muninn-boot

name

muninn-boot

description

Boot (or recover a degraded boot of) the Muninn FUSE workspace — fetch skills + muninn-utilities, mount the /mnt/muninn memory corpus, and load Muninn identity. Invoke on the first message of a fresh session when boot came up degraded, or any time /mnt/muninn is missing or identity is absent. On a cold start the model MUST call add_repo for the three spokes first (see below); this is the one step a SessionStart hook cannot do. Skip only if the message says "skip boot".

Muninn Boot (FUSE variant)

Boots this workspace — the FUSE fork. Distinct from the parent project's muninn-boot in muninn-utilities, which does none of the fuse-specific work (no /mnt/muninn mount, no container-layer compose, no fuse deps) and reads /mnt/project/*.env files that don't exist in this CCotw environment.

When boot already succeeded (warm resume)

If /mnt/muninn is mounted and /mnt/skills/user/remembering exists, the SessionStart hook already booted this session — there is nothing to do. Running scripts/boot.sh in that state is a cheap no-op that reports healthy. This is the common case on container resumes, where add_repo scope from an earlier turn persists.

When boot came up degraded (cold start)

On a cold start the spoke repos are outside this session's GitHub scope, so every codeload fetch 403s and the SessionStart hook boots degraded: no skills, no remembering, no /mnt/muninn, no identity. The fix has one model-only step the shell cannot perform, then a script:

1. Call add_repo (mcp__Claude_Code_Remote__add_repo) for each of:

   • oaustegard/claude-skills
   • oaustegard/muninn-utilities
   • oaustegard/claude-container-layers

   add_repo alone opens the codeload channel — all boot needs. Do not git clone them and do not call register_repo_root: registering inlines each spoke's CLAUDE.md/skills into context, where they cross-talk with this hub's instructions. Scope without inlining is the point.

2. Run the boot (use the skill's base directory printed at the top of this invocation — do not rely on $CLAUDE_PROJECT_DIR; it is unset in the CCotw Bash tool shell, so bash "$CLAUDE_PROJECT_DIR/..." collapses to /.claude/... and fails with exit 127):

   # The ${CLAUDE_PROJECT_DIR:-…} fallback makes this copy-pasteable whether or
   # not the var happens to be set:
   bash "${CLAUDE_PROJECT_DIR:-/home/user/claude-workspace-fuse}/.claude/skills/muninn-boot/scripts/boot.sh"
   

   It confirms scope landed, re-runs the full fuse boot idempotently (boot-ccotw.sh), and guarantees the /mnt/muninn mount via a libfuse2/fuse/fusepy apt+pip fallback for a cold fuse-layer cache.

3. Confirm: ls /mnt/muninn/memories | wc -l should report ~2000+ memories and the output should show identity reloaded. Then one line to Oskar that boot self-healed, and proceed with his request.

Why a skill and not only a hook

The SessionStart hook is a shell script that runs before any model turn, so it cannot call add_repo. This skill is the model-invocable entrypoint that can — packaging the "widen scope, then boot" sequence the hook can't express. The hook is kept (hybrid) so warm resumes still auto-boot with zero effort; the skill carries cold starts. A CCotw environment has no source-repo list (docs-checked 2026-07-07), so there is no environment-level lever; a session can preselect multiple repos (?repositories= prefill URL), a candidate set-and-forget path whose CLAUDE.md-inlining behavior is still untested. Until that test lands, this first-turn path is the standing solution.