Skip to content

FEATURES.md

Latest commit

 

History

History
91 lines (59 loc) · 4.61 KB

FEATURES.md

File metadata and controls

91 lines (59 loc) · 4.61 KB
Raw

FEATURES.md

Roadmap and feature-addition strategy for aiyou. If you are an agent picking up work on this repo, this file is the source of truth for what gets built and when.

Product thesis (the one thing we do)

Extract data from past user interactions to know the user better, to serve them better.

aiyou turns scattered AI chat history into one normalized store, and distills it into a portable profile (SKILL.md) that every agent the user runs can consume. Nothing near-term may expand beyond this thesis.

There are exactly two axes of near-term extension:

  1. More surfaces in — where chat data lives: new harvest adapters, new import sources.
  2. More surfaces out — where the profile gets consumed: new install targets for the generated skill.

Near-term (v0.3 → 1.0): integrations & hardening only

A. More surfaces in — harvest adapters

Each is a candidate adapter following the recipe in AGENTS-MAINTENANCE.md (recon first — trust the on-disk format, not this list):

Candidate Where the data likely lives Notes
Windsurf ~/.codeium/windsurf state dbs same state.vscdb pattern as cursor
Amp ~/.config/amp/sessions generic scan already points here; verify real format
aider ~/.aider.chat.history.md, ~/.aider/ single markdown log — needs a small custom parser
Warp ~/.warp their blocks/agents store
Jules / Droid / Factory respective home dirs confirm local persistence before promising
Zed (verified format) ~/.config/zed/conversations generic scan guesses; verify against real files
Continue (real format) ~/.continue/sessions current parser is best-effort; verify
Claude Code extras todos, shell snapshots sessions are covered; extras optional

B. More surfaces in — import sources

  • Perplexity data export
  • GitHub Copilot account export (if/when available as dumps)
  • Format drift tracking: ChatGPT/Claude/Gemini export schemas change; fixtures must be refreshed when real dumps break parsers (the importers are tolerant, but identity rules must stay stable)

C. More surfaces out — profile install targets

The skill currently installs to claude, opencode, crush, codex, --agents-dir. Add:

  • AGENTS.md / CLAUDE.md injection (append a managed <!-- aiyou:begin -->…<!-- aiyou:end --> block — idempotent, never clobber user text)
  • Gemini CLI GEMINI.md context
  • goose, amp, continue, zed rule/context files
  • Cursor rules (.cursor/rules)
  • A plain --print of the skill to stdout for piping anywhere

D. Core hardening (compatibility promises)

  • Store format versioning: version the .jsonl session header ("fmt": 1) and the fingerprints cache; migrations for future changes. Goal: version updates never break the store — this is the gate to 1.0.
  • aiyou doctor — per-source discovery report (found / empty / error / skipped, plus resolved paths) so "why is source X empty" is one command.
  • Profile quality: redaction coverage tests, incremental correctness (only changed chats re-mapped), evidence quotes spot-checked.
  • Stability: sync idempotency tests on every adapter; adapters must degrade gracefully.

Near-term anti-scope: no new subsystems, no daemon, no network beyond the configured profile endpoint.

Deferred until ~1.0: utils built on the store

Semantic search, analytics (token/cost/session stats, activity heatmaps), an MCP server exposing the store+profile to agents, HTML/PDF digests, encryption-at-rest — these are second-layer utilities on a mature store.

Gate for starting this layer (all three required):

  1. Core battle-tested across real machines and OS variants
  2. Multiple minor releases with zero store-compatibility breaks
  3. Adapter suite mature (the harvest list above largely done)

Do not pull items from this layer forward because they look fun. The store's trustworthiness is the product.

Non-goals (permanent)

  • No cloud sync, no aiyou-hosted anything, no auto-push to remotes
  • No back-mutation of harness data — sources are read-only, always
  • No vector DB / embeddings in the core
  • No feature that sends unredacted data off-machine

How a feature gets done

  1. It must map to a section above (A/B/C/D) — or it's not near-term.
  2. Follow the recipes in AGENTS-MAINTENANCE.md (adapter recipe, import recipe, install-surface pattern).
  3. Definition of done: failing-test-first implementation, cargo test green, cargo clippy --all-targets -- -D warnings clean, README + docs/index.html updated if user-facing, store-compatibility preserved, committed with feat:/docs: prefix on master.