[shekharprateek]

Grouped because they share one test file and none is big enough to warrant its own PR. Each regression test was confirmed to fail against main before the fix and pass after — verified by reverting the sources while keeping the tests:

sources reverted to main          fixes restored
  BedrockSlugPrefix        FAILED   →  OK
  RetryGuardsFinishedWork  FAILED   →  OK
  DisjointAgentVllmCache   FAILED   →  OK

1. Token totals inflated ~2× for codex and strands on vLLM

codex, strands and strands-harness report input_tokens as the whole prompt, and their counts are declared disjoint, so everything downstream sums input + output + cache_read + cache_write.

That's correct when the cache figure comes from the agent's own API. When it comes from vLLM's Prometheus counters instead, cache_read is a subset of input_tokens, so the cached prompt gets counted twice:

	1000-token prompt, 800 cached
Before	1000 + 100 + 800 + 200 = 1900
After	200 + 100 + 800 + 0 = 1100

_summary_metrics now moves the cached part out of input and records no separate cache write — the uncached prefill is the remaining input — and sources.input_tokens states the subtraction. A run that never reported an input count keeps the old shape, since there's no whole prompt to split.

This matters beyond the token count: on the self-hosted path, cost per task is derived from tokens processed.

2. A retry deleted finished work

_summary_is_retryable only refused on ok or error_max_turns. A task that wrote all six artifacts and then exited on a non-stop reason — omp can end on stopReason toolUse straight after its final summary — was classed transient. And a retry clears the task's artifacts before re-running.

So the one case where the work was already complete is the case where it got thrown away. A task with design and patch both on disk is now never retried: the judge scores the files, not the exit status.

3. model_to_slug missed three of Bedrock's four region prefix forms

The regex was ^[a-z]{2}\., so us. and eu. stripped while global., apac. and us-gov. did not — and global. is the form this repo's own docs recommend, for the 10% lower cross-region rate.

The slug names the artifact folder the /swe skill writes to, so a global.-prefixed model wrote its artifacts under global.anthropic.claude-.../ while the harness and the judge looked under claude-.../.

us.anthropic.claude-opus-5       → claude-opus-5
global.anthropic.claude-opus-5   → claude-opus-5   (was: unchanged)
apac.anthropic.claude-opus-5     → claude-opus-5   (was: unchanged)
us-gov.anthropic.claude-opus-5   → claude-opus-5   (was: unchanged)
moonshotai.kimi-k2-thinking      → moonshotai.kimi-k2-thinking   (correctly untouched)

That last case is deliberate: a bare <vendor>.<model> has no region component, so there's no prefix to strip (issue #205). The fix is scoped to the four real region forms rather than "strip anything before the second dot".

Checks

642 tests pass, 7 new. ruff, ruff format, mypy, bandit (0 issues at every severity) clean. The one failure, test_judge_common.test_lowercase_preferred_on_collision, fails on main too — it assumes a case-sensitive filesystem.

Remaining from the audit

.gitignore not ignoring run output or runner.yaml (while four docs claim it does), CLAUDE.md not importing AGENTS.md, and the omp fuzzy model-match guard. Then versioning.