Skip to content

Dev - #4

Merged
evanpelle merged 107 commits into
mainfrom
dev
Feb 1, 2025
Merged

Dev#4
evanpelle merged 107 commits into
mainfrom
dev

Conversation

@evanpelle

Copy link
Copy Markdown
Collaborator

Merge thread split & perf improvements

@evanpelle
evanpelle merged commit c1d9090 into main Feb 1, 2025
BrunoJurkovic pushed a commit to BrunoJurkovic/OpenFrontIO that referenced this pull request Mar 2, 2025
Added CTRL+CLICK hotkey to instructions
@Celant Celant mentioned this pull request Jun 30, 2026
3 tasks done
joshuabradley012 pushed a commit to joshuabradley012/OpenFrontIO that referenced this pull request Aug 30, 2026
…ST/ACT log, calibrate.py, simWars and hystRetreats flags

Estimate.ts and tests/playbook/estimate.test.ts are back from 95f4b63 with three
calibration inputs: lossScale (× attacker loss per tile), speedScale (× tiles per
tick) and extraDefenderTroops (the target's allies' possible pile-in). Params gain
estLossScaleNation/Human/Bot and estSpeedScale (1.0 = the raw replay), fed into
every estimator call.

Calibration logging is always on (log only): every war wave and every tribe's
first click logs EST <target> wave=n troops= tilesEst= lossEst= ticksEst= wins=
class= others=; when the attack has left outgoingAttacks() the bot logs ACT with
the target's tile loss, our tile delta, sent − last troops seen, ticks, class and
how it ended (dead/done/retreat/fast). scripts/lab/calibrate.py (stdlib) pairs
them across a results dir and fits per defender class the log-ratio least-squares
lossScale / speedScale with spread, prints a table and the Params JSON blob;
--selftest on a synthetic fixture.

Flags (default off until the 30-game Medium A/B):
- simWars: the B1 Estimate-driven target/size pick (bisection, 20 % margin) on
  the calibrated scales; a running wave on the target counts as part of the
  estimate (engine merge rule) and, with trustWars, the target's allies'
  nationWouldSend as part of its army.
- hystRetreats: every 100 ticks 'continue' (600-tick estimate: survivors × 0.75
  + tiles × 60) vs 'retreat now' (troops × 0.75); continue must win by
  0.1 + 0.2 × clamp(maxBsr − 1, 0, 2) and lose twice in a row; a wave under
  retreatBelowRatio × the target's troops that cannot win is lost outright.
  retreatBelowRatio was read nowhere before. Works with realRetreats.

Golden hash regenerated: the calibration log adds EST/ACT lines to the bot log
(the hashed material). Decisions are unchanged — the golden material minus the
EST/ACT lines is byte-identical to the previous hash's, and a 6-minute africa
Medium smoke game's non-log rows and log (minus EST/ACT) are identical before
and after.

Tests: estimate (restored + scales), simWars, hystRetreats, calibration.
Docs: PlaybookBotPlan.md "openfrontio#4 — Calibrating the estimator".

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PoCSUbBqvfPV6bpb2kyfAc
evanpelle pushed a commit that referenced this pull request Sep 25, 2026
The pick logic moves to openfrontio/.github (#4) so other repos can use
it; this file keeps only the trigger, like claude-code-review.yml.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
evanpelle added a commit that referenced this pull request Sep 25, 2026
## Summary
- New workflow `cherry-pick-milestone.yml`: when a PR with milestone
`vNN` is merged, or a merged PR is given a `vNN` milestone, its merge
commit is cherry-picked (`-x`) onto branch `vNN`. Milestones like
`Backlog` and ones with no matching branch are skipped.
- Clean picks push straight to the release branch and comment the new
SHA on the original PR, matching how `v34` is updated by hand today.
- Picks already on the branch are skipped: either a commit with the same
`(#NNNN)` exists, or the pick would change nothing.
- On conflict, Claude (same OAuth secrets and backup-account fallback as
the review workflow) resolves the markers using only file tools and
read-only git. The workflow checks that no markers remain, then opens
`[vNN] <title>` from `cherry-pick/vNN/pr-NNNN` for human review. If
Claude decides the change can't apply to the older branch, it writes
down why and the workflow posts that on the original PR.
- Claude's instructions are built only from trusted values (PR number,
merge SHA, and a milestone checked against `vNN`). Runs one at a time
per release branch so two merges can't race to push. PRs opened with
`GITHUB_TOKEN` don't trigger CI, and the conflict PR body says so.

This briefly became a thin caller for a shared workflow in
openfrontio/.github (#4, now closed). OpenFrontIO is the only repo with
`vNN` milestones and release branches, so it stays here until another
repo needs it.

## Test plan
- `actionlint`: clean apart from info-level SC2016 on intentional
backticks inside single quotes. It accepts the `milestoned` trigger.
- Ran the pick step locally against real `v34` for the last 12 commits
on main: 5 clean, 5 already present, 2 conflicts, no false skips. Also
checked the empty-pick case with a PR number that isn't on the branch:
skipped with a clean tree.
- CodeQL: an environment-injection alert on the first version was fixed
in 2ebd801. No alerts are open.
- The Claude-resolution and PR-opening steps are untested outside
Actions. The first real conflict will be their first run.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@evanpelle evanpelle mentioned this pull request Nov 11, 2026
3 tasks done
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant