Configure OpenClaw Forge full-brief evaluation

[sanafayyaz315]

Summary

• Configure the openclaw-forge submission's GLM-5-3-Flash agent with maxTokens: 128000 and set its separate preflight request to 512 tokens.
• Collect the published brief.json, reject OpenClaw's no-final-summary fallback, and require a brief with an evidenceId and scope: full for the morning-briefing case.
• Show commented, optional submission-repo, harness-repo/revision, and prepare-stage llm-base-url overrides in the existing OpenShell PipelineRun example; note where to change its sandbox image digest.

Why

The Flash output budget was too small for a larger briefing run. The evaluation also needs to distinguish a completed full brief from an attention brief or a fallback response. These settings are confined to the openclaw-forge submission; the shared Pipeline and Evaluate Task are unchanged.

Verification

• The changed YAML files parse successfully, and the full-brief check returns true for scope: full and false for scope: attention or missing scope.
• Pre-commit passed on the final commit.
• Earlier Gateway evaluations with test-specific run configuration produced full briefs at caps 30, 100, and 400; the cap-400 run found 104 available emails. These runs predate this cleaned PR commit and do not establish content quality or a live rerun of this exact diff.

Scope

The branch excludes the test-only inbox-cap parameter, morning-only case copy, DEBUG/raw-stream settings, personal change log, and saw-mpk Gateway PipelineRun template. The template remains available locally for testing. The published_brief check verifies full scope and an evidence ID; it does not score the brief's content. Content-quality evaluation is the next step.

Companion harness PR: GuyZivRH/agent-eval-harness#9

[GuyZivRH]

lgtm