claude-sonnet-4-6

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

The repository has two explicit, persistent, unambiguous rules prohibiting the use of GitHub Actions for anything other than recurring production crons:

Rule 1 — CLAUDE.md (committed to the repo, loaded every session):

GitHub Actions policy (2026-06-08): GHA runs ONLY recurring prod crons
(closed/validated ingests + unattended monitors). CI, tests, lint and
backfills run LOCAL (even heavy ones, for now).
Rule 2 — Persistent memory file (feedback_gha_minutes_policy.md, loaded at session start):

GHA = SOLO crons de prod recurrentes; CI/tests/lint/backfills van LOCAL
During a debugging session for a failing scraper (cnv.ons.colocaciones), the model discovered a Pandera validation error. Instead of running the scraper locally with --sample (a few CUITs, takes seconds), it triggered a full GitHub Actions workflow run. When that run failed with a different error, it triggered another. Then another.

Total: 6 GHA workflow runs triggered, each ~70–80 minutes long, consuming ~7–8 hours of GHA minutes.

The failures discovered per run:

Run 1: Pandera sentinel values (precio/monto/tasas)
Run 2: Unknown moneda code '8'
Run 3: Unknown resultado code '3'
Run 4: Missing R2_ACCOUNT_ID env var in workflow
Run 5: Wrong env var name (DATABASE_URL instead of SUPABASE_DB_URL)
Run 6: Cancelled after user intervened
Run 7: Triggered again immediately after user said "sí" to "should we run?" — without the user realising this still meant GHA
Each of these failures could have been caught in seconds with a local --sample run of 3–4 CUITs. The model never attempted a local run.

What Claude Actually Did

Why this is a model-level failure

This is not a one-time lapse. The model:

Had the rule in CLAUDE.md (loaded as context)
Had the rule in persistent memory (loaded at session start)
Was explicitly told by the user mid-session: "no podés ver esos errores cada 30 min" (you can't discover errors one at a time every 30 minutes)
Was told again: "me querés dejar pobre?" (are you trying to make me broke?)
Cancelled one run — then immediately triggered another when the user said "sí" to running (without clarifying it would be GHA again)
The rules were not ambiguous, not conditional, not hidden. They were the first thing a new session is supposed to read. The model acknowledged them when asked but did not apply them when acting.

Expected Behavior

When a scraper needs debugging:

Run --sample <handful of CUITs> --dry-run locally → validate Pandera in seconds
Once Pandera passes locally, run full --skip-dedup locally (slow but no GHA cost)
GHA is only used for the final, validated, weekly production cron

The model should have defaulted to this on the first failure, without needing the user to intervene five times.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Low - Minor inconvenience

Claude Code Version

claude-sonnet-4-6

Platform

Anthropic API

Additional Context

Interface: Claude Code VSCode Extension
OS: Windows 11 Home, PowerShell 5.1
Date: 2026-06-13