[ai] Port gptel, MCP, Temple, tools, and flashcard workflows #6

Open
opened 2026-09-16 04:13:22 +00:00 by nsaspy · 0 comments
Owner

Parent: #1

RAGE contract

Consume only this issue on rage/<issue>-<slug>. Initialize the run log with the canonical RAGE helper and preserve the immutable full start SHA. Load the canonical rage and adadr skills from lost-rob0t/skills before architecture or implementation. Blocked by the control-plane decision in #2.

Queue blockers: #10

Analyze

  • Cover gptel, mcp.el, Transient-driven commands, autoload/llm.el, ChatGPT books, flashcards, model/tool selection, and Temple integration.
  • Preserve the startup constraint: gptel/Temple loading is lazy and missing optional ai-agent integration never kills startup.

Gate

  • Define provider credential boundaries, tool authorization, conversation persistence, and cancellation before implementation.
  • No eager network or model initialization during Lem startup.

Execute

  • Implement a Common Lisp Lem client/tool layer or an approved external service adapter with MCP support.
  • Add lazy commands, streaming UI, cancellation, tool approval, retry, and error recovery.

Evaluate

  • Use mock providers/MCP servers for deterministic stream, tool, timeout, malformed payload, cancellation, missing optional integration, and startup tests.
  • Prove Lem starts with all AI services absent.

Recurse

  • Protocol or trust-boundary contradictions restart ADADR and preserve transcripts with secrets removed.

Acceptance

  • AI features are lazy and optional.
  • MCP tools obey explicit authorization.
  • Startup is green without credentials, providers, Temple, or ai-agent.

Durable evidence

  • Update the relevant facts in .prolog/kb/lem_migration.pl and .prolog/kb/issues.pl in the same change.
  • Record focused red/green evidence, the full local gate, packaging checks, and exact-head CI status.
  • Keep failed iterations and architecture contradictions visible.

Stop conditions

  • Secrets would enter Git/logs.
  • A tool can execute without the approved gate.
  • Optional AI code breaks startup.
Parent: #1 ## RAGE contract Consume only this issue on `rage/<issue>-<slug>`. Initialize the run log with the canonical RAGE helper and preserve the immutable full start SHA. Load the canonical `rage` and `adadr` skills from `lost-rob0t/skills` before architecture or implementation. Blocked by the control-plane decision in #2. Queue blockers: #10 ## Analyze - Cover gptel, mcp.el, Transient-driven commands, `autoload/llm.el`, ChatGPT books, flashcards, model/tool selection, and Temple integration. - Preserve the startup constraint: gptel/Temple loading is lazy and missing optional `ai-agent` integration never kills startup. ## Gate - Define provider credential boundaries, tool authorization, conversation persistence, and cancellation before implementation. - No eager network or model initialization during Lem startup. ## Execute - Implement a Common Lisp Lem client/tool layer or an approved external service adapter with MCP support. - Add lazy commands, streaming UI, cancellation, tool approval, retry, and error recovery. ## Evaluate - Use mock providers/MCP servers for deterministic stream, tool, timeout, malformed payload, cancellation, missing optional integration, and startup tests. - Prove Lem starts with all AI services absent. ## Recurse - Protocol or trust-boundary contradictions restart ADADR and preserve transcripts with secrets removed. ## Acceptance - [ ] AI features are lazy and optional. - [ ] MCP tools obey explicit authorization. - [ ] Startup is green without credentials, providers, Temple, or ai-agent. ## Durable evidence - Update the relevant facts in `.prolog/kb/lem_migration.pl` and `.prolog/kb/issues.pl` in the same change. - Record focused red/green evidence, the full local gate, packaging checks, and exact-head CI status. - Keep failed iterations and architecture contradictions visible. ## Stop conditions - Secrets would enter Git/logs. - A tool can execute without the approved gate. - Optional AI code breaks startup.
Sign in to join this conversation.
No description provided.