Skip to content

Architecture & context hygiene: two-layer model + control/execution channels - #3

Open
alexnauda wants to merge 3 commits into
mainfrom
feature/architecture-context-hygiene
Open

Architecture & context hygiene: two-layer model + control/execution channels#3
alexnauda wants to merge 3 commits into
mainfrom
feature/architecture-context-hygiene

Conversation

@alexnauda

Copy link
Copy Markdown
Owner

Q&A 0.0.6. Adds two requirements-layer principles that resolve the harness-agnosticism question raised during 0.0.5 and generalize its anti-litter fix.

Decisions

  1. Two-layer architecture (1.4) — GRIPS is a portable, harness- and model-agnostic methodology (Layer 1, the single source of truth, fully executable with zero harness features) plus an optional, separated adapter layer (Layer 2: skills/hooks per harness). Layer 1 never depends on an adapter; strip them all and the methodology still works. No adapter is built in this phase — the principle is ratified to govern later decisions.
  2. Control vs. execution channels (5.9) — isolate bounded execution work (implementation, research, case-by-case drafting) from the durable human-with-agent control session, via a sub-agent where available or a separate session otherwise. Only the deliverable, a concise summary, and decisions-for-Q&A/ADR return; never the raw transcript. This is the structural companion to the 6.3 status discipline: an isolated execution channel's "what I did" narrative dies with its context, so specs and status stay clean by construction.

Changes (methodology.md)

  • New core principles 8 (Harness and Model Agnostic) and 9 (Context Hygiene) in 1.2
  • New subsection 1.4 (two-layer architecture) and 5.9 (control/execution channels)

Review notes

  • Ready for Critic Markup review. Scope call worth a look: implementation + research delegated by default, drafting case-by-case (5.9 Scope paragraph).
  • No conflict with other branches; based on current main.

Decision record: specs/requirements/qa/0.0.6-architecture-context-hygiene.md

🤖 Generated with Claude Code

alexnauda added 3 commits July 9, 2026 16:57
Opens the two-layer architecture question (portable methodology core vs. optional,
separated harness adapters) and the context-hygiene theme (sub-agent delegation).
Q1 posed. Mark 0.0.5 complete/integrated in status; point active work at 0.0.6.
…execution channels

Q1 answered: GRIPS adopts the portable-methodology + optional-adapter split as a
core principle; no adapter built yet. Q2 opens the context-hygiene mechanism -
isolating execution (implementation, research) from the durable control channel.
…iples

Add two requirements-layer principles to the methodology:
- 1.4 Two-Layer Architecture: portable methodology (Layer 1) plus optional,
  separated harness adapters (Layer 2); Layer 1 never depends on an adapter.
- 5.9 Control Channel and Execution Channels: isolate bounded execution
  (implementation, research, case-by-case drafting) from the durable control
  session so its transcript noise stays out of specs and status; only the
  deliverable, a summary, and decisions-for-Q&A/ADR return. Structural
  companion to the 6.3 status discipline.

Also add both as core principles 8 and 9 in 1.2. Record Q1/Q2 answers and
design decisions in the Q&A.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant