Phosphor

A local-first terminal with first-class AI — bring your own model, glowing since the VT100.

_{Based on Warp (via Zap / OpenWarp); evolving independently as its own project.}

Note

This is a personal project, mostly for fun and tinkering. It's a playground for poking at Phosphor's BYOP (bring-your-own-provider) AI stack — prompt rendering, caching behavior, and provider plumbing — not a polished product. Things here are experimental and may change, break, or get thrown away. If you want the real thing, go upstream: zerx-lab/zap. Everything below the upstream overview is what this fork has been messing with.

Most of it is being tested against FastFlowLM (FLM) as a local BYOP endpoint, and several of the changes below exist specifically to play nicer with it — especially the prompt-cache / environment-context work, which is tuned for FLM's text-only partial-prefill (KV-cache) behavior.

Phosphor is an open, local-first terminal with first-class AI and agent support. Plug in any AI provider, bring in any CLI agent, manage SSH hosts inside the terminal — with keys, history and agent state staying on your machine by default.

What Phosphor adds over upstream Warp

• No mandatory cloud — no account, login, Drive sync or cloud agent history required.
• BYOP AI providers — any OpenAI-compatible endpoint, plus native OpenAI / Anthropic / Gemini / DeepSeek / Ollama protocols, and Google Vertex AI (Gemini + Claude, authenticated through gcloud — no static key). Keys stay local.
• Third-party CLI agents — DeepSeek-TUI / Codex CLI / Claude Code / Google Antigravity (agy) wired into Blocks and the notification center.
• Built-in SSH host manager — manage hosts, configs and sessions inside the terminal, with tmux integration.
• Editable system prompts — minijinja templates rendered on the client.
• Rendering fixes — tuned Markdown pipeline; CJK soft-wrap caret and bold subpixel fixes.
• Localized UI — English / Simplified Chinese / Japanese out of the box, community-extensible.
• Privacy defaults — Cloud Agent / Computer Use / Referral / telemetry off by default.

What this fork is playing with

All experimental, spread across feature branches. Roughly in order of how much fun they were:

Google Vertex AI (BYOP)

• First-class Vertex AI provider — pick "Vertex AI" as a provider type, enter your GCP project + location (and optionally a service-account email to impersonate), and add models like gemini-2.5-flash or claude-sonnet-4-6. The publisher (Gemini vs. Claude) is routed by model name automatically.
• No static key — Vertex uses short-lived OAuth2 bearer tokens, minted via the gcloud CLI (Application Default Credentials / the active account / SA impersonation) and cached in-process. Nothing to paste, nothing stored.
• Reuses the whole BYOP path — model picker, reasoning tiers, attachment caps — dispatched by model family. Includes a fix so streaming Claude on Vertex uses :streamRawPredict (Google's streaming method) rather than the unary endpoint.

Prompts & providers

Same FastFlowLM motivation as above, from the other direction: small models served by FLM have tight context windows, and a bloated built-in system prompt eats budget the actual conversation needs. Being able to hot-swap a leaner prompt per model — live, without a rebuild — is the whole point here.

• Per-profile, per-prompt system prompt overrides. Every prompt slot in a profile can be set to Auto (pick a built-in by model family), a specific built-in family (default / anthropic / lean / gpt / beast / codex / gemini / kimi / trinity / local / troubleshooting), or a custom file. This covers the agent slots (base / coding / full-terminal-use / computer-use) and the auxiliary prompts (title generation, prompt suggestions, input completion, relevant files, workflow metadata, next command) — each picked independently in the profile editor's new System Prompts section.
• Custom prompt files, hot-loaded. Drop .j2 / .md files into a custom/ folder under your prompt template directory and they show up in every picker. They render through the same minijinja env as the built-ins (so {% include "partials/..." %} still works) and are re-read live — no rebuild. Every override degrades gracefully (bad name, missing file, .. traversal, syntax error → falls back to the built-in / auto pick).
• Active-AI prompts folded into the hot-reload env. The command-suggestion / input-completion / relevant-files / next-command / workflow-metadata templates used to be baked into a separate environment; they now hot-reload from the prompt directory like everything else.
• lean.j2 — a trimmed agent system prompt to A/B against the verbose default, for exactly this: keeping the system prompt small enough that a small FLM model doesn't blow its context window before the conversation starts.
• troubleshooting.j2 — a built-in example of a task-focused prompt: a diagnose-and-fix agent (observe the failure, one hypothesis at a time, change one thing, verify). Pair it with a "Troubleshooting" profile. It's opt-in per slot, never auto-picked by model family — a template to copy and riff on.
• Tool-list dedup — when structured tools are sent, the redundant # Available Tools text block is suppressed to save prompt tokens.

BYOP wire inspector

• A live capture window for outbound/inbound LLM traffic: the exact system prompt, structured tools, the environment block, one-shots (title gen, active AI), and streamed responses. Filterable, pausable, copyable — opened from the left panel header. Handy for seeing what actually goes over the wire.

Prompt-cache & environment-context stability

Mostly driven by testing against FastFlowLM (FLM): it partial-prefills a stable prefix, so anything volatile near the front of the prompt tanks the cache. Note this assumes a text-only FLM model — VL/MoE engines that can't partial-prefill don't benefit from the tail-block design.

• Volatile environment context moved out of the system prompt into a tail block (and appended as a standalone message) so upstream KV-cache breakpoints stay stable instead of being invalidated every turn.
• Environment context is frozen/sticky per conversation rather than refilled from shifting block metadata, so replays stop drifting.
• Tolerate null optional tool args; resolve tool-call ownership so assistant text can't orphan results.

Terminal UI (TUI)

• zap-tui-oss — a keyboard-driven terminal frontend, ported from upstream Warp's warp_tui crate and rewired onto the BYOP stack (no cloud orchestration). It shares the GUI's app identity, so your models, config and providers load unchanged; it boots interactive with shell/path Tab-completion and the full non-cloud slash-command set (/model, /profile, /prompts, /init, /compact-and, /queue, /fork, /fork-and-compact, /fork-from, /rewind with file-revert, …). Cloud-only surfaces (codebase search, cloud conversation history, credits/usage) are intentionally dropped for BYOP.
• Build/run it with cargo run -p warp_tui (the crate is a non-default workspace member; the GUI build is unaffected). Agent runs go through the same BYOP path as the GUI (chat_stream / oneshot / the prompt-override system).

Odds and ends

• Title-generation language fix (this branch's namesake) — stop Chinese-biased few-shot examples leaking into tab titles.
• English-only BYOP tool schemas and skill wrapper.
• Logging — the GUI always logs to a file now, with a ZAP_LOG_STDOUT escape hatch for stdout.
• Perf — reuse HTTP clients, unblock async DNS, trim webfetch/history overhead, cap file_glob results and back off failed retries.

Migrating from Zap, OpenWarp, or Warp

If you used this project under an earlier name (Zap, or originally OpenWarp), or are coming from upstream Warp, see docs/migrate-from-warp.md to bring your settings across. Note: on-disk config, secrets and data still live under the zap app id for now, so existing installs keep working unchanged — the storage rename is intentionally deferred (only the branding has changed so far).

Roadmap

See docs/roadmap.md.

Acknowledgements

• Warp — the upstream terminal Phosphor is built on.
• Zap / OpenWarp (zerx-lab) — the BYOP Warp fork this project descends from.
• DeepSeek-TUI — first-class CLI agent partner.