Todd Thomas todd-dsm

Cascadian vs Symphony — Feature Comparison

Two-directional comparison. Part 1 scopes to Symphony's feature set (github.com/openai/symphony — SPEC.md + the Elixir reference implementation, analyzed from a local checkout 2026-07-12), rating Cascadian against it. Part 2 reverses the question: a strict tally of Cascadian's feature set rated against Symphony, with coverage percentages. Ratings throughout: ✓ has it, ≈ partial/different mechanism, ✗ absent. In Part 1, * = coded in Symphony but undocumented in any of its doc surfaces (easter egg).

Part 1 — Symphony's feature set, Cascadian rated

Architectural frame: Symphony is a standalone daemon that polls an issue tracker (Linear) and drives an external coding agent (Codex app-server) in isolated per-issue workspaces. Cascadian is an in-harness process plugin — the orchestration (PE/ONP), quality gates, and issue lifecycle run inside Claude Code itself. Many ✗ rows are architecture, not absence of thought.

Solution Requirements

Issues	Analyst-x	Analyst-y	Solutions
	(HRIS-1)	(HRIS-2)
Expensive Licensing	[x]	[x]	Reduces licencing costs
HR staff complains about data accuracy	[x]	[x]	data is now complete and accurate
Applicants don't like it	[x]	[ ]	Better UX and Data-entry
Manual On/Offboarding	[ ]	[x]	batch processing excel upload
IT & Device Management Bottlenecks	[x]	[ ]	EZ Mobile Device Management

Frequency	Placement	Example
Every task	Auto-loaded (L1)	Core principles, project identity
Every session	Session init	Current state, active plan
Per domain	Agent/skill prompt	Domain patterns, standards
Per decision	On-demand	Decision logs, lessons learned
Rarely	External	Vendor docs, API references

Layer	Bad (pointer only)	Good (motivating)
Metadata	"Handles PDF files"	"Extracts text, fills forms, merges documents -- 12 validated patterns for common failures"
SKILL.md	"See FORMS.md for details"	"Form filling has 3 failure modes that require specific handling. See FORMS.md"
Reference	(terminal)	(terminal)

Level	Speed	Size	Agent Analog
L1 Cache	Fastest	Smallest	Auto-loaded principles (CLAUDE.md)
L2 Cache	Fast	Medium	Session context (quickstart.md, active plan)
L3 Cache	Moderate	Larger	Domain knowledge (agent prompts, standards)
Main Memory	Slower	Large	Reference files (decisions, lessons, docs)
Disk	Slowest	Largest	External resources (MCP tools, web)

Approach	Framing	Reliability
Rule-based	"Read infrastructure-patterns.md before writing Terraform"	Agent may skip if confident, may not notice the instruction
Goal-aligned	"15 validated patterns exist that prevent failures you'd otherwise encounter"	Creates awareness of a gap, which drives retrieval

Factor	Human Analog	Agent Manifestation
Limited working memory	~7 chunks (Miller, 1956)	Context window capacity
Session boundaries	Sleep resets executive state	Each session starts with no episodic memory
No introspection	Unconscious incompetence	Cannot "know what it doesn't know"
Decision fatigue	Ego depletion (Baumeister)	Every retrieval decision is a decision that might be skipped
Recency bias	Availability heuristic (Tversky & Kahneman)	Recently loaded context dominates

	{
	"Container": {
	"Name": "$containerName",
	"UID": "13ed01be-f035-41bb-9307-76fd06ddad9c",
	"HASH": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
	"provenance": {
	"source": "https://github.com/tukaani-project/xz",
	"version": "1.0.0",
	"proximity": "2"
	}

	```
	panic(cpu 0 caller 0xfffffe0024b36cf0): busy timeout[1], (60s): 'AppleAPFSMedia' (a,4020001) @IOService.cpp:5829
	Debugger message: panic
	Memory ID: 0x6
	OS release type: User
	OS version: 24G90
	Kernel version: Darwin Kernel Version 24.6.0: Mon Jul 14 11:30:29 PDT 2025; root:xnu-11417.140.69~1/RELEASE_ARM64_T6000
	Fileset Kernelcache UUID: 0843B4A77BE94220E9BFBC6024B71273
	Kernel UUID: D013020A-F19B-3730-8283-F7E47107D27A
	Boot session UUID: 8B5D8AA0-0E37-4BF4-A2CA-0F2AFDBBAB59

	~/.config/rsync/backups 2>&1 \| tee /tmp/backups.log

	Running LIVE backup

	Backing up special files:
	+ IFS=,
	+ read -r program source_path
	+ [[ -z # backup random files and follow the format; $HOME is assumed ]]
	+ [[ # backup random files and follow the format; $HOME is assumed = \#* ]]
	+ continue