Claude-Mind: Skill for Claude Code

/drain-temp — Temp Store Drain Engine

Encodes the value out of agents/<agent>/temp/ into the one long-term retrieval surface (the knowledge tree + supplementary stores), then archives each processed file to temp/drained/. This is the mechanism that makes temp/ a STAGING area rather than a slush directory: every working doc either drains into knowledge or is discarded — nothing lives in temp/ permanently.

Hybrid skill: the user invokes it ("drain temp", "encode everything in temp"); the aspirations loop invokes it when aspirations-precheck flags temp_drain_needed (temp/ accumulated past the drain threshold). Writes to the tree / reasoning bank / guardrails — requires assistant or autonomous mode.

Why this exists

A working document written to temp/ is in none of the retrieval stores until it is drained — invisible to /prime and retrieve.sh. Draining is how that knowledge becomes durable and findable. See core/config/conventions/temp-store.md.

Phase 0: Load Routing Context

Bash: load-conventions.sh temp-store learning-routing
→ Read the returned paths not already in context. learning-routing.md carries the
  "where does this learning go?" decision tree used in Phase 2 classification.

Phase 1: Enumerate Undrained Files

1. Resolve the bound agent's temp dir via the canonical path helper (never
   hardcode an agent name — _paths.sh exports AGENT_DIR for $MIND_AGENT):
   Bash: source core/scripts/_paths.sh; TEMP_DIR="$AGENT_DIR/temp"; echo "$TEMP_DIR"
2. List undrained working docs — files DIRECTLY under temp/ (NOT under drained/):
   Bash: ls -1 "$TEMP_DIR"/*.md "$TEMP_DIR"/*.json 2>/dev/null
   (drained/ is the archive subdir — never re-drain it; the glob above does not
   descend into it.)
2b. ALSO list the THIRD-CLASS residue — everything matching NEITHER the
   *.md/*.json glob above NOR Phase 1.5's ephemera suffixes:
   Bash: find "$TEMP_DIR" -maxdepth 1 -type f ! -name '.*' \
       ! -name '*.md' ! -name '*.json' ! -name '*.log' ! -name '*.txt' \
       ! -name '*.py' ! -name '*.sh' ! -name '*.err' ! -name '*.raw' \
       ! -name '*.out' ! -name '*.bak*'
   These are NOT auto-drained; CLASSIFY each one in Phase 2 alongside the docs
   (same routing table, same 3b probe where the shape is probe-able).
   WHY AT CENSUS TIME AND NOT LATER: Phase 2.5's stamp gate condemns exactly
   this population, so meeting it there means meeting it as a list of files you
   are about to DELETE. Measured twice on the same box (2026-08-27 and
   2026-08-30, echo/cc-03): both passes correctly retracted the stamp per
   guard-4864, and substantially the same five files came back — because a
   correctly-refused stamp protects them from deletion while nothing ever
   classifies them, so they can never be legitimately discharged either. Two
   correct mechanisms composing into a permanent blind spot. Retracting is safe
   and INSUFFICIENT; this step is what lets the stamp pass on merit. The 08-30
   pass classified all five in minutes and one turned out to be a working
   compile gate that existed nowhere else on the box. (rb-9792, g-001-104.)
3. IF no drainable files (.md/.json): set docs_count=0 and skip Phase 2-3, but
   STILL run Phase 1.5 — pure ephemera (.log/.txt) may need purging even when no
   docs remain (the g-115-1727 case: 7 ephemera survived a full doc-drain).
   Report "temp/ is clean" + DONE only if Phase 1.5 ALSO purges nothing.
4. IF --file <name> given: restrict the list to that one file under "$TEMP_DIR"
   and SKIP Phase 1.5 (single-file drain is a targeted op, not a full sweep).
5. Sort oldest-first (timestamped filenames sort lexically = chronologically).