Skip to content

Roadmap

Jump to bottom
jstuart0 edited this page May 5, 2026 · 4 revisions

Roadmap

Shipped in v0.4.0 (2026-05-05): PostgreSQL backend at parity with SQLite — dual-dialect schema, Drizzle migration runner, advisory-lock rolling-deploy serialization, PostgresSearchBackend (ILIKE), two-job GHA CI, deploy/overlays/postgres/ Kustomize overlay. See CHANGELOG for full details. Follow-ups: pgvector, tsvector search, jsonb payload, LISTEN/NOTIFY transcript sync, FOR UPDATE SKIP LOCKED leaser.

Shipped in v0.3.0 (2026-05-05): audit-remediation campaign — HOST bind hardening, signup default off, settings allowlist, Authentik trust secret, CSP report-only, backup sidecar, restore runbook, and 15 other hardening items.

Shipped in v0.2.0-pre.3 (2026-04-27): projects registry with cwd-based session resolution, template ↔ project linkage with live inheritance, AI Ask command surface (12 action_request kinds — launch / resume / pin / archive / stop / delete / edit-project / delete-project / edit-template / delete-template / add-channel / alert-rule / freeform-rule / bulk), project-level alert rules with periodic sweep + first-run backfill, search-result event highlight, telemetry install classification.

Living document. Priorities shift based on user signal (issues, Ask usage, what breaks for real people). "Next" = probably this quarter; "Later" = on the list but no commitment.

🚀 Near-term (next few weeks)

Reliability

  • Transient-retry at the LLM adapter level — auto-retry once on rate-limit / 5xx / Ollama cold-start errors so cold starts don't surface as user-visible failures. ([issue TBD])
  • Multi-arch Docker builds — linux/arm64 for Raspberry Pi / M-series Linux VMs.
  • Postgres follow-ups (from v0.4.0): pgvector event embeddings, jsonb raw_payload + GIN index, LISTEN/NOTIFY transcript sync, tsvector full-text search, FOR UPDATE SKIP LOCKED leaser, legacy-init removal after soak period.

Integrations

  • Mistral adapter (#1)
  • Cohere adapter (#7)
  • Slack channel (#2)
  • Discord channel (#3)

DX

  • CI on PRs — GitHub Actions running biome + typecheck + tests (#9).
  • Diagnostic CLI (bun run diagnose) (#6).
  • Linux install script (#4).

UX

  • Syntax-highlighted code blocks in Ask markdown rendering (#8).
  • SVG a11y pass (#5).

📅 Medium-term (quarter-ish)

Session control

  • True "stop session" action from Ask (right now the Ask assistant is read-only).
  • Session groups — explicit grouping beyond the automatic project color tint; think Linear cycles for agents.
  • Long-running process awareness — detect when the agent is running a dev-server loop and separate that from real work.

AI

  • Retrospective memory — give the watcher context from past sessions in the same repo, not just the current session.
  • Batched classifier — right now the classifier computes per-session health on demand. Batch it on a schedule so the dashboard can pre-render health chips without a fan-out per tile.
  • Cost telemetry — per-provider, per-feature. Right now spend is tracked per-session for the watcher; Ask and classifier calls aren't charged back to any budget.

Channels

  • Email channel for digest delivery. Less interactive, but useful for passive-catch-up flows.
  • Webhook channel — generic POST-this-JSON-wherever for users who want to pipe HITL into a custom endpoint.

Deployment

  • Kube-native HPA — watcher leaser scales horizontally already; the ingest path is the remaining bottleneck.
  • ARM docker builds — published multi-arch.

🔮 Later / speculative

  • Desktop app — Electron or Tauri wrapper so the dashboard lives in the system tray. Useful mostly for "bring notifications to the user" — Telegram already does that remotely.
  • VS Code / Cursor extension — show live AgentPulse session state for the current workspace next to your editor.
  • Multi-user teams — watcher configs and HITL ownership scoped per-user. Today the instance is assumed single-user.
  • Fine-tuning export — with user consent, export the transcript + outcome of sessions into a training format (OpenAI JSONL, Claude messages) for fine-tuning a local model on "how this user actually codes with Claude".
  • Replay mode — re-render a session as a slideshow. Useful for retrospectives ("why did we take that path?") and onboarding new team members.

What's explicitly NOT on the roadmap

  • Running the agents themselves. AgentPulse observes and orchestrates sessions — it doesn't BE the agent. Claude Code and Codex CLI each already run fine; the goal is visibility and control around them, not replacement.
  • Cloud SaaS hosting. The project is self-hosted by design. If you want "AgentPulse as a service" that's a fork you're welcome to ship.
  • IDE-integrated chat UI. That's Cursor's / Copilot's job. AgentPulse is a separate window you glance at.

How priorities get set

Three inputs, roughly in this order:

  1. Things that break on real instances (mine, other early users). These jump the queue.
  2. Unlocking integrations (new LLMs, new channels). Expands the addressable population.
  3. Polish + DX. Everything in this bucket matters but also compounds over time — never urgent until a contributor ships it.

If you care about something not on this list, open an issue with the enhancement label — if the write-up is concrete enough it becomes a good-first-issue candidate directly.

Clone this wiki locally