TLDR

Status: QA harness/mock-provider issue, not a proven broad Codex runtime tool dropout.

The original report overclaimed that Codex drops planned tool calls for most default tool fixtures. The corrected architecture is:

• Codex-native workspace tools (read, write, edit, apply_patch, exec, process, update_plan) are intentionally owned by Codex native app-server behavior, not duplicated as OpenClaw dynamic tools.
• OpenClaw-owned integration tools remain dynamic openclaw bridge tools and are valid parity rows.
• Mock provider providerPlanToolCalls are fixture intent, not actual runtime transcript tool calls.

Product impact if OpenClaw moved fully to Codex today: P4 as originally filed. This issue does not prove a Codex runner product bug.

QA impact: P1/P0 depending on lane. False product failures would undermine the parity gate, so PR #80323 corrects the harness and report language.

Latest Beta.5 Proof

Validated on PR #80323 at OpenClaw v2026.5.10-beta.5:

PR head: 3336dec6419c9cc9a87dc7cfa6f48118ca2d838e
Remote proof run: https://github.com/electricsheephq/openclaw-local-test/actions/runs/25719383976
Confidence tracker: #80936

Artifact-backed results:

{
  "tool-defaults-direct": { "total": 20, "passed": 20, "skipped": 0, "failed": 0 },
  "openclaw-dynamic-tools-direct": { "total": 8, "passed": 8, "skipped": 0, "failed": 0 },
  "tool-defaults-searchable": { "total": 20, "passed": 15, "skipped": 5, "failed": 0 },
  "confidence-report": { "pass": true, "zeroUnknowns": true }
}

The 5 searchable skips are explicitly report-only because the mock provider does not yet model Codex searchable/deferred OpenClaw dynamic tool declarations honestly. That is a QA/mock limitation, not a product bug claim.

What Actually Was Wrong

The old tool-defaults interpretation mixed three layers:

1. Codex-native workspace behavior: must be tested by user-visible native outcomes, not by expecting duplicate OpenClaw dynamic tools.
2. OpenClaw dynamic integration tools: must be tested through actual dynamic bridge calls/results.
3. Provider mock plans: must be displayed as fixture intent only, never as actual runtime tool calls.

Correct Fix In PR #80323

• Adds capability/bucket metadata for runtime tool rows.
• Suppresses false hard failures for Codex-native workspace tools when the only difference is missing duplicate dynamic exposure.
• Keeps OpenClaw dynamic integration tools hard-gated in --codex-tool-loading direct mode.
• Keeps searchable/deferred loading as a report lane until the mock provider can model deferred discovery.
• Shows counts.skipped in qa-suite-summary.json so report-only rows are not mistaken for passes.
• Separates provider-plan tool calls from runtime transcript tool calls.

Remaining Work

• Keep this issue open until [qa-lab] Complete Codex vs Pi runtime parity harness phases 2-5 #80323 lands or until maintainers decide where to track searchable/deferred mock-provider fidelity.
• Do not file product bugs from these mock-only rows unless native/live Codex behavior reproduces the failure independently.

Links

• Parent RFC/tracker: Codex-vs-Pi runtime parity QA harness (RFC + tracking) #80171
• PR: [qa-lab] Complete Codex vs Pi runtime parity harness phases 2-5 #80323
• Confidence proof: [QA-lab] Codex runtime parity beta.5 confidence proof for PR #80323 #80936
• Live/Testbox proof tracker: [QA-lab] Complete live-frontier token-efficiency and Testbox parity proof #80397