# Pelican CLI observations — 2026-09-15

Twelve independent Codex CLI sessions, requested models `gpt-5.6-sol` and
`gpt-6-astra`, reasoning `low` and `high`, three attempts per setting. All attempts
are included; no drawing was improved or retried. All twelve produced SVGs.

Exact user prompt: `Generate an SVG of a pelican riding a bicycle`.

## Conditions and limitations

- Codex CLI 0.154.0, existing authenticated account, read-only temporary folders.
- Explicit flags: `--ignore-user-config --ignore-rules --ephemeral
--skip-git-repo-check --sandbox read-only -c project_doc_max_bytes=0`.
  Each invocation also set `-m` and `-c model_reasoning_effort` to its recorded values.
- Default runtime instructions, available skills and tools remained in the environment.
  Every session reported: “Skill descriptions were shortened to fit the skills context
  budget.” Some sessions called commands, including environment/skill inspection.
  These observations are not a tool-free or instruction-free model evaluation.
- At most two sessions ran concurrently. Elapsed time includes client startup,
  network, tool overhead and inference; it is not model latency.
- Requested model names are recorded, not independently verified backend snapshots.
  No before/after baseline exists, so this series cannot establish degradation.
- Three attempts per setting cannot support a population-level quality ranking.

## Artifacts

`manifest.json` records UTC timestamps, CLI settings, elapsed seconds, token usage
reported by the CLI, tool-command counts, warning/error counts, extraction method,
and SHA-256 of each original SVG. Raw event streams remain local because they can
contain runtime instructions and local paths. The public manifest omits those details.

The SVG element was extracted unchanged from each final response. Sol low run 3
returned an SVG data URL; percent-decoding it produced the archived SVG. Preview
PNGs preserve the drawing and add white margins to a 720 × 540 canvas; SVGs were
not repaired. The 1200 × 1360 comparison image includes all attempts, grouped by
requested model/effort with run numbers across columns.

Inspect the visual outputs directly. A recognizable drawing does not establish
correctness on software engineering tasks. The site keeps these observations
separate from its six historical examples.