arcodange-org/factory
1
0
Fork 0
Code Issues 4 Pull Requests Actions Packages Projects Releases Wiki Activity

docs(prd): harness portability — Claude as escalation tier, not hard dependency #28

Merged
arcodange merged 2 commits from arcodange/harness-portability into main 2026-07-15 18:50:24 +02:00
Conversation 0 Commits 2 Files Changed 3
+13 -1
2 changed files with 13 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit 5aea83749d - Show all commits
vibe/PRD/ai-back-office/agent-architecture.md
+1 -1
View File
@@ -185,7 +185,7 @@ To be settled by POC evidence, each closing with a short ADR:
| # | Decision | Options (leaning) |
| --- | --- | --- |
| D1 | Work queue | telegram-gateway's planned Postgres durable queue (**leaning, operator-endorsed 2026-07** — already roadmapped, transactional, one less system) vs. flat files in git vs. Redis |
| D2 | Orchestration runtime | Claude Agent SDK headless for cluster-triggered jobs + **hermes** for M4-side lanes (**leaning** — hermes already runs skills + cron there) vs. bespoke TS orchestrator (erp `test/` Deno codebase) vs. pure CronJobs + scripts |
| D2 | Orchestration runtime | Claude Agent SDK headless for cluster-triggered jobs + **hermes** for M4-side lanes (**leaning** — hermes already runs skills + cron there) vs. bespoke TS orchestrator (erp `test/` Deno codebase) vs. pure CronJobs + scripts. **Operator direction 2026-07-15: the harness itself must be multi-runtime — Claude is an escalation tier, not a hard dependency; the same protocol must run on Mistral or on hermes + Ornith/MLX locally.** Verifier roles migrate first (bounded, read-only — and cross-family refutation is *stronger* verification than same-family); builder roles are benched on a scoped issue before trust ([model fleet › harness portability](model-fleet.md#harness-portability)) |
| D3 | KM monthly invoice firing | enable Dolibarr template auto-fire (`frequency>0`) vs. agent-fired via sandbox+promote (**leaning** — keeps the gate + mention audit in-line) |
| D4 | PA — e-invoicing platform (*plateforme agréée*, ex-PDP) | **Leaning: Qonto** (operator direction, 2026-07 — the capital-deposit bank, DGFiP-registered PA, e-invoicing included in every plan, and the fleet's richest existing API integration); POC-6 validates reception + API pull before the ADR — **must close before 2026-09-01** ([C12](challenges.md#c12--e-invoicing-reform-unknowns)) |
| D5 | OCR provider for scanned docs | Mistral OCR (EU cloud) vs. local vision model on M4 vs. Tesseract baseline |
vibe/PRD/ai-back-office/model-fleet.md
+12
View File
@@ -44,6 +44,18 @@ The format guarantee never rests on the model:
…and regardless of tier, every payload passes the same deterministic validators downstream ([agent architecture](agent-architecture.md#atom-contract)). Constrained decoding guarantees *shape*; validators guarantee *truth conditions* (arithmetic, checksums, plausibility).
## Harness portability
The **atoms** above are multi-model by construction. The **harness** — the orchestration layer that executes backlog issues (builder sessions), verifies them (cold readers, locate-tests) and files PRs — ran its first full loop on Claude (erp#38 → PR erp#62, 2026-07-15). **Operator direction (2026-07-15): that layer must not have Anthropic as a hard dependency either — the same loop should run on Mistral, or on hermes + Ornith / MLX-served local models.**
The protocol was built for this: everything an executing agent needs lives in **files and issues** — AGENTS.md orientation, per-issue Execution footers (first command, done-means), acceptance tests defined in the issue itself, deterministic validators around every output. Nothing load-bearing lives in a vendor's context window. Any runtime that can read the forge and run a shell can attempt an issue; the acceptance gates don't care who built the thing.
Migration ladder (evidence-gated like everything else — erp#63):
1. **Verifier roles first** — locate-tests, cold-reader audits, refutation passes: bounded, read-only, cheap. Re-run the already-passed erp#38 locate-test on **Ornith** (hermes, `127.0.0.1:18080`) and on an **MLX-served candidate** (Qwen3 14B/30B-A3B, Mistral Small 3.x class); require verdict parity with the Claude run. Once parity holds, **prefer cross-family verification**: a different model family refuting the builder is stronger evidence than the builder's family agreeing with itself.
2. **Scoped builders benched** — one small, tightly-footered issue (a #56-class fix) attempted end-to-end by a Mistral-driven runtime against the unchanged acceptance gates; measure completion rate and intervention count.
3. **Claude stays the default** for multi-file builders, ambiguity and fiscal reasoning **until the bench says otherwise** — a routing-policy fact (complexity ceiling above), not an allegiance.
## Degraded modes
| Outage | Keeps working | Queues | Lost until recovery |