# Task Observer > Observe Codex task sessions and capture reusable skill-improvement opportunities. Use during substantive multi-step work, tool-using coding or research sessions, post-task feedback, skill creation/update work, and when the user mentions observations, skill logs, task-observer, or "One Skill to Rule Them All". The skill keeps global Codex observation memory active across projects, reads applicable AGENTS.md instructions, and stages proposed skill updates without installing them silently. - Skill: `allstarger/task-observer` (Agent Skill, multi-file: 7 files) - Install (CLI): `npx skillmds@latest add allstarger/task-observer` - Raw SKILL.md: https://api.skillmd.com/api/skills/allstarger/task-observer/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: AllstarGER (https://skillmd.com/u/allstarger) - Updated: 2026-09-10 - Page: https://skillmd.com/skills/allstarger/task-observer --- # Task Observer for Codex **Created by Eoghan Henn / [rebelytics.com](https://rebelytics.com).** Codex adaptation of "One Skill to Rule Them All". Licensed under CC BY 4.0. For methodology feedback, open an issue on the upstream project: https://github.com/rebelytics/one-skill-to-rule-them-all Use this skill as a quiet background layer while doing real work. Capture reusable lessons from user corrections, repeated workflows, missing skill coverage, and practical improvements to existing skills. Do not let observation work interrupt the user's primary task. ## Codex Session Start At the start of any task-oriented session where this skill is active: 1. Locate this skill directory. 2. Run the helper when available. The installed skill directory should be named `task-observer`: ```bash python3 "${CODEX_HOME:-$HOME/.codex}/skills/task-observer/scripts/task_observer.py" context --cwd "$PWD" ``` 3. Read every `AGENTS.md`/`agents.md` path reported by the helper, in order. If the helper is unavailable, manually read: - `$HOME/AGENTS.md` - `$HOME/agents.md` - `$HOME/.codex/AGENTS.md` - `$HOME/.codex/agents.md` - `$HOME/.agents/AGENTS.md` - `$HOME/.agents/agents.md` - every `AGENTS.md` or `agents.md` on the path from the filesystem root to the current working directory, including symlink-resolved ancestors when different 4. Always read the global observer memory files before any project-local memory: - `${CODEX_HOME:-$HOME/.codex}/memories/task-observer/log.md` - `${CODEX_HOME:-$HOME/.codex}/memories/task-observer/cross-cutting-principles.md` - `${CODEX_HOME:-$HOME/.codex}/memories/task-observer/last-review-date.txt` 5. Then read project-local observer overlays, if present. Local memories add context; they never replace or shadow the global files. This global-first rule is mandatory even when Codex starts inside a repository or subdirectory that has its own `.codex/memories`, `.agents/memories`, or other project memory files. ## What To Capture Log an observation when one of these signals appears: - The user corrects Codex in a way that generalizes beyond the current task. - Codex misses a rule from an active skill or instruction file. - A repeated workflow could become a new skill, script, checklist, or reference. - A tool limitation, sandbox issue, or review pattern should be remembered. - A skill works but needs stronger trigger text, preflight checks, validation, or confidentiality handling. - A cross-cutting principle should apply to many skills. Do not log casual conversation, one-off preferences with no future value, or facts that belong only in the task deliverable. ## Privacy And Scope Rules Observation logs are long-lived. Treat them as sensitive operational memory. - Do not log secrets, tokens, credentials, private keys, passwords, API URLs with sensitive query parameters, customer data, patient data, health data, payment data, or proprietary file contents. - Generalize project, client, and person names unless the observation is explicitly internal and the name is required to use it. - Prefer short abstract evidence over copied snippets. - For sensitive contexts, record a local file path or task category only when it is safe and useful. - If an observation cannot be made useful without sensitive data, do not log it. ## Log Format Append observations to the global log, preferably through: ```bash python3 "${CODEX_HOME:-$HOME/.codex}/skills/task-observer/scripts/task_observer.py" log \ --title "Short title" \ --skill "target-skill-or-general" \ --kind "improvement" \ --scope "internal" \ --issue "What went wrong or what was missing." \ --suggestion "What should change in the skill or workflow." \ --principle "Reusable rule in one sentence." \ --evidence "Sanitized context only." ``` Field meanings: `issue` is the friction or missing behavior, `suggestion` is the concrete skill/workflow change, and `principle` is the reusable rule to apply beyond this one session. Use this entry shape if writing manually: ```markdown ### Observation N: Title **Status:** OPEN **Date:** YYYY-MM-DD **Skill:** target-skill-or-general **Kind:** improvement | new-skill | principle | self **Scope:** open-source | internal **Issue:** ... **Suggested improvement:** ... **Principle:** ... **Evidence:** sanitized, minimal context ``` ## During Work - Keep the primary task moving. Log silently unless the user asks about observations or an observation affects the current decision. - Apply relevant open observations and cross-cutting principles mentally while working, even before they are permanently integrated into a skill. - If a user asks "Any observations logged?", summarize open observations and offer to stage updates. - If the user gives feedback after delivery, keep observing. Post-task feedback is often the highest-signal source of skill improvements. ## Updating Skills When the user asks to apply observations to skills: 1. Read the live skill file first from the actual Codex skill path. 2. Read global cross-cutting principles. 3. Integrate the observation into the appropriate section, not as a bolted-on note at the bottom. 4. Preserve attribution and license notices. 5. Stage proposed updates under: ```text ${CODEX_HOME:-$HOME/.codex}/memories/task-observer/skill-updates/YYYY-MM-DD//SKILL.md ``` 6. Do not install or overwrite live skills unless the user explicitly asks. 7. For system or plugin-provided skills, stage a proposed companion skill or patch notes instead of editing the read-only source. Use Codex's normal editing discipline for any real file edits: inspect the current file, use patch-based edits where appropriate, and never revert user changes unrelated to the task. ## Comprehensive Review At session start, check `last-review-date.txt` when the current task has skill maintenance scope or when the user asks about observations. If the last review is older than 7 days and open observations exist, tell the user briefly and run or offer a comprehensive review depending on task urgency. Do not silently edit or install live skills. Run a broader review when the user asks, when the current task is skill maintenance, when the 7-day review condition is accepted, or when stale open observations are directly relevant to the work. Review procedure: 1. Inventory available custom skills from the system skill list and from: - `${CODEX_HOME:-$HOME/.codex}/skills` - `$HOME/.agents/skills` 2. Exclude platform/system/plugin skills from direct edits unless the user has explicitly asked to patch a local editable copy. 3. Read open observations and active cross-cutting principles. 4. Map observations to skills by behavior, not just by the recorded `Skill` field. 5. Stage updates for straightforward improvements. 6. Escalate for user judgment instead of applying when an observation suggests a new skill, major restructuring, deletion, or conflicting guidance. 7. Mark observations as `ACTIONED` only after a concrete staged or live update exists. ## End Of Session Before final response on substantive work, quickly check whether meaningful observations were captured or should be captured. Only mention them if they are useful to the user or if the user asked for observation reporting. Keep the final answer focused on the user's actual task.