Avoid redundant bulk range comparisons in team-memory requests

[Changyong Gong (chagong)]

Cause

The first coordinator run after merging #51 failed in agent processing, not source validation, dispatch, Actions variables, or Azure OIDC: failed job.

Preflight completed and OIDC login succeeded. The team-memory agent then repeated complete-range verification through the unpaginated, patch-bearing github-compare_commits tool. The range 4175ea71e170938826fb847e6bd5108f0f5597cf..0bc63baa3b6426b1f22daabe80e4ab004801f39b has one commit and 17 changed files; a read-only measurement returned 187,415 bytes, exceeding the MCP tool's 128-KiB transport limit. The agent stopped with a valid needs-review receipt and reported that no wiki publication was attempted. The action correctly failed an incomplete maintenance batch.

Change

• Clarify in the caller-owned batch request that invocation preflight already validated the complete fast-forward commit inventory and eligible merged PR list. Do not repeat that range validation with compare_commits or make its optional bulk response a prerequisite for analysis.
• Keep independent agent revalidation of repository/branch/PR merge identities, source/wiki evidence, privacy, destination policy, paired write preconditions and publication confirmation.
• Explicitly start PR evidence with list_pull_request_files(per_page=1, page=1), recheck PR metadata after paging, and verify relevant final source with full-SHA get_file reads. Rebase merge tips remain insufficient whole-PR evidence.
• Preserve needs-review/partial failure receipts and no-retry behavior; do not turn tool limits into successful maintenance.
• Document the failure distinction and recovery: after merging the fix, start a new authorized manual coordinator run for PR Use artifact-free team-memory reconciliation with standalone dispatch #51. Rerunning the old run retains its original workflow/action revision.

This is an action-request fix on the merged main baseline. No hosted-agent/tool code or size limits changed, and no runtime deployment is needed for the new caller instruction. No live rerun, agent invocation, deployment, wiki write, workflow enablement, credential change, or Java Pack migration was performed.

Validation

• Regression tests first reproduced the missing range-validation handoff, then passed after the fix.
• 130 focused tests passed on Python 3.12, including direct/coordinated handoffs, full range fail-closed discovery, direct/manual/issue-loop compatibility, standalone dispatch, and retained incomplete receipts without retrying.
• Python compilation, git diff --check, and self-review passed.
• Published commit: f01b76b4896c324486e2cde1ab54910f8bbd68ad.
• Fresh hosted CI passed all six jobs for that exact revision: CI run 37930917380, including full application tests on Python 3.13, MCP tests/package builds on Python 3.12/3.13, and workflow validation. CLA also passed.

Prompt/contract tests validate the caller payload and preserved safeguards, not the deployed model's future tool choices or successful wiki publication. Those require an explicitly authorized fresh operational run.