elophanto/EloPhanto v2026.10.01 on GitHub

129 commits since v2026.08.11: 200 files, +46,873/-1,133.

Upgrade before 2026-10-14

Codex retires gpt-5.5 on 2026-10-14. This release keeps older configs working:

At call time, gpt-5.5 maps to gpt-6.1-sol and gpt-5.5-mini maps to gpt-6-luna.
./update.sh (or elophanto config migrate) rewrites config.yaml, including direct-OpenAI entries.

./update.sh

Long, unattended runs (long-run autonomy, phases 0–5).
- Fixed twelve ways an unattended run stalled, lied or lost its work.
- Every goal keeps a run ledger, and every interrupted run leaves a handoff.
- The agent deliberates before expensive calls.
- Progress is verified by something other than the model that did the work.
- Long work keeps its turn across days.
- Failures teach, and the operator gets a daily health digest.
Learning that is measured. Checkpoint plans are scored, and lessons earn their place. A benchmark replays the agent's own verified work, playbooks are tuned against it, and training data can be exported.
Models. GPT-6.1 Sol is the Codex default. The thinking steps get their own route on the strongest model.
Agent Society. A live isometric campus where residents move as the agent works, in its own isolated window.
Competitive intel.
- An executive deck in a consulting-house style.
- Voice of customer.
- A weekly one-page brief with mid-week alerts.
- A regulatory register.
- Logged-in observation the agent drives itself.
- Research that reads how old a page is.
Search. The dated Search.sh contract: date controls, source dates and conflicts.
CI is green again: ruff, mypy across core/ tools/ cli/, and 3,941 tests.

Full details: CHANGELOG.md.