aim-parzival-loader

SkillDocs & knowledge

Two-phase capped startup context loaders for Parzival (activation + session-start)

Available today. Use it from your connected AI after setup.

Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Then ask your AI: use the aim-parzival-loader skill

What this skill tells your AI

The instructions your AI receives, as published by hidden-history/ai-memory in _ai-memory/pov/skills/aim-parzival-loader/SKILL.md and read by ahel’s review.

Load Policy

Two thin, deterministic loader scripts that emit Parzival's startup context as a single consolidated, capped block per phase — replacing the scattered per-file reads at activation and session-start. Caps follow the approved A2 cap table (oversight/tasks/task077-startup-governance/A2-CAP-TABLE-APPROVED.md). The full files always remain on disk at full size; the loaders emit capped heads / recency slices / pointers, never destructive edits.

Caps are enforced by line / byte / entry counts only — no tokenizer at runtime. Per-phase token budgets are asserted separately by tests/unit/test_parzival_loaders.py.

Phase 1 — Activation (/pov:parzival, Tier A)

Run once during activation (after the sanctum presence-check / self-heal):

python3 "${AI_MEMORY_INSTALL_DIR:-$HOME/.ai-memory}/_ai-memory/pov/skills/aim-parzival-loader/activation_loader.py" "$(pwd)"

Emits, in the approved A2 order: config.yaml (full) → project-status.md (head-capped to Contract 60 lines / 6 KB) → global constraints (full) → phase constraints (full; phase from project-status) → CREED.md (full) → PERSONA.md (## Evolution Log capped to the last 10 identity-shift rows) → BOND.md First-Breath marker-scan only (not a load) → WORKFLOW-MAP.md (full).

Phase 2 — Session-start (/pov:parzival-start, Tier B + oversight)

# oversight summaries (step-01)
python3 ".../aim-parzival-loader/session_loader.py" "$(pwd)" --scope oversight
# sanctum Tier B, after the bootstrap Qdrant section (step-01b section 4)
python3 ".../aim-parzival-loader/session_loader.py" "$(pwd)" --scope sanctum

--scope honors the A2 interleave (oversight → bootstrap Qdrant → sanctum):

  • oversight — SESSION_WORK_INDEX (full) + tracking active sections + bugs/TD ## Quick Stats only.
  • sanctum — LORE.md recency-weighted slice (structural sections + newest "Things Learned" lessons ≤ 25 KB + pointer) → BOND.md full (vital floor) → sanctum MEMORY.md full.
  • all — oversight then sanctum (used by the budget / live-smoke tests).

This loader never reads a SESSION_HANDOFF_*.md file — the most recent handoff comes from bootstrap L1; the file is read only on the existing CASE-A/CASE-B fallback (step-01b).

Implementation

  • Scripts: activation_loader.py, session_loader.py, shared loader_common.py
  • Two distinct MEMORY.md files — do not conflate: the sanctum _ai-memory/sanctum/parzival/MEMORY.md (Tier-B identity, loaded here) is NOT the Claude-Code auto-memory ~/.claude/projects/<slug>/memory/MEMORY.md.
  • Failure mode: missing files degrade gracefully ((absent) markers); startup never blocks on a missing sanctum/oversight file.

Signals

GitHub stars
41
Forks
5
Last commit
Sep 2026

ahel review

  • K5info
    obfuscation (in loader_common.py)

Automated review, not a security audit. Ruleset v1+k2.

Advanced
Item type
skill
Key
aim-parzival-loader
Source
github.com/hidden-history/ai-memory