The Dynamic, Fluid Companion Framework to Interpretable Context Methodology (ICM)
Composable Context Architecture (CCA) is the fluid companion to Interpretable Context Methodology (ICM) (arXiv:2603.16021).
Where ICM excels at structured, repeatable, rigid pipelines (e.g. Soundingboard's 5-stage novel production studio, or multi-phase software delivery), CCA governs the phase before optimization: the messy, non-linear stretch of creative exploration, research, architecture, and knowledge discovery where a rigid pipeline would trap the human in the wrong phase.
AI should amplify human agency, not absorb it.
The human remains the sole source of intent, meaning, judgment, taste, and authorship. The machine brings capability, memory, analysis, challenging questions, structural transformation, and execution. Neither pretends to be the other:
- The human sets direction; the machine helps explore.
- The human makes judgments; the machine provides capability.
- The human can change their mind at any point, and the machine adapts to the changing intent rather than anchoring to the old one.
- The machine can challenge assumptions without owning the decision.
- The work stays inspectable, the process stays attributable, and the human holds the red pen at all times.
-
The Tripartite Chain of Custody:
$$\text{Deterministic Tool Observation} \longrightarrow \text{Agent Craft Framing} \longrightarrow \text{Author Sovereign Decision}$$ Tools make mechanical observations (syntax checks, token metrics, ASTs); agents frame observations into structured options or Socratic questions; the human signs the gate. -
Separation of Technical Integrity from Creative Uncertainty:
-
Technical Integrity: Syntax errors, unclosed brackets, token budget violations (
$>3,000$ tokens), or state leaks intoblackboard.mdare fatal errors that halt the engine immediately. -
Creative Uncertainty: Ambiguous ideas, open plot holes, incomplete requirements, and raw brainstorming are first-class citizens quarantined with brackets (
[unverified],[option A | option B]) and never block progress.
-
Technical Integrity: Syntax errors, unclosed brackets, token budget violations (
-
Lossless Cache Restoration:
The markdown file is the record; the index is a cache. If
.cca/or any cache file is deleted,cca reindexlosslessly reconstructs the system state from disk in under 50ms. -
Zero Runtime Dependencies:
No npm runtime packages (
dependencies: {}). Built strictly with Node.js standard library built-ins (node:fs,node:path,node:child_process,node:crypto). A CCA workspace will run identically a decade from now.
┌────────────────────────────────────────────────────────────────────────┐
│ 1. The Person Layer (~/.config/cca/me.md) │
│ Voice, Constraints, Fluency, Harness Prefs, Auditable Patterns │
│ • Outlives all projects • Read once per session as ambient context │
│ • System can ONLY propose to staging; Human promotes │
└───────────────────────────────────┬────────────────────────────────────┘
│ Ambient Read
▼
┌────────────────────────────────────────────────────────────────────────┐
│ 2. Workspace Primitives (Project-Scoped) │
│ │
│ blackboard.md (Human Thinking Canvas — UNPOLLUTED by Machine Output) │
│ │ │
│ ▼ │
│ context.md (Transient Action Pipe, <3,000 tokens flushed) │
│ │ │
│ ▼ │
│ .cca/pipe.md (Intermediate Stream Buffer between Blocks) │
│ │ │
│ agents.md (Harness, Prompt & Authority Constraint Registry) │
└───────────────────────────────────┬────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────────────┐
│ 3. Composable Cognitive Blocks (26 Unix Filters) │
│ α (Generators) │ β (Inquirers) │ γ (Distillers) │ δ (Verifiers)│
└────────────────────────────────────────────────────────────────────────┘
Scoped per person and sits strictly above all workspace primitives. It survives project boundaries and compilations.
- Sections: Voice & Working Style, Standing Constraints, Domain Fluency Map, Harness Preferences, and Observed Patterns.
- Auditable Staging: The system can only write cross-project observations to a
## Proposedholding area with inline provenance comments (<!-- proposed: YYYY-MM-DD, workspace: X, accepted -->). Deleting removes fact and provenance simultaneously. - Echo Chamber Prevention: Workspace-level
preferences.mdoverrides ambientme.md. An incognito flag (--fresh) enables unconditioned discovery.
-
blackboard.md: The live human thinking canvas. Intermediate block outputs stream through.cca/pipe.mdand are never dumped directly here. Only human-approved milestones or final yields reach the blackboard. -
context.md: Transient, single-purpose pipe mounting strictly minimal data fragments for the active action and flushed cleanly after every use ($<3,000$ tokens). -
agents.md: Declarative registry of agent personas, authority boundaries, tool parameters, and harness targets.
Every skill operates as an interactive, creative, temporary ICM micro-pipeline with its own CONTEXT.md defining token envelopes, schemas, and authority contracts:
| Skill | Type | Core Role | Who Supplies Content | Who Decides |
|---|---|---|---|---|
/spark |
Low-stakes strawman generator to defeat blank page | Machine (Strawman) | Human (Reaction) | |
/dig |
Socratic inquiry rooted in human's exact words | Human (Content) | Human | |
/bracket |
3 orthogonal candidate pathways (A, B, C, D) | Machine (Options) | Human (Selection) | |
/prune |
Constraint sieve & ruthless elimination of bad options | Machine (Sieve) | Human (Culling) | |
/distill |
Invariant extraction & tension crystallization | Machine (Extraction) | Human (Validation) | |
/unroll |
Dependency deconstruction & first principles | Machine (Topology) | Human (Validation) | |
/prism |
Multi-stakeholder archetypal lenses | Machine (Lenses) | Human (Judgment) | |
/bridge |
Cross-domain structural isomorphism | Machine (Analogies) | Human (Inspiration) | |
/tether |
Logic, continuity & assumption falsification | Machine (Audits) | Human (Resolution) | |
/blast |
Boundary mapping & pre-mortem stress testing | Machine (Attacks) | Human (Hardening) | |
write-me |
Person calibration interview | Human (Interview) | Human |
-
The Creative Diamond (
$\beta \rightarrow \alpha \rightarrow \gamma$ ): Ground problem$\rightarrow$ explore widely$\rightarrow$ ruthlessly converge. -
The Red-Team Gauntlet (
$\gamma \rightarrow \delta \rightarrow \alpha$ ): Isolate core claim$\rightarrow$ stress-test to failure$\rightarrow$ generate hardening forks. -
The Perspective Interrogation (
$\alpha \rightarrow \beta \rightarrow \beta$ ): View through outsider eyes$\rightarrow$ isolate deepest tension$\rightarrow$ Socratic probe. -
The Strawman Catalyst (
$\alpha \rightarrow \beta \rightarrow \gamma$ ): Generate intentionally flawed prototype$\rightarrow$ capture human reaction$\rightarrow$ extract invariants.
- Node.js
$\ge 18.0.0$ - Git
Clone the repository and install dev dependencies:
git clone https://github.com/richardelder2/Composable-Context-Architecture.git
cd Composable-Context-Architecture
npm testcca status— Inspect the health of workspace primitives, ambientme.md, and pipe buffer.cca route— Analyzeblackboard.mdand synthesize an appropriate cognitive topology.cca run— Step through a dynamic block pipeline interactively with human gate approvals.cca compile --to=icm— Ingest Git micro-commits, isolate recurring workflows, and scaffold rigid ICM stages.cca me— Inspect, stage, and approve cross-workspace person patterns.cca doctor— Verify invariant compliance, token discipline, and cross-platform integrity.
- Architecture Specification: docs/CCA_SPECIFICATION.md
- Methodology: Companion to Interpretable Context Methodology (ICM, arXiv:2603.16021).
- License: MIT © 2026 Richard Elder & Antigravity