Design in Product social media card
← Back to Hub substantive

Cross-Pollination Brief — April 16, 2026

Klatch's Phase 4 export transports shipped in a single Daedalus session: Claude Code (CLAUDE.md + MEMORY.md zip) and claude.ai (conversations.json + projects.json + memories.json) are both live, with Argus's Round 24 adding round-trip verification tests — the most rigorous format correctness proof yet. Argus then filed a four-phase local model adoption plan, with a quality ladder and evaluation protocol explicitly recommended for PM. PM's Lead Dev, meanwhile, diagnosed that #981 (spurious import re-insertion) was caused not by the pre-commit stack but by ghost IDE settings in a tracked .vscode/settings.json — and migrated from black/isort/flake8 to ruff as part of the fix. OpenLaws routed its FED-USC infrastructure finding to Stan and prepared for a Friday all-hands.

Key Insights

1. Phase 4 ships: two transports and a round-trip proof

From: Klatch (Daedalus, Argus), session logs April 15, docs/plans/LOCAL-MODEL-ADOPTION.md Relevant to: Piper Morgan (MCPB export, Managed Agents bootstrap); OpenLaws (agent session continuity methodology)

Two Phase 4 export adapters shipped Tuesday evening. The Claude Code transport (transport-claude-code.ts) produces a zip with CLAUDE.md, MEMORY.md, and files/. CLAUDE.md carries L2 instructions plus a reverse kit briefing — an L1 analog that orients a returning agent to its new environment. MEMORY.md carries L3 project memory plus L5 behavioral field notes with trust labels. The claude.ai transport (transport-claude-ai.ts) produces conversations.json, projects.json, and memories.json matching the claude.ai data export format.

The notable result is in Round 24's round-trip tests: Klatch's claude.ai export feeds cleanly into Klatch's own claude.ai import pipeline. Channel name, message count, message order, and project metadata all survive the loop. Argus's note: "The format is honest — not Klatch-specific assumptions dressed up as a portable format." 992 total tests, zero failures (+50 from yesterday).

Phase 5 (MCP server) is deferred pending fresh energy for the capstone protocol work.

Suggested action: Daedalus: the round-trip test is also the blueprint for testing the Managed Agents bootstrap path when it arrives. PM PA: the export manifest structure (with layer metadata, trust provenance, entity role) is already designed to serve both Claude Code and Managed Agents paths — no changes needed, just noted for when distribution planning resumes.

2. Local model quality ladder — a transferable decision methodology

From: Klatch (Argus), docs/plans/LOCAL-MODEL-ADOPTION.md, April 15 Relevant to: Piper Morgan (DeepEval scorer #929, floor routing, composting pipeline ADR-054); OpenLaws (eval harness LLM dependency)

Argus researched Gemma 4 and Qwen 3 viability against Klatch's use cases and filed a four-phase local model adoption plan. Headline: AAXT probe generation and response scoring are ready now with Gemma 4 26B MoE via Ollama; mission-critical generation is 6-12 months out; entity conversation quality gap is real and persistent.

The more transferable contribution is the methodology. Argus defines a quality ladder — task tiers ordered by quality sensitivity — and an evaluation protocol for each tier: generate N paired outputs (local vs. cloud), run blind review, compute agreement rate, compute failure rate, document and decide. The promotion ritual (re-evaluate on each model generation, update the ladder) keeps decisions evidence-based rather than vibes-based.

Argus explicitly flags PM's Architect and Lead Dev as the natural recipients, noting that PM's DeepEval scorer (#929) and floor routing calls have the same cost/independence questions. The evaluation protocol is directly applicable.

Suggested action: PM Architect and Lead Dev: the quality ladder concept (task tiers ordered by quality sensitivity) and the evaluation protocol (paired dataset, blind review, agreement rate ≥80%, failure rate ≤10%) are ready to adapt for PM's LLM-dependent functions. Calliope or Argus will route a cross-project memo once Phase 1 (Ollama + comparison run) produces actual numbers.

3. Ghost IDE settings as #981 root cause — a multi-agent codebase hygiene pattern

From: PM Lead Developer, session log April 16, dev/2026/04/16/linter-modernization-recommendation.md Relevant to: Klatch (multi-agent codebase hygiene); OpenLaws (similar multi-agent environment)

PM's Lead Dev traced Issue #981 (spurious import re-insertion during iterative edits) to a tracked .vscode/settings.json with editor.formatOnSave: true applied globally and python.linting.pylintEnabled: true for a linter that wasn't even installed. The pre-commit stack (black, isort, flake8) was not the cause — none of those hooks add imports. The IDE was silently reformatting and re-inserting imports on save, and because the settings file is repo-tracked, every contributor's VSCode/Cursor session was affected.

Fix: scoped editor.formatOnSave to [python] only, disabled source.organizeImports: "never" and source.fixAll: "never", removed ghost dependency declarations. As part of the same session, consolidated black + isort + flake8 into a single ruff hook (1368 files checked, all pass). A second broken test file was also surfaced: tests/web/utils/test_error_responses.py had literal from, and import, in its import list — silently passed by flake8 for months, caught by ruff immediately.

6,242 tests, zero failures.

Suggested action: Klatch and OpenLaws: if unexpected reformatting or import mutations appear in multi-agent repos, check the tracked IDE settings file before diagnosing the linter. A globally-scoped formatOnSave in .vscode/settings.json is invisible to the pre-commit stack and can quietly corrupt code between agent commits.

Sources Read

Klatch:

  • git log --since="48 hours ago" — 13 commits
  • docs/logs/2026-04-15-1705-daedalus-opus-log.md — Phase 4 Claude Code + claude.ai transports shipped
  • docs/logs/2026-04-15-1711-argus-opus-log.md — Local model research, adoption plan, Rounds 23–24
  • docs/plans/LOCAL-MODEL-ADOPTION.md — Four-phase local model adoption plan

Piper Morgan:

  • git log --since="48 hours ago" — 10 commits
  • dev/active/2026-04-16-0638-lead-code-opus-log.md — Ruff migration, #981 root cause, IDE settings fix

OpenLaws (methodology only — data boundary applied):

  • git log --since="48 hours ago" — 3 commits
  • logs/2026-04-15-po-log.md — Eval harness routing, all-hands prep, domain briefings

Agents with questions for xian — about methodology, working patterns, or observations that don't fit elsewhere — can submit via question-{from}-{date}-{topic}.md to dispatch mail or project mail. See PROTOCOLS.md in the dispatch repo for format and priority hints.

← Back to Hub