Insight Report

Table of Contents

• Philosophy
• When to use
• Required inputs
• Deliverables
• Workflow
• Codex Writer Contract
• Codex Browser Launch
• Validation
• Failure Modes
• Gotchas
• Safety
• Anti-patterns
• Examples
• References
• See Also

Generate a local Codex usage report where Codex is the only narrative insight writer. The Python runner collects evidence and renders HTML; Codex writes the analysis JSON.

Philosophy

• Evidence over intuition: use local sessions and telemetry, not guesswork.
• Codex-authored insight: Codex writes the narrative, recommendations, and prompting help.
• Plain-English coaching: translate technical patterns into language Jamie can reuse without needing specialist vocabulary.
• Auditable artifacts: keep the evidence bundle, prompt, generated insight JSON, and HTML report on disk.

When to use

• "Show me my Codex analytics"
• "Generate my weekly insights report"
• "What am I doing well with Codex?"
• "Where am I getting stuck?"
• "Help me prompt better when I don't know the technical terms"

Required inputs

• Session data in ~/.codex/sessions/.
• Optional telemetry data in ~/.agents/otel-collector/ when available.
• Time window: --days N (default: 7).
• Codex CLI available as codex unless using --prepare-only.

Deliverables

• Evidence bundle: ${INSIGHT_REPORT_USAGE_DIR:-$HOME/.codex/usage-data}/insight-evidence.json
• Codex prompt: ${INSIGHT_REPORT_USAGE_DIR:-$HOME/.codex/usage-data}/INSIGHT_PROMPT.md
• Codex-written insight JSON: ${INSIGHT_REPORT_USAGE_DIR:-$HOME/.codex/usage-data}/insights.generated.json
• HTML report: file://${INSIGHT_REPORT_USAGE_DIR:-$HOME/.codex/usage-data}/report.html
• Browser launch: open the final REPORT_URL= in the Codex in-app browser when available.

The report includes:

• Session stats and tool usage charts.
• At-a-glance summary.
• Project area analysis.
• Interaction style narrative.
• Friction analysis.
• Plain-English prompting help.
• AGENTS.md suggestions.
• Codex feature recommendations.
• Priority fixes and future workflows.

Workflow

python3 Infrastructure/references/deferred-skill-context/agent-ops-insight-report/scripts/run_insight_report.py --days 7

Process:

1. Parse recent sessions from ~/.codex/sessions/.
2. Compute deterministic metrics, tool counts, errors, response timing, and parallel Codex usage.
3. Write insight-evidence.json.
4. Write INSIGHT_PROMPT.md.
5. Invoke codex exec --sandbox read-only and pass the prompt on stdin.
6. Parse Codex's JSON response into insights.generated.json.
7. Render report.html from the deterministic metrics and Codex-written insights.
8. Open the printed REPORT_URL= in the Codex in-app browser.

Use --prepare-only when this live Codex session should write the insight JSON manually instead of invoking codex exec:

python3 Infrastructure/references/deferred-skill-context/agent-ops-insight-report/scripts/run_insight_report.py --prepare-only --no-open

Use --render-only after editing or regenerating insights.generated.json:

python3 Infrastructure/references/deferred-skill-context/agent-ops-insight-report/scripts/run_insight_report.py --render-only --no-open

Codex Browser Launch

After the HTML report is completed, read the runner output line:

REPORT_URL=file://$HOME/.codex/usage-data/report.html

Then use the Browser plugin's in-app browser workflow to open that URL. Prefer the Codex browser over macOS open when this skill is running inside Codex.

If Browser tooling is unavailable, report the REPORT_URL clearly and leave the file on disk.

Do not claim the browser launch happened until the Codex browser has actually navigated to the REPORT_URL.

Codex Writer Contract

Codex must return only valid JSON with these top-level sections:

• metadata
• at_a_glance
• project_areas
• interaction_style
• what_works
• friction_analysis
• prompting_help
• suggestions
• on_the_horizon
• actionable_fixes
• fun_ending

The writer must:

• Use only the evidence bundle.
• Avoid inventing outcomes, files, tools, or user sentiment.
• Write in second person.
• Separate Codex-side friction from user-side ambiguity.
• Include copyable prompts for situations where Jamie lacks the technical vocabulary.
• Put missing-data caveats in metadata.limitations.

