Best possible solution:

Land the status cold-path optimization after rebuilding global session rows once per unique store without agentIdOverride, with shared-store regression coverage; split the task-health surface unless maintainers explicitly want it in this PR.

Do we have a high-confidence way to reproduce the issue?

Yes, source-backed: the PR body gives before/after timing output for the affected 2026.5.7 install, and current main still has the status model parsing path that can manifest-normalize without an explicit allowManifestNormalization: false. I did not rerun the live timing locally in this read-only review.

Is this the best way to solve the issue?

No, not as submitted: the manifest-normalization and CLI-provider-cache pieces are narrow, but the global session aggregation change breaks shared-store semantics. The safer repair is to reuse loaded stores without reusing per-agent-overridden rows for the global list.

Full review comments:

• [P2] Deduplicate shared session stores — src/commands/status.summary.ts:264
  When session.store lacks {agentId}, every agent resolves to the same store file. Appending rows already built with agentIdOverride makes global sessions.count and sessions.recent include each stored session once per configured agent, with shared rows stamped as the current agent. Build the global rows once per unique store without the override, or dedupe by store/key.
  Confidence: 0.92

Overall correctness: patch is incorrect
Overall confidence: 0.9

Acceptance criteria:

• pnpm test src/commands/status.summary.test.ts src/commands/status.summary.runtime.test.ts src/status/status-message.test.ts
• pnpm exec oxfmt --check --threads=1 src/commands/status.summary.ts src/commands/status.summary.test.ts src/commands/status.summary.runtime.ts src/commands/status.summary.runtime.test.ts src/status/agent-runtime-label.ts src/agents/model-selection.ts
• pnpm build
• If task-health changes remain in the branch, also run pnpm test src/commands/tasks.test.ts src/gateway/server-methods/tasks.test.ts

What I checked:

• PR head proof and scope: The PR body reports live before/after timings on an affected 2026.5.7 install: openclaw status improved from 20.32s to 2.27s while openclaw gateway status stayed about 1.3s, and the head patch includes status cold-path changes plus separate task-health changes. (837c1af00688)
• Current-main status model parsing still enters manifest normalization: Current main passes allowPluginNormalization: false for status model parsing but does not pass allowManifestNormalization: false into parseModelRef for configured or persisted status model refs. (src/commands/status.summary.runtime.ts:35, 726ecc4987e7)
• Model normalization contract: parseModelRef normalizes through normalizeProviderModelId; manifest normalization is controlled separately from plugin normalization, so disabling plugin normalization alone does not keep status parsing off the manifest-normalization path. (src/agents/model-selection-normalize.ts:39, 726ecc4987e7)
• Current-main global status aggregation contract: Current main builds per-agent rows with agentIdOverride, then builds the global recent list from each unique store path without the override, preserving one global row per stored session key. (src/commands/status.summary.ts:257, 726ecc4987e7)
• Shared session stores are supported: When session.store does not include {agentId}, resolveStorePath returns the same resolved path for every agent; current tests also document shared stores with both agent:main:* and agent:ops:* keys as legitimate. (src/config/sessions/paths.ts:294, 726ecc4987e7)
• Shared-store regression in the PR diff: The PR replaces the unique-store global pass with allSessionsByAgent.push(...sessions), where sessions were already built with agentIdOverride; a shared store is therefore counted once per configured agent in the global status list. (src/commands/status.summary.ts:264, 837c1af00688)

Likely related people:

• vincentkoc: Recent GitHub history shows fix(status): show runtime in CLI sessions on the status summary path, and earlier task health/audit work in the task area that the PR touches. (role: recent status and task area contributor; confidence: high; commits: 46c99cff0b2f, c52fac836cc8, e57b3618fc8d; files: src/commands/status.summary.ts, src/commands/status.summary.runtime.ts, src/status/agent-runtime-label.ts)
• steipete: Recent history shows status model-resolution fast-path work and several task audit/runtime refactors adjacent to the PR's status and task surfaces. (role: recent status model-resolution and task audit contributor; confidence: high; commits: e61eba11e671, 1fb44f0aadd4, 32ebaa37574e; files: src/commands/status.summary.runtime.ts, src/agents/model-selection.ts, src/tasks/task-registry.audit.ts)
• tmimmanuel: GitHub history shows feat(gateway): add SDK task ledger RPCs, which introduced the gateway/SDK task summary surface extended by the PR's task-health commits. (role: gateway task ledger surface introducer; confidence: medium; commits: 2945948a5ee4; files: src/gateway/server-methods/tasks.ts, src/gateway/protocol/schema/tasks.ts, packages/sdk/src/types.ts)

Remaining risk / open question:

• I did not run timing or test commands because this review was read-only; the performance reproduction is based on the PR's live-output proof plus source inspection.
• The task-health gateway/SDK/docs changes are outside the stated status cold-path fix and may need to be split or explicitly accepted separately.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 447206623700.