pi-muse-code-context-fold

Pi extension that registers a muse-code model provider backed by the Muse Code CLI (muse exec), and keeps it coherent when you switch models mid-session (e.g. grok → muse → grok) by resuming a real muse session instead of re-sending the whole conversation as text on every turn.

Originally a companion extension for pi-muse-bridge (see ## Origin below). As of 0.3.0 it no longer depends on that package: it registers the muse-code provider itself.

For Muse Code CLI, not Muse OAuth.

Requirements

• Pi 0.85.1 (verified; other 0.8x versions likely work)
• The muse CLI on PATH (or PI_MUSE_BINARY pointing at it), logged in
• No pi-muse-bridge needed or wanted; remove it if installed (pi remove npm:pi-muse-bridge) to avoid two extensions registering the same muse-code provider id.

Install

# from source (until published)
pi install ~/pi-muse-code-context-fold
# or after cloning your fork
pi install git:github.com/zereight/pi-muse-code-context-fold

Then reload the running Pi session:

/reload

Behavior

• Registers the muse-code provider directly (pi.registerProvider("muse-code", ...)), with models read from Muse's local model catalog.
• Tracks one muse --session-id per running Pi process (src/session.ts). The first muse-code turn folds the full prior Pi conversation (other models included) into the prompt, since there's nothing to resume yet. Every later muse-code turn just sends the newest user message, because the resumed muse session (verified against the real muse CLI: two separate muse exec --session-id <same uuid> calls land in one durable ~/.local/share/muse/sessions/.../session.jsonl) already remembers everything before it.
• Bounded fold: keeps the newest 20 turns and 12,000 chars (MAX_PRIOR_TURNS, MAX_PRIOR_CHARS in src/fold.ts). Any single prior message (e.g. a huge tool result) is additionally capped at 2,000 chars (MAX_MESSAGE_CHARS) so one big tool dump can't crowd the real conversation out of the budget.
• Non-text blocks (toolCall args, images) are skipped, not serialized.
• PI_MUSE_SANDBOXED=1 or --muse-sandboxed runs Muse with its sandbox and approval prompts enabled instead of --yolo.

Limits

• One muse session per Pi process, not per saved Pi session file. If Pi restarts and you /resume an old conversation, this cold-starts a new muse session and folds once more on the next muse-code turn. That's an acceptable simplification, not a bug: the alternative (persisting the mapping to disk, keyed by Pi's session id) adds real complexity for a case that just costs one extra fold.
• The very first switch into muse-code in a Pi session always arrives as folded text, not a resumed Muse session, because no earlier turn ran through Muse yet.
• Does not install the agents/muse-spark.md subagent definition into ~/.pi/agent/agents/ the way pi-muse-bridge's /muse-setup command did. This package only owns the muse-code chat provider used by /model; Pi's separate Task-subagent delegation feature is out of scope here. agents/muse-spark.md in this repo is just the system prompt used for muse-code chat turns.

Origin

Forked from pi-muse-bridge by ferdousbhai (MIT). 0.1.0–0.2.0 were a companion extension that only hooked Pi's context event to patch pi-muse-bridge's prompts from the outside. 0.3.0 ports pi-muse-bridge's provider/runtime/catalog code directly into this package (still MIT, same author's original design) so it can resume a real muse session natively instead of working around another package's one-shot-only behavior.

Test

node --test test/*.test.mjs

License

MIT