PR1 is now ready for review and the non-draft CI path has started running on 9539b56c.

I wasn’t able to use GitHub’s reviewer-request mutation from this fork context, so I’m tagging likely runtime maintainers here instead: @steipete @vincentkoc.

This slice is intentionally narrow and GPT-5-first:

• embedded-Pi only
• no auth classification
• no permission truthfulness changes
• no continuation/liveness scope

Tracked by #64227 and scoped by #64228.

Greptile Summary

This PR adds an opt-in strict-agentic execution contract for embedded Pi agents running OpenAI/Codex GPT-5-family models. When enabled, the runtime detects plan-only turns (text-only responses that promise future action without actually calling tools), retries up to twice with an "act now" steering instruction, and fails closed with an explicit blocked-state payload if retries are exhausted. The update_plan tool auto-enable is also tightened so it only activates for strict-agentic GPT-5 runs or explicit tools.experimental.planTool: true, removing the previous OpenAI/Codex blanket auto-enable.

Confidence Score: 5/5

Safe to merge; the only finding is a minor P2 inconsistency in the blocked-state return.

All logic paths for the strict-agentic retry/block cycle are correct and well-tested. The GPT-5 guard, per-agent contract resolution, and update_plan auto-enable gating are consistent across all call sites. The single P2 comment (missing finalAssistantVisibleText in the blocked-state meta) is a minor cosmetic inconsistency that does not affect runtime correctness or user-facing behavior.

src/agents/pi-embedded-runner/run.ts around the blocked-state early return (line ~1562) — finalAssistantVisibleText is omitted from meta unlike every other exit path.

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/run.ts
Line: 1562-1581

Comment:
**Missing `finalAssistantVisibleText` in blocked-state meta**

Every other early-exit return in this function (the retry-limit path at ~L1488, the incomplete-turn path at ~L1611, and the success path at L1645) includes `finalAssistantVisibleText` in `meta`. The blocked-state return omits it, so the planning text that caused the agent to be blocked is not preserved for callers that use this field for logging or display. The variable is already in scope (computed at L1450), so adding it here is a one-liner.

```suggestion
            return {
              payloads: [
                {
                  text: STRICT_AGENTIC_BLOCKED_TEXT,
                  isError: true,
                },
              ],
              meta: {
                durationMs: Date.now() - started,
                agentMeta,
                aborted,
                systemPromptReport: attempt.systemPromptReport,
                finalAssistantVisibleText,
              },
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "agents: catch structured plan-only turns" | Re-trigger Greptile}

CI is green on 310c27ab, and the follow-up fixes from the initial review pass are already folded in.

Scope is still intentionally narrow to #64228 only:

• embedded-Pi strict-agentic same-turn behavior for openai / openai-codex GPT-5-family runs
• revised update_plan semantics and gating
• no permission, replay, or generic provider-contract expansion in this slice

This PR is ready for maintainer review.

Rebuilt this branch from a clean upstream/main history and force-pushed it back over the polluted branch state. The PR is now back to the intended strict-agentic slice only, with the earlier review fixes preserved.

Current intended scope remains:

• embedded-Pi strict-agentic same-turn behavior for openai / openai-codex GPT-5-family runs
• revised update_plan gating and non-chatty semantics
• no auth, permission, replay, or harness scope in this PR

This should address the dirty-branch bot warning and make the diff reviewable again.

Addressed the latest review pass on this branch.

Changes in this push:

• use the resolved agentDir when marking auth profile usage inside the embedded runner
• trim modelId before the GPT-5 strict-agentic guard regex so whitespace does not disable the retry/fail-closed path

Local validation:

• CI=1 pnpm exec vitest run src/agents/openclaw-tools.update-plan.test.ts src/agents/tools/update-plan-tool.test.ts src/agents/pi-embedded-runner/run.incomplete-turn.test.ts src/agents/system-prompt.test.ts src/agents/tool-catalog.test.ts

One note on lint:tmp:no-raw-channel-fetch: the current failure points at extensions/browser/src/browser/client-fetch.ts, which is outside this PR diff (git diff --name-only upstream/main...HEAD does not include that file). I’m treating that as shared/mainline debt unless CI shows a branch-specific regression here.

Rebased this branch onto current upstream/main and pushed refreshed head cb1453de29.

Why this mattered:

• the branch was 17 commits behind main
• the current checks-node-test failure was coming from stale plugin-install test expectations on the older base, not from the strict-agentic runtime slice itself

Validation on the rebased head:

• CI=1 pnpm exec vitest run src/agents/openclaw-tools.update-plan.test.ts src/agents/tools/update-plan-tool.test.ts src/agents/pi-embedded-runner/run.incomplete-turn.test.ts src/agents/system-prompt.test.ts src/agents/tool-catalog.test.ts
• CI=1 pnpm exec vitest run src/plugins/install.test.ts

Scope is still intentionally unchanged: this is the same strict-agentic GPT-5-first embedded-Pi slice, just refreshed onto current main so CI and maintainer review are looking at a clean base.

Latest state on this slice:

• Current head cb1453de29
• Required checks are green, including checks-node-test
• The stale bot findings on blocked-state metadata / resolved paths / GPT-5 model-id trimming were addressed on the rebased head

Scope stays intentionally narrow:

• owns strict-agentic same-turn execution for embedded-Pi GPT-5-family openai / openai-codex
• does not own auth/runtime classification, permission truthfulness, replay/liveness redesign, or parity benchmarking

At this point this should be the first merge candidate for the program.

Current head cb1453de29 is green on the required matrix, including checks-node-test and check-additional.

The earlier bot findings around blocked-state meta, resolved paths, and trimmed model-id behavior are addressed on this rebased head. Scope remains intentionally narrow: strict-agentic execution only, GPT-5-first and embedded-Pi-scoped.

This should be the first merge candidate. If the current head looks good from a maintainer perspective, I’d love formal review on this slice.

@steipete finished these now for your review Chat GPT major upgrade fixes (make GPT 5.4 = as good as Opus 4.6 in Openclaw). cut down to 4 PR's now.

Landed via rebase merge. Thanks @100yenadmin.

Commits on main:

• 09b1117271 agents: add strict-agentic execution contract
• e22f60faea docs: note strict-agentic execution contract

Validation before merge:

• pnpm check
• pnpm config:schema:check
• pnpm config:docs:check
• pnpm plugin-sdk:api:check
• pnpm build
• OPENCLAW_LOCAL_CHECK=1 pnpm test src/agents/openclaw-tools.update-plan.test.ts src/agents/pi-embedded-runner/run.incomplete-turn.test.ts src/config/zod-schema.agent-defaults.test.ts
• PR checks green: Labeler, Workflow Sanity, Install Smoke preflight

Thanks @steipete @vincentkoc ! 🖤 let me know what you think of the rest of the GPT upgrades:

Merge order

PR A (#64241) - merged
PR B (#64439) -need review
PR C (#64300) -need review
PR D (#64441) -need review

PR D is the proof layer. It should not block review of runtime-correctness slices.

Completion gate
The project is only complete when all of these are true 🚀
PR A, PR B, and PR C are merged and stable
PR D is merged and produces a readable parity report plus a machine-readable verdict
GPT-5.4 no longer stalls after planning
GPT-5.4 no longer fakes progress or fake tool completion
GPT-5.4 no longer gives false /elevated full guidance
replay/liveness failures are surfaced as explicit states, not silent disappearance
the parity gate shows GPT-5.4 matches or beats Opus 4.6 on the agreed metrics

gotta crash out but gave you edit permissions on all these @steipete