agents-stack v2.0 — spiral routing, method layer, and a 16→7 phase reduction

What changed

agents-stack v2.0 replaces the v1.6.3 linear pipeline (16 phases: 8 worker phases + 8 infrastructure phases) with a 7-phase spiral architecture organized into three layers.

New architecture

Direction Layer:   thesis → challenge
Method Layer:      response → synthesis    ★ new in v2.0
Action Layer:      contract → build → audit

Two new capabilities

Method Layer (response → synthesis). In v1.6.3, the pipeline jumped directly from adversarial review to execution with no design stage for deciding how to respond to gaps revealed by the challenge. v2.0 inserts a method layer that lets the system change its approach before committing to a build contract.

Spiral turn. In v1.6.3, audit could only return PASS or FAIL — bugs meant retrying the same contract. In v2.0, audit can surface deeper insight (“this result reveals a problem with the underlying thesis”), triggering a return to the Direction Layer at increased depth (depth++). The pipeline can now discover that a defect traces back to a wrong assumption, not just bad code.

Phase merge/split summary

v1.6.3	v2.0	Change
`project-initializer`	`init.sh` (bash)	Replaced — doesn’t need an LLM
`generator-brainstorm` + `generator-proposal`	`thesis`	Merged — same cognitive operation
`evaluator-contract-review`	`challenge`	Same gate, delegates judgment to oracle/council
`generator-execution`	`contract` + `build`	Split — two distinct mental operations
`adversarial-live-review`	`audit`	Same invariant, 76% shorter SKILL.md
`state-update`	Orchestrator routing	Absorbed — mechanical state sync
`compound-capture`	`plan.md` Lessons	Absorbed — cross-sprint learning is a paragraph

Removed subsystems

using-agents-stack/scripts/ — 3,138 lines of Python validation/dispatch scripts (logic now embedded in the orchestrator)
6 Python maintenance scripts
12-tier evidence priority system → simplified to 4 tiers
roadmap.md, current-focus.md, ideas.md → merged into three sections of plan.md
Task decomposition JSON with symbol declarations (oracle review found limited real value after LLM generation)

Principle

Every deletion traces back to a single question: does this step need an LLM agent, or is it just ceremony? Moving files to archive folders, appending a lesson to a document, running dispatch validation — these are filesystem operations, not reasoning tasks. v2.0 treats “no phase” as the default; an LLM worker only exists when the task genuinely requires reasoning under uncertainty.