Work Records / validate (push) Has been cancelled

Details

BINKY-WP-0006: retarget Railiance rhythm to llm-connect/OpenRouter

Reject Claude Code (or any coding agent) on railiance01. Mail-scan stays
deterministic; triage/briefs use llm-connect HTTP → structured JSON →
deterministic apply. Document in integrations/railiance-llm-rhythm.md and
rewrite WP-0006 tasks accordingly; timers no longer require_claude.

2026-07-22 00:00:04 +02:00

4.1 KiB

Raw Blame History

Railiance rhythm via llm-connect (not a host coding agent)

Status: design note 2026-07-21 (BINKY-WP-0006).
Supersedes the residual “install Claude on railiance01” idea from the BINKY-WP-0004 cutover notes.

Principle

Railiance executes infrastructure + hosted LLM inference. It does not host Claude Code, Codex, or another interactive coding harness.

Layer	On Railiance	Not on Railiance
Secrets	OpenBao AppRole / in-cluster Secret	Paste into chat
Deterministic I/O	email-connect scan, git commit/push with deploy key	—
LLM	llm-connect → OpenRouter (open-weights / cheap models)	Claude CLI agentic session
Policy	Allow-listed apply steps (Blue/Green)	Model free-form shell

This matches agent-harness INTENT (session policy shell) and llm-connect INTENT (provider abstraction): the server profile uses llm-connect HTTP; laptop coding agents remain optional for interactive work only.

Mail path (canonical)

timer / CronJob
  → agent-harness mail-scan          # deterministic, AppRole IMAP
  → agent-harness mail-triage        # NEW: llm-connect JSON + apply
  → hub progress + local commit

Phase 1 — scan (exists, proven)

Module: agent_harness/mailscan.py
Creds: tenants/binky/company-email/imap via AppRole
Output: mailmeta/reports/email-channel-evidence-report-*.csv
Hub: binky_mail_intake with counts only

Phase 2 — triage (to build)

Prompt (metadata only): newest CSV rows + current mail-log tail + queue templates.
llm-connect POST (same contract as activity-core llm_client.py):
- URL: http://llm-connect.activity-core.svc.cluster.local:8080
- Provider: OpenRouter (already configured on railiance)
- Model: cost-efficient; prefer open-weights (pick at T03; examples to evaluate: Llama/Qwen/Mistral class on OpenRouter — measure quality on fixture CSV before locking).
Response: strict JSON, e.g.

{
  "log_entries": [
    {"date": "2026-07-21", "from": "…", "subject": "…", "action": "ignore|queue|suspicious"}
  ],
  "queue_items": [],
  "notes": "one-line summary"
}

Apply (deterministic Python): append mail-log; add queue YAML only if schema-valid; suspicious → log only; git commit; hub event.
Never pass message bodies to the model if the scanner can omit them; never execute model-suggested shell.

Daily / weekly (same shape)

Slot	Deterministic inputs	LLM output	Apply
Daily brief	queues, open risks, yesterday brief	structured sections	write `briefs/YYYY-MM-DD-daily-brief.md` + commit
Weekly review	briefs week, SuccessMilestones	structured prep	write weekly prep file + commit

No tool-using coding agent required if apply is code-owned.

What already runs on railiance01

Component	Status (2026-07-21)
`llm-connect` Deployment in `activity-core`	Running; OpenRouter; ConfigMap model `google/gemini-2.5-flash` (may change)
OpenRouter API key Secret	Present (operator custody)
`agent-harness mail-scan` on host	Proven
Host systemd timers	Interim cadence; must call llm path, not `require_claude`
Claude CLI on host	Must not become a dependency

Anti-patterns

Installing Claude Code / npm agent stacks on railiance for rhythm.
Defaulting runner.run_task to AgenticClaudeCodeAdapter for server Jobs.
Sending full email bodies or secrets to OpenRouter.
Dual “coding agent policy” vs “server policy” that drifts (same JSON schema and apply code for all environments).

Implementation homes

Work	Repo
HTTP adapter + mail-triage command + tests	agent-harness
Model/profile ConfigMap, smoke	llm-connect / activity-core overlay
Timers, binky task prompts, dogfood	binky-control (this doc + BINKY-WP-0006)

4.1 KiB Raw Blame History