Task Observer

Task Observer

Capture sanitized, reviewable observations from real work so shared skills and portable workflow lessons can improve without turning every task into a mandatory observation run.

This skill is inspired by Eoghan Henn / rebelytics.com's one-skill-to-rule-them-all task-observer methodology and the Codex-native adaptation by AllstarGER. The upstream project is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Preserve attribution when copying or adapting this skill:

• Original methodology: https://github.com/rebelytics/one-skill-to-rule-them-all
• Codex adaptation evidence: https://github.com/AllstarGER/one-skill-to-rule-them-all

Activation

Use $task-observer only when the user requests observation, asks whether a session produced reusable workflow lessons, or asks to review/apply open observations.

Do not:

• inject this skill into every workflow or batch-worker prompt by default;
• treat this skill as a prerequisite for $pr-batch, $verify, $autoreview, or any other shared workflow;
• rely on this skill to load or trigger other skills; or
• continue observing after the user asks to stop.

Host-sensitive behavior must be availability-checked. Prefer the helper in this skill when it exists. If a host cannot write files or expose a persistent memory path, produce a short handoff summary instead of inventing a platform-specific storage location.

What To Capture

Capture only compact, sanitized observations that point to reusable process improvement:

• user corrections that reveal missing or unclear skill rules;
• repeated manual steps that could become a portable helper or checklist;
• workflow gaps where UNKNOWN, degraded coordination, validation, or review state needed clearer handling;
• simplification opportunities where a skill can remove ceremony or reduce false positives;
• self-improvement notes about this observer skill's own activation, privacy, or staging behavior; and
• cross-cutting principles that belong in shared pack guidance.

Prefer the smallest observation that can drive later review. Do not copy raw files, private issue text, customer examples, logs, stack traces, or proprietary content into observation memory.

Privacy Rules

Never store:

• secrets, tokens, passwords, private keys, session cookies, or credentials;
• customer, patient, payment, cardholder, health, diagnosis, prescription, or other regulated data;
• private URLs that include sensitive query parameters;
• proprietary file contents, private prompt text, or non-public source snippets;
• full GitHub issue, PR, review, or comment bodies; or
• anything the user marked confidential or temporary.

When a useful signal is mixed with sensitive material, discard the sensitive material and write only a generic pattern, such as "review helper should treat missing private coordination status as UNKNOWN."

Memory Helper

The optional helper writes under a safe observation path, defaulting to:

${CODEX_HOME:-${CLAUDE_HOME:-$HOME/.codex}}/memories/task-observer

If CODEX_HOME is unset and only a Claude home exists, the helper uses ${CLAUDE_HOME:-$HOME/.claude}/memories/task-observer.

Use it for local runs:

TASK_OBSERVER_SKILL_DIR="${TASK_OBSERVER_SKILL_DIR:-.agents/skills/task-observer}"
if [ ! -x "$TASK_OBSERVER_SKILL_DIR/bin/task-observer" ]; then
  for agent_home in "${CODEX_HOME:-}" "${CLAUDE_HOME:-}" "$HOME/.codex" "$HOME/.claude"; do
    [ -n "$agent_home" ] || continue
    if [ -x "$agent_home/skills/task-observer/bin/task-observer" ]; then
      TASK_OBSERVER_SKILL_DIR="$agent_home/skills/task-observer"
      break
    fi
  done
fi

"$TASK_OBSERVER_SKILL_DIR/bin/task-observer" init
"$TASK_OBSERVER_SKILL_DIR/bin/task-observer" status --json
"$TASK_OBSERVER_SKILL_DIR/bin/task-observer" append \
  --kind skill-improvement \
  --skill pr-batch \
  --summary "Worker handoffs should preserve degraded private state as UNKNOWN." \
  --source "session-note"
"$TASK_OBSERVER_SKILL_DIR/bin/task-observer" list

The helper appends observation stubs only. It does not edit live skills, workflows, documentation, or memory registries outside its own observation directory.

Staged Update Behavior

Observations are recommendations, not applied changes. When reviewing open observations:

1. Re-read the target skill, workflow, helper, or docs page from the current checkout before proposing edits.
2. Group observations by target and discard stale, duplicate, repo-specific, or low-value notes.
3. Convert durable, portable workflow lessons into docs/solutions/ entries only when they satisfy that library's criteria.
4. Stage any skill or workflow edits as normal repo changes and wait for the user's explicit request before overwriting live installed skills or personal memory.
5. Run the relevant helper tests and .agents/bin/validate before publishing changes.

Never overwrite installed skills, user-global skills, or live personal memory as a side effect of observation review. A direct user request is required before applying staged recommendations outside the current repo worktree.

Relationship To docs/solutions/

Task-observer memory is short-lived working evidence. docs/solutions/ is the durable shared lessons library. Promote an observation into docs/solutions/ only when the lesson is portable across consumer repositories, supported by repeatable evidence, and stable enough to outlive a single session.

Repo-specific command choices, release tracker state, customer context, and one-off memory notes stay out of docs/solutions/.

Closeout Checklist

• Observation memory contains only sanitized summaries.
• Any proposed edits remain staged in the current worktree until explicitly approved.
• Host-sensitive paths, metadata, and helper availability are reported as UNKNOWN when not verified.
• Validation evidence names the exact helper tests and repo gate that ran.

On Cursor, memory and session-start hooks stay availability-checked. Do not treat Claude Desktop archive or Codex goal monitors as Cursor surfaces.