Roadmap

Likely next

Things that probably make sense once there’s user signal or the next 1-2 sessions of focused work.

Additional adapters. Aider first (plain markdown in .aider.chat.history.md, cheapest adapter we could land); opencode next (similar storage shape to Claude Code); continue.dev after that. Cursor’s SQLite storage is meaningfully harder and waits until demand justifies it.
crates.io publish once the surface feels stable. Tag v0.1.0, cargo publish, users install with cargo install portaconv instead of cloning.
Schema promotion for production-observed record types. The spike’s initial six record types were joined by five more surfaced on real corpus data (permission-mode, attachment, last-prompt, custom-title, agent-name). Most should stay in the extensions bag, but custom-title in particular deserves promotion to Conversation.title when present, overriding the first-user-message heuristic.

Maybe (ideas worth recording, not prioritized)

pconv search <query> — a separate verb for full-content search, streaming through JSONLs with no index. Deliberately distinct from list --grep (which stays substring-over-title+cwd) so the scope creep of FTS doesn’t bleed into the discovery filter.
pconv copy <id> — bake in clipboard support with OS auto-detection (clip.exe, pbcopy, wl-copy, xclip). Removes the | clip.exe pipe dance. Small win; currently piping works.
pconv export --to <file> — shorthand for pconv dump <id> > <file>. Arguably sugar; > already works.
Fuzzy-matched list --find — nucleo-ranked (same crate portagenty uses) for “kinda remember the project name” flows. The dep is real weight though; punt unless someone asks.
Snapshot tests for markdown renderer — insta is already a dev-dep and unused. Lock down the rendered shape so format regressions surface on PR diff.
Property tests for path-rewrite edge cases — paths with spaces, UNC paths, paths adjacent to URLs. Catch the subtle regex cases my hand-rolled tests miss.
Cache warm-up on first install — the first pconv list on a big corpus is 40s; subsequent runs are 3s. A pconv cache warm subcommand that pre-fills in the background would make the first interactive call feel instant.
Shell completions — pconv completions bash|zsh|fish emits a completion script. clap makes this trivial; hold until someone hits a dead-end tab-completing --sort.

Portagenty-side (lives in the other repo)

These ideas aren’t portaconv’s to build — they’d ship as PRs on portagenty. The receiving-side contract (what flags pconv accepts, what output format it emits) is settled and documented here; the consumer is upstream.

pa init --with-agent-hooks scaffolds a workspace with pconv pre-wired: .mcp.json (portaconv MCP entry), .claude/commands/*.md (slash commands like /pull-context that call pconv), and .claude/skills/*.md (capability manifest so the agent self-discovers what’s available). Optional flag; skipped if pconv isn’t on PATH. The in-session friction killer — agent inside a pa-launched session reaches its own workspace history without copy-paste or leaving the terminal.
pa convos shim. Small wrapper that shells out to pconv with workspace context pre-filled. pa convos list → pconv list --workspace-toml <current>. pa convos last → pconv dump --latest --workspace-toml <current>. If pconv isn’t installed, print an install hint and exit. ~100 LOC + an install check.
Auto-maintain previous_paths in the workspace TOML. When pa’s auto-re-register-on-walk-up detects that a workspace id has shown up at a new path, append the old cwd to the TOML’s previous_paths array. portaconv already reads this field, so the bridging works the moment pa starts writing it — no portaconv version bump needed. Keeps the entire story inside the committable TOML; no cross-tool registry coupling.
pa://convos/<workspace-id> protocol-handler route. pa already registers a pa:// URL scheme for cross-device deep links; this slot would resolve to pconv list filtered to that workspace. Nice-to-have, low demand.

Won’t do (design decisions, recorded so nobody re-argues them)

These are parked permanently unless the foundation shifts. The answers below came from the original design Q&A and live in knowledgebase/03-design-decisions.md.

Interactive TUI viewer. Different niche. jhlee0409’s claude-code-history-viewer covers browsing well; portaconv stays CLI-pipe-friendly.
Daemon / file watcher / auto-sync. The whole thesis is extract-on-demand, not continuous-sync. A daemon brings back every problem the research doc catalogued (race windows, divergence, content-layer poisoning). Non-starter.
Writes to any tool’s storage. Read-only by construction. Every feature is additive; not one byte goes back to ~/.claude/projects/.
FTS bolted into --grep. --grep is substring-over-title- and-cwd and stays that way. Full-content search is a separate verb (see “Maybe”) if / when it lands.
pconv import or any flow that writes markdown back into a tool’s native format. If you want to commit context, use pconv dump <id> > docs/agent-context/whatever.md and git-add it. Git is the store.

What’s already shipped

For context — the following are done and live:

Core schema + Claude Code adapter
list + dump (markdown/json) + mcp serve
Path-rewrite transforms (wsl-to-win, win-to-wsl, strip)
Tier A list filters (--since / --sort / --reverse / --limit / --grep)
dump --latest --workspace-toml auto one-liner
dump --tail N length control with self-documenting truncation
Full MCP schema parity (every CLI flag mirrored as a tool argument)
Per-file list cache (~13× speedup on 2607-JSONL corpus)
v0.2 issues #1 #2 #3 all closed
Defensive previous_paths read — portaconv bridges old-path sessions the moment portagenty starts writing the field; no portaconv version bump required when that lands upstream
dump --file <path> — manual override for picking a specific backing JSONL when a sessionId lives in multiple files (WSL- vs Windows-encoded dirs). Paired with list --show-duplicates --format json for discovery. Mirrored on get_conversation MCP tool as file.

If something here looks like it should be on a different tier, open an issue and argue for it.