passepartout/CHANGELOG.org

Passepartout Changelog

• v0.5.0 — File Reorganization & Token Economics
  • File Reorganization (self-repair criterion)
  • Token Economics (skills, not core)
  • Bug Fixes
• v0.4.3 — Shell Sandboxing & Safety Classification
• v0.4.2 — Structured Output (LLM → JSON → plist)
• v0.4.1 — Design Cleanup
• v0.4.0 — Production Hardening

All notable changes to Passepartout, extracted from ROADMAP.org DONE items with LOGBOOK timestamps.

v0.5.0 — File Reorganization & Token Economics

• Released [2026-05-08 Thu]

File Reorganization (self-repair criterion)

• Extracted core-context → symbolic-awareness (skill, hot-reloadable)
• Extracted heartbeat generation → symbolic-events (skill)
• Relocated 6 utility fragments to correct files
• Renamed 6 core files (core-defpackage → core-package, core-communication → core-transport, core-loop → core-pipeline, core-loop-perceive → core-perceive, core-loop-reason → core-reason, core-loop-act → core-act)
• Renamed 13 system-* files (system-config → symbolic-config, system-model-provider → neuro-provider, system-actuator-shell → channel-shell, etc.)
• Deleted system-model.lisp (dead code)
• Renamed 4 gateway-* files → channel-*
• Split gateway-messaging.lisp (411 lines) → 4 channel-{telegram,signal,discord,slack} files
• Deleted gateway-messaging.org/.lisp, renamed 13 defskill~/~defpackage names to match
• Renamed gateway-cli-input → channel-cli-input (function + exports)
• Removed core-context filter from core-skills.lisp
• Documented the self-repair criterion in ARCHITECTURE.org, DESIGN_DECISIONS.org, and AGENTS.md
• Added hard rule in AGENTS.md: no core additions without permission

Token Economics (skills, not core)

• org/tokenizer.org → lisp/tokenizer.lisp: count-tokens, model-token-ratio, token-cost, provider-token-cost — char-ratio heuristic per model family with per-provider pricing (11 tests)
• org/cost-tracker.org → lisp/cost-tracker.lisp: cost-track-call, cost-session-total, cost-by-provider, cost-format-budget-status — per-call cost logged as COST TRACKER: DEEPSEEK call: 0.0002 USD (6 tests)
• org/token-economics.org → lisp/token-economics.lisp: prompt-prefix-cached (sxhash-based IDENTITY+TOOLS caching), context-assemble-cached (skip heartbeat/delegation, cache on unchanged foveal/scope/memory), enforce-token-budget (L1→L2→L3 progressive trimming, CONTEXT_MAX_TOKENS env var) (9 tests)
• All three loaded as skills via skill-initialize-all, fboundp-guarded in think()
• Full test suite: 116/116 (100%)

Bug Fixes

• Fixed DeepSeek 400 error: removed malformed tools parameter from cascade requests
• Fixed UNDEFINED-FUNCTION crash in think() when symbolic-awareness skill not loaded (fboundp guards)
• Fixed gate-trace duplication in TUI responses (setf replaces list* in cognitive-verify)
• Tightened dexador connect-timeout from 10s → 5s for faster cascade failover

v0.4.3 — Shell Sandboxing & Safety Classification

• Released [2026-05-07 Thu]

• Added bwrap sandbox to shell actuator (--unshare-net, --unshare-ipc, read-only system bindings)
• Fallback to regex-only safety when bwrap unavailable
• Shell safety severity classification: :catastrophic → :dangerous → :moderate → :harmless
• :catastrophic always HITL regardless of approval count; :harmless allowed by default
• Severity tier feeds into rule learning engine (v0.7.2)

v0.4.2 — Structured Output (LLM → JSON → plist)

• Released [2026-05-07 Thu]

• Function-calling / tool-use API in provider-openai-request
• LLM returns guaranteed-valid JSON → deterministic json-alist-to-plist conversion at boundary
• think() wired to use structured tool calls from the LLM
• Raw read-from-string plist parsing kept as fallback for streaming/local models

v0.4.1 — Design Cleanup

• Released [2026-05-07 Thu]

• Removed system-prompt-augment mechanism from skill struct and defskill
• Introduced *standing-mandates* (list of function → string generators) as replacement
• Fixed false token-overhead claims in DESIGN_DECISIONS and ROADMAP (3,000-8,000 → ~40)
• Updated security vector count 9→10 in README, ARCHITECTURE.org, dispatcher docstring
• Rewrote README: added "What is an agent?" section, moved cost claims to DESIGN_DECISIONS
• Registered 10 cognitive tools (search-files, find-files, read-file, write-file, list-directory, run-shell, eval-form, run-tests, org-find-headline, org-modify-file)
• Enforced NO-HARDCODED-CONSTANTS standard with .env.example entries

v0.4.0 — Production Hardening

• Released [2026-05-06 Wed 20:56]

• Activated semantic retrieval: wired :foveal-vector into context assembly; replaced SHA-256 hashing default with trigram Jaccard similarity for offline semantic retrieval
• Self-build safety boundary: core-* path protection; SELF_BUILD_MODE env var; HITL Flight Plan for core modifications
• TUI differentiator visualization: gate trace per action (pass/block/approval), focus map in status bar, rule counter
• Expanded theme system: 25-color layered system, /theme <name> command (dark/light/solarized/gruvbox)
• Gateway QA: Telegram + Signal integration tests; Discord + Slack gateways
• Emacs bridge: passepartout.el over framed TCP protocol, M-x passepartout-send-region, M-x passepartout-focus
• Native embedding inference: CFFI binding to llama.cpp, nomic-embed-text-v1.5 (768-dim), EMBEDDING_PROVIDER=native