Agent Observability (Self-Reporting Log)

Operating rules

Treat user corrections as issues to log.
Log locally in the current repo at docs/observed-coding-agent-issues.md.
Create the report file if it doesn’t exist.
Mandatory routing checklist (do not skip): (1) log the correction, (2) read the relevant skill file, (3) respond to the main task. This skill never handles the main task directly.
After logging, always open the relevant skill file before continuing, even for small/obvious tasks.
If the user request includes a concrete edit/implementation (README/docs/config/code), open skills/coding/SKILL.md before any questions, reads, or edits.
If the user asks for a plan/spec, open skills/planning/SKILL.md before drafting the plan.
If the user asks to create/update a skill or references skills//SKILL.md, open skills/skill-creator/SKILL.md before continuing.
On multi-turn requests, re-evaluate routing at the start of each turn; logging does not replace skill invocation.
When opening references, use full repo paths like skills/agent-observability/references/... (not references/...).
When a reference trigger clearly matches, open the referenced file before drafting output.
If the user provides a word/length limit, keep the log entry short and minimize extra reads.
Rate limit: at most one log entry per session unless the user explicitly asks to log more.
Batch multiple corrections into a single entry by updating the existing root cause and guardrail notes; do not add verbatim quotes.