@eclipse-docks/extension-ai-system

Cognitive chat runtime for Eclipse Docks: sessions, model routing, tool broker with approval modes, and a UI-agnostic event bus. Consumers use @eclipse-docks/extension-ai-system/api.

Status

The harness (prompt → specialist loop → tools → approvals → streaming UI) is the working product path.

Composer supports attachments (paperclip / drag-drop / Ctrl+V, up to 8). Payloads live on the session. The shared in-memory root is domain-agnostic: callers register live file callbacks that resolve payloads on read (no copies). AI places refs under in-memory/<sessionId>/… and optionally projects the mount to the active session folder. The root is not mounted by default. Image data-URL files are presented for vision on read.

Supported: text files (≤~100k chars stored) and visual media (jpeg/png/gif/webp/svg/ico/bmp/avif/tiff, ≤4MB). SVG keeps source text plus a preview data URL. Audio/video rejected. Storage bootstrap commands (connect_opfs, load_workspace, …) are not exposed as LLM tools.

ICOM Core (runtime/core) and ACT-R memory (runtime/memory + session buffers) are wired into each turn but do not yet drive model policy, tool policy, or prompt content in a meaningful way. Treat them as scaffolding until the todos below land.

See also docs/internal/icom-agentic-runtime.md for the layer map.

Finalize implementation

ICOM Core — make motivation real

• Feed intent / dominant need / valence into the specialist system prompt (not only goal ≈ user text)
• Let needs/valence influence routing (e.g. safety → prefer ask / lower-risk tools; clarity → explore/plan)
• Update need urgency from turn outcomes (tool failure, user correction, success) via bumpNeed
• Surface TOPIC.coreTick in the UI (optional debug panel: conscious valence, need, intent)
• Decide whether subconscious state affects anything beyond blending, or drop it until needed

ACT-R memory — close the loop

• Inject top retrieved chunks into the LLM context (system or dedicated message), not only the retrieval buffer for coreTick
• Call decay() on a schedule (per turn or idle) so activation ranking matters over time
• Use imaginal / motor buffers or remove them from BufferMap until there is a real producer/consumer
• Persist utility-relevant episodic structure (what worked) and retrieve by task similarity, not only keyword overlap
• Pin / forget UX for workspace-scoped chunks (or API for extensions)

Specialist roles & Plan/Build

• Role picker in docks-cognitive-view (plan / explore / build / general)
• Default role from ICOM intent / need instead of hard-coded build
• Explicit Plan → approve → Build handoff (checkpoint + user gate), not only read-only tool filters
• Compact / title / summary specialists invoked automatically with economy models

Tool broker & approvals

• Persist allow/deny rules and approval mode across reloads
• Cache risk classifications per tool (or signature) to avoid a classify call every batch in auto_safe
• Show why a tool was classified critical in the approval UI
• Align WebMCP / other cmd_* callers with first-class command tool names

Model router & transport

• Persist router utility scores with cognitive state
• Move tool + reasoning to Responses API (or equivalent) so reasoning_effort and tools can coexist
• Re-enable effort levels when tools are required, once transport allows
• Expose manual effort override consistently with Models.dev capabilities

Sessions, context & compaction

• Inject compactionSummary into provider messages; LLM summarize on compact (extractive fallback)
• Auto-compact threshold tunable in settings; compact summary should include ACT-R goal/retrieval highlights
• Keep context-usage estimate in sync with what is actually sent to the provider (retrieved chunks, enhancers)
• Checkpoint restore should rehydrate Core state, not only session buffers/history

Consumer API & docs

• Document ./api surface (runtime, TOPIC, enhancers, external providers, settings helpers)
• Fix stale references (cmd_* tool prefix, command ids) in internal docs
• e2e coverage for approval modes, external provider registration, and prompt enhancers

Cleanup

• Remove or quarantine any leftover pre-cognitive modules if still present in the tree
• Trim public exports to the stable façade; keep Core/memory internals package-private unless consumers need them