l10n-brazil-agents

Multi-agent pipeline for autonomous migration and refactoring of Odoo l10n-brazil modules. All LLM inference is local via Ollama (qwen3-coder). No external APIs.

Architecture

issue/pr → issue_analyzer → codebase_analyzer
                               ↓
                      migration_agent (l10n-brazil)
                               ↓
                       refactor_agent (perf only)
                               ↓
                   ┌── execution_agent (k8s sandbox) ──┐
                   │           ↓                       │
                   │    debug_agent → fix_agent ───────┘
                   │    (retry loop, max 9 attempts)
                   └──────────────────────────────────
                               ↓
                          test_agent
                               ↓
                           ci_agent → Forgejo PR

All agents share a single AgentState object, checkpointed to JSON after each step.

Requirements

• Python 3.11+
• Ollama running locally with qwen3-coder pulled
• minikube running with l10n-agents namespace (or existing k8s cluster)
• kubectl in PATH and kubeconfig configured

Setup

git clone https://git.pop.coop/popsolutions/l10n-brazil-agents
cd l10n-brazil-agents

pip install -r requirements.txt

cp .config/settings.env.example .config/settings.env

Edit .config/settings.env with your credentials (never commit this file).

Install the pre-push security hook:

cp scripts/security-check.sh .git/hooks/pre-push
chmod +x .git/hooks/pre-push

Ollama setup

ollama pull qwen3-coder
ollama serve

K8s sandbox setup

minikube start --cpus=4 --memory=8g
kubectl apply -f k8s/sandbox.yaml
kubectl get pods -n l10n-agents

Usage

Single module migration

python graph.py \
  --repo /path/to/l10n_br_account \
  --issue "Migrate l10n_br_account from 16.0 to 18.0" \
  --from-version 16.0 \
  --to-version 18.0

Resume from checkpoint

python graph.py \
  --repo /path/to/l10n_br_account \
  --resume /tmp/l10n-agents/ci_agent.json

Run unit tests (no k8s, no LLM needed)

python -m pytest tests/ -v

Output

• Migrated files written in-place to --repo
• Migration scripts at <repo>/migrations/<version>/
• PR opened at https://git.pop.coop/oca-mirror/<module>/pulls/
• Full pipeline summary printed to stdout
• Checkpoints at /tmp/l10n-agents/*.json

Environment variables

All credentials in .config/settings.env (gitignored):

Variable	Description	Default
OLLAMA_URL	Ollama endpoint	http://localhost:11434
LLM_MODEL	Model name	qwen3-coder
FORGEJO_URL	Forgejo instance	https://git.pop.coop
FORGEJO_TOKEN	API token	required
SMTP_HOST	Mail server	mail.popsolutions.com.br
SMTP_PORT	SMTP port	465
SMTP_USER	SMTP user	ia@pop.coop
SMTP_PASS	SMTP password	required
NOTIFY_EMAILS	Alert recipients	lpc@pop.coop,m@pop.coop
K8S_NAMESPACE	Pod namespace	l10n-agents
MAX_RETRIES	Debug loop limit	9

Agent responsibilities

Agent	Input	Output
issue_analyzer	Issue text	Technical plan + affected files
codebase_analyzer	Repo files	Bottleneck map + dependency graph
migration_agent	Files + plan	Migrated code + migration scripts
refactor_agent	Codebase map	Performance-optimized code
execution_agent	Code + scripts	Logs + errors from k8s pod
debug_agent	Errors + logs	Classified error report
fix_agent	Error report	Corrected files
test_agent	Running pod	Test results + perf metrics
ci_agent	Final state	Commit + Forgejo PR

Deterministic rules

knowledge/migration_rules.py contains regex-based rules applied before any LLM call. These handle: api decorator changes, openerp→odoo imports, l10n-brazil field renames, fiscal pattern updates. Add new patterns here first — cheaper and faster than LLM.

Scaling

Each agent pod is isolated in the l10n-agents k8s namespace with CPU/memory quotas. To process multiple modules in parallel, run multiple graph.py instances pointing to different repos. The checkpoint system prevents work duplication on failure.

For datacenter deployment on Proxmox:

• Run Ollama on the bare-metal node with GPU
• Run the agent pipeline as k8s jobs on the blade cluster
• Use the Proxmox storage backend for checkpoint persistence

Security

• All credentials in .config/settings.env — never in code
• Pre-push hook blocks credential patterns
• Whitelist-based .gitignore
• No code executed outside k8s sandbox pods
• Pods auto-deleted after pipeline completes