Repository navigation
Releases
Releases · MedAdemBHA/mythical-agent-month
v1.2.0: plan figure, plugin, evals
- Bold Plan figure = likely × 1.3 until calibrated. On the reference run that's +14% in-sample and −19% held-out, versus −12% / −37% for the raw likely value.
"calibrated": trueor per-taskminutesswitch it back to likely and hide the warning. - No-AI comparison is now a wide, unmeasured range and stays out of headlines.
- Claude Code plugin:
/plugin install mythical-agent-month --marketplace MedAdemBHA/mythical-agent-month - Eval suite: 6 cases for
claude plugin eval, including a must-not-apply case. - With vs without: one everyday prompt, both answers verbatim (examples/with-vs-without.md). Plain Claude already calls the 6-week estimate too long; the skill adds structure.
Still n = 1. The timing data for the held-out test isn't public yet.
v1.1.0: honest accuracy
- Held-out test on the reference run (fit on phases 0-3, predict 4-9): agent-hours -24%, wall-clock -37%. See examples/holdout.md. The skill now tells the agent to plan on the upper half of its range until you calibrate.
- One no-AI baseline (size x risk), the same in both modes.
- Measured risk multipliers (0.3 / 2.7) and size table; no hand-tuning.
- Narrower trigger: teams without AI get a normal estimate.
- The formula in the docs matches the script; the low end of the range is realistic.
- Manual install option; installer for Claude Code, Codex, Cursor, Gemini CLI and Copilot.
n = 1. A second, independent run is the missing evidence: run scripts/calibrate.py and open a PR.
v1.0.0
First release. Install: upload mythical-agent-month.skill in claude.ai, or copy skills/mythical-agent-month to ~/.claude/skills/.