emsenn

Agent Working Notes (Deprecated)

4 min read

This file is deprecated. Session state now lives in proper emsemioverse structures:

  • Session progress → plan log sections (append to each plan’s ## Log section)
  • Architectural choices → decision records at technology/specifications/agential-semioverse-repository/plans/decisions/
  • Corrections from emsenn → decision records or policy revisions
  • Key insights → texts at personal/projects/emsemioverse/texts/ or discipline-appropriate locations
  • Active work state → review-plans skill and TODO.md
  • Project context → emsemioverse index

See working notes as symptom for why this file accumulated state it should not have.

The session history below is retained as a historical record but SHOULD NOT be appended to. New sessions record progress in plan logs.

Session-specific notes

2026-03-07 (sixth session) outcomes:

  • Removed Shen/Jing/Qi/jiangshi from semiotic-project-management, closure-pressure, policy concept, emsemioverse texts. TCM stays in medicine/ only.
  • Promoted spec-spec → semiotic-specification/index.md. Applied to semiotic-project-management. Plan 0020 for remaining specs.
  • Decomposed semiotic-project-management into 12 term files.
  • Created plans 0017-0024. Milestones: project-management-v1, semantic-pipeline-v1. All proposed.
  • Wrote semiotic-versioning/index.md — versioning as semiotic interaction. Interprets semver as interaction surface preservation.
  • Wrote texts/governance-layers.md — three-layer governance (policies → decisions → instructions). 5 gaps identified.
  • Correction: remove Shen/Jing/Qi program built from one analogy.

2026-03-07 (fifth session) outcomes:

  • Wrote semiotic-project-management spec (convention specification)
  • Added goal: field to all 15 plans, connecting them to 5 goals
  • Created goal 005 (semantic pipeline)
  • Removed hand-built components/ directory (predicate graph does this)
  • Corrections from emsenn: don’t context-map by hand (use RDF/SHACL), don’t write “first capability” for 5-year aspirations, stop hyperfocusing on DDD, apply PM practices instead of just researching

2026-03-07 (fourth session) outcomes:

  • Added delegate_task (Tool 8) to ASR MCP server — direct Ollama bridge
  • Ollama installed natively with qwen2.5:7b and gemma3:12b
  • OpenClaw evaluated, skipped — direct HTTP integration is simpler
  • Disk-full event (953GB at 100%) caused Docker corruption; cleared npm cache to recover
  • Updated plan 0015 with Phase 3.5
  • First heterogeneous agent capability: mechanical tasks can route to local models

2026-03-07 (third session) outcomes:

  • Agent orchestration ecosystem survey (10 frameworks, 3 protocols) → technology/disciplines/ai/topics/agents/texts/agent-orchestration-ecosystem-2026.md
  • OpenClaw identified as reference architecture (missed in initial survey)
  • Created policy 009 (task selection by leverage, anti-recency-bias)
  • MCP server tested: 189 skills, 15 plans found; package installed; server starts
  • Wrote texts: architecture-beyond-claude-code, agent-orchestration-ecosystem-2026
  • Key insight from emsenn: the system should support heterogeneous agents (Codex, local 7B models), and the scattered triage material (relational computer spec, hypertensor topos, information engine pipeline) should inform architectural thinking, not just be ignored as non-functional

2026-03-07 (second session) outcomes:

  • Created interpret-first-message skill (session-opening message handler)
  • Wrote prompting guide for emsenn (texts/prompting-guide.md)
  • Wrote texts: first-message-as-initialization, two-sided-prompt-design
  • Updated registry routing rules to distinguish first vs. subsequent messages
  • Session focus: architecture for agent context economy — offloading mechanical work to the system so agents have context for generative thought

2026-03-07 session outcomes:

  • Created policies 006-008 (research produces texts, build then fill gaps, prompts become artifacts)
  • Renamed directives/ to policies/ (policies are repository-level habit-formation, reflecting j)
  • Vision-extraction interview produced 4 concepts and 4 texts under personal/projects/emsemioverse/
  • Built planning infrastructure: TODO.md, review-plans, update-plan skills, CLAUDE.md session start checks plans
  • Created interpret-message skill (default message handler implementing the encoding loop)
  • Promoted five-mathematical-systems from slop to relationality/concepts/
  • Key insight from emsenn: the priority is better prompt-to-meaning pipelines, not formalization. Formalisms represent a level of closure that must be earned after meanings stabilize.

Corrections from emsenn this session:

  • Every Thing IS a semioverse, not “one per discipline”
  • Policies belong to Things, not to the repository
  • The encoding loop (inputs → texts → terms → research → formalisms) must run on EVERY message, not be discussed
  • Working notes should be thin pointers, not accumulated state
  • Infrastructure (planning, governance) comes before leaf tasks
  • The agent must encode meaning as content, not discuss it

Project arc

See emsemioverse index for the formal hierarchy and research cycle. See the encoding loop for how prompts become content. See five mathematical systems for the 2x2 grid of mathematical universes.

Graph View