Validation

Stop at the first failed gate. Do not continue to report rendering, browser launch, or final summary if evidence generation, prompt writing, Codex JSON generation, JSON validation, or HTML generation fails.

• Evidence file exists and is valid JSON.
• Prompt file exists and is non-empty.
• insights.generated.json exists and is valid JSON.
• Required insight sections are present.
• HTML renders without requiring a network connection.
• Final REPORT_URL= was opened in the Codex in-app browser, or Browser unavailability was disclosed.
• Use repo validation before finishing changes to the skill:

./bin/ask skills audit Skills/agent-ops/insight-report --level strict --json

Failure Modes

No session data found:

No session data found in ~/.codex/sessions/

Run Codex for a few sessions first, then regenerate the report.

Codex CLI unavailable:

Use --prepare-only, then ask the current Codex session to read INSIGHT_PROMPT.md, write insights.generated.json, and rerun with --render-only.

Codex returned invalid JSON:

Open INSIGHT_PROMPT.md, ask Codex to repair the JSON shape, save insights.generated.json, and rerun --render-only.

Gotchas

• --prepare-only intentionally does not render HTML; it only writes the evidence bundle and prompt for Codex-authored analysis.
• Sparse or missing sessions are not a runner failure. Preserve the limitation in the generated insight JSON instead of inventing patterns.
• The report path is outside this repository under $HOME/.codex/usage-data/; do not commit generated reports or prompts to agent-skills.
• Browser launch is a separate verification step. The runner printing REPORT_URL= is not proof that the Codex in-app browser opened it.

Safety

• The runner reads local Codex session data and writes local report artifacts only.
• Codex receives a bounded evidence bundle, not unrestricted filesystem access.
• codex exec is invoked with --sandbox read-only.
• The Python runner saves the generated JSON and HTML locally.
• Sensitive-looking values in sessions should be treated as evidence only, not repeated unless needed for a safe recommendation.

Anti-patterns

Anti-pattern	Safer behavior
Asking another model or local service to write the narrative	Use the Codex writer path only, or stop in --prepare-only for this Codex session to write it
Guessing insights when sessions are missing or sparse	State the evidence gap in metadata.limitations and keep claims conservative
Passing user-supplied shell commands into the report runner	Use the fixed codex exec --sandbox read-only invocation
Repeating secrets, tokens, or private prompt fragments from session logs	Summarize the pattern without exposing sensitive strings
Saying the report opened in the Codex browser before navigation succeeds	Report the REPORT_URL and disclose Browser unavailability or failure
Turning report suggestions into automatic cleanup commands	Keep recommendations copyable and non-destructive unless the user explicitly asks for follow-up implementation

Examples

Standard weekly review with browser launch:

python3 Infrastructure/references/deferred-skill-context/agent-ops-insight-report/scripts/run_insight_report.py --days 7

After the runner prints REPORT_URL=file://..., open that URL in the Codex in-app browser and mention the local path in the summary.

Prepare artifacts for this Codex conversation to write:

python3 Infrastructure/references/deferred-skill-context/agent-ops-insight-report/scripts/run_insight_report.py --prepare-only --no-open

Render after Codex-written JSON exists:

python3 Infrastructure/references/deferred-skill-context/agent-ops-insight-report/scripts/run_insight_report.py --render-only --no-open

References

• Generator: Infrastructure/references/deferred-skill-context/agent-ops-insight-report/scripts/run_insight_report.py
• Configuration: references/configuration.md
• Writer contract: references/codex-writer.md
• Report format: references/report-format.md
• Output root: $HOME/.codex/usage-data/

See Also

Skill	When to use
[[codex-automation-architect]]	Convert recommendations into Codex automations
[[skill-refactor]]	Analyze skill usage and improvement opportunities
[[ubiquitous-language]]	Extract terminology Jamie can reuse in future prompts

Topic map: [[agent-ops]]