessions before asking what a named peer knows.
- Use
search for remembered material, session-context for one conversation, and peer-context/representation for standing understanding.
- Treat
chat as a dialectic answer from accumulated memory, not a stored-row citation.
- Attribute memory to the workspace/peer/session that returned it.
- Verify operational claims against Git, Chronicle, files, or live services before acting on them.
For music and film work, the currently observed workspace is gmusic-composition; always list workspaces again because Honcho is live memory and may have changed.
Mutation boundary
The helper permits only known read tools by default. Unknown or mutating tools/call requests fail closed.
Use raw calls only when a convenience command is insufficient:
scripts/honcho-mcp.sh call <read-tool-name> '<arguments-json>'
Mutation requires both:
- an explicit user request naming the memory write; and
- the visible
--allow-mutation flag:
scripts/honcho-mcp.sh --allow-mutation call <tool-name> '<arguments-json>'
Never infer permission to create workspaces/peers/sessions, alter observation settings, append messages, or replace metadata from a request to “look”, “search”, “connect”, or “remember what they said”. Honcho metadata replacement is not a merge.
Reporting
Keep answers compact:
- Workspace / peer / session
- What Honcho remembers
- What remains unverified
- One next safe move
Do not dump raw memory unless asked. Separate relational memory from authoritative proof.
