Why this should exist

#79334 is the right narrow seam for image-first structured extraction, but the maintainer feedback there surfaced the broader platform need: many plugins need bounded host-owned inference, not raw OAuth credentials and not product-specific core routes.

Today plugin authors that need typed model output for knowledge-base enrichment, CRM extraction, finance ingestion, support triage, or import pipelines still face two awkward choices:

• ship bespoke bridges that shell out into host runtime behavior
• ask for product-specific SDK seams one plugin shape at a time

A generic host-owned inference surface would keep auth, provider routing, timeouts, and safety in OpenClaw while letting plugins request structured work without ever receiving tokens.

Proposal

Add a general runtime API under api.runtime.llm, something like:

• api.runtime.llm.completeStructured(...)

That is a better long-term fit than broadening media-understanding further or exposing raw OAuth credentials to plugin code.

Desired behavior

• text-only structured completion
• image-plus-text structured completion
• optional jsonMode and jsonSchema
• host-controlled model and auth profile
• bounded timeout
• controlled success and error envelopes
• provider and model metadata in the response

Existing building blocks

• api.runtime.llm.complete(...) already exists for trusted plugin text completions
• api.runtime.modelAuth.getRuntimeAuthForModel(...) already resolves runtime auth
• #79334 proves a bounded provider-owned structured extraction lane for media inputs

This follow-up would generalize the pattern without requiring each plugin to invent its own bridge.

Example use cases

• KB plugins: turn text or screenshots into normalized evidence records
• CRM plugins: extract people, companies, dates, action items, and summaries from text plus images
• Finance plugins: receipt and invoice field extraction
• Support plugins: screenshot plus reproduction-note triage
• Migration/import plugins: map raw content into plugin-owned JSON before storage

Boundary

OpenClaw core owns

• auth resolution
• provider runtime exchange
• model/profile policy
• bounded execution
• typed generic inference API

Plugins own

• routes
• prompts
• schemas
• storage
• follow-on behavior

Non-goals

• no raw OAuth token or refresh token exposure to plugins
• no product-specific routes or schemas in core
• no tool-using long-running agent lane
• no replacement of #79334; that PR should stay the media-understanding seam

Acceptance criteria

• no user model API key required when the host runtime already has auth
• text-only structured requests work
• image-plus-text structured requests work when the chosen provider supports them
• JSON/schema validation failures return controlled errors
• host policy still gates model/profile/agent overrides
• secrets are not returned or logged
• existing api.runtime.llm.complete(...) behavior does not regress

Relationship to #79334

That PR should remain intentionally narrow: image-first structured extraction via media-understanding.
This issue tracks the broader host-owned plugin inference seam so future plugins do not need to keep stretching media-understanding to cover every structured workload.