[plan-12] Provider & Extensibility Roadmap — net-new capabilities, not defects

Why this plan exists

Plans 01–11 closed the defect backlog: every "it silently does the wrong thing" bug across hooks, spawning, worker lifecycle, install, observer security, env isolation, server runtime GA, OpenCode capture, data-pipeline integrity, build hygiene, and observer output fidelity is now fixed, tested, and gated in CI.

What remains in the tracker is a different class: feature requests — net-new capabilities the product does not yet have and that nothing is currently doing wrong. They were intentionally kept open through triage rather than force-folded into a defect plan. This plan is the single master that routes them, so the backlog reflects "planned roadmap" instead of "untriaged pile." Each item below is independently shippable; an executing agent can /do one at a time.

Note: the OpenAI-compatible base-URL cluster (#2382 OpenRouter base URL, #2590 custom provider base URL, #2622 DeepSeek, #2393 LM Studio model inheritance) was already implemented — a single configurable CLAUDE_MEM_OPENROUTER_BASE_URL turned the OpenRouter provider into a generic OpenAI-compatible client. Those four are closed. Vertex (#2522) is the only provider request that survives here because it needs a different auth/transport, not just a base URL.

Children

Providers

#2522 — Vertex AI support for the Gemini provider (GCP service-account / ADC auth + Vertex endpoint; distinct from the OpenAI-compatible base-URL path which is done).

Ingestion & filtering

#2690 — backfill / ingest existing Claude Code session JSONL files (one-time import of pre-claude-mem history into the observations pipeline).
#2498 — incremental scan skill for changes made outside a Claude session.
#2463 — tool_response-level filter (file-extension / size / content heuristic) to prevent Read-of-binary / Read-of-playwright-residual blowups.
#2423 — per-directory disable support (finer-grained than project exclusion).

Observability & introspection

#2566 — MCP grammar-introspection routes + worker provider retry telemetry + audit log.
#2513 — clarify logger audit policy and codify the CI behavior around it.

UX / semantics

#2645 — i18n support for startup UI labels.
#2467 — PreToolUse:Read context injection encourages the LLM to treat Read as a turn boundary in connected tool sequences (a product/semantics decision: should PreToolUse:Read inject at all, or only annotate?). Deferred from plan-01 because it is a behavior-design call, not an IO-discipline defect.

Suggested execution order (independent; pick by value)

#2463 tool_response filter — small, high daily value, prevents real blowups. Extends the existing truncation in the observation path with extension/size guards.
#2423 per-directory disable — thin addition over the existing CLAUDE_MEM_EXCLUDED_PROJECTS / shouldTrackProject machinery.
#2690 backfill ingest — a CLI/skill that walks ~/.claude/projects/**.jsonl and replays them through the observation pipeline (reuse the transcript parser).
#2566 retry telemetry + audit log — partially seeded: plan-05 already added an observer tool-attempt audit log; extend with provider retry counters + MCP grammar-introspection routes.
#2522 Vertex — new auth/transport branch on GeminiProvider; largest provider lift.
#2498 incremental scan, #2645 i18n, #2513 logger audit policy, #2467 Read-as-turn-boundary — schedule by demand.

Out of scope

Anything already shipped in plans 01–11 (defects).
The OpenAI-compatible base-URL cluster (#2382/#2590/#2622/#2393) — done.

Non-actionable tracker/meta issues (handle separately, not part of this plan)

#2601 — tracking issue for "8 PRs ready for review"; superseded by the consolidated branch work in plans 01–11. Close once that branch merges.
#2646 — "Are we really sure claude-mem works with Claude Code?" — a question, not a defect; answer in-thread (the green CI suite + lifecycle tests are the evidence).
#2418 — "Enable OpenHarness integration / allow PR from fork" — repo-policy/CI config, not a code change in this tree.