8.5 Session install, phase one

2026-09-25 · Finished in development, not yet in a release. Principles and consent tiers: Natural language-first & consent tiers.

What you can see

• Find a skill with one sentence: say "install me a skill that organizes PDFs" in your own session. The assistant searches the Teloa official catalog (including upstream entries from Claude Code, Codex, DSH, OpenClaw, ClawHub and Hermes), lists addable candidates and explains why an entry cannot be added (unsupported, built-in and needs no install, or a connector that goes through the connection flow instead).
• Paste a link: https://github.com/<owner>/<repo>[/tree/<ref>/<dir>], .../blob/<ref>/<dir>/SKILL.md, and market links https://market.teloa.ai/<entry>/ (with /en/) are resolved directly. GitHub sources are pinned to one commit; at most 3 distinct sources are resolved per user turn.
• One card per skill: the confirmation card states the source, license, file manifest and fingerprint; "Allow once" completes "add and install on this host". Skills that go online or ship scripts list the hosts and tools declared in their text, with data egress marked separately.
• Two cards for an industry solution: first add it to your content library, then confirm loading it into your space; the loading card lists the roles, knowledge, skills, data sources, connections and work templates that will be created.
• MCP connectors: the card lists the provided tools (read-only / write), the hosts accessed and the upstream address; local-process recipes name the npm package and version to be installed (install scripts ignored, integrity checked). Connectors without credentials connect right away; those needing credentials are only registered, and you fill the secret under Market → Connections → Connection settings before connecting.

Boundaries

• Secrets never enter the chat: any parameter whose value looks like a secret (common token prefixes or a high-entropy string of 32+ characters; a heuristic that does not guarantee catching every form) or whose key name looks like a credential is rejected outright, and the rejection never echoes the value. Structured identifier fields (fingerprints, entry ids, skillPath and the like) only get the prefix check; a false positive tells you to use a description or link instead.
• AI colleagues working in tasks (including twins, group chats and sub-agent sessions) cannot search, add, install or connect; they can only say what they need in their reply. Sending request records into "Needs you" comes in phase two.
• After adding, the host re-verifies the confirmed manifest file by file against the real install preview; on mismatch it does not install and the receipt says "added, not installed".
• Ambiguous names only return a candidate list, never a confirmation card; a market link pointing at a missing entry fails outright instead of guessing by name.
• Code-execution plugins have no session entry point in this phase and still go through the pages.

Undo entry points

• Skills: "Disable" on the capabilities page. Uninstalling skills and deleting content library entries come in phase two.
• MCP connections: Market → Connections → Delete.
• Industry solutions: Space → Industry solutions → Unload.

Traceability

Every add, install, connect and load has a stable request ID and a receipt: adding is the market content import receipt, installing is the installation record (bundleHash / trustHash / version / time), connecting is the connection record, loading is the industry load record. The confirmer is the current owner and the confirmation time is the receipt time.