Greptile Summary

This PR combines auth/runtime failure classification (ProviderRuntimeFailureKind) with truthful full-access surfacing for the Codex provider. The core classification logic in errors.ts, the new resolveEmbeddedFullAccessState helper, and the OAuth scope normalization are solid. Two introduced bugs need attention before merging:

• In handleAgentEnd, a redundant direct call to classifyProviderRuntimeFailureKind (without a parsed HTTP status) overrides observedError.providerRuntimeFailureKind — the correctly-classified value that does use the parsed status — causing JSON error payloads to be misclassified (e.g., auth_scope → unknown) in lifecycle logs.
• In commands-system-prompt.ts, only .available is read from resolveEmbeddedFullAccessState; .blockedReason is discarded, so the system prompt always renders "runtime constraints" for unavailable full access instead of the actual reason ("host policy", "channel", etc.).

Confidence Score: 4/5

Two P1 bugs (classification override in lifecycle log, lost fullAccessBlockedReason in commands prompt) should be fixed before merging.

The core classification machinery (errors.ts, sandbox-info.ts, OAuth normalization) is correct. The two P1 issues are both straightforward one-to-three-line fixes with clear root causes, so the overall structural approach of the PR is sound.

src/agents/pi-embedded-subscribe.handlers.lifecycle.ts and src/auto-reply/reply/commands-system-prompt.ts

This is a comment left during a code review.
Path: src/agents/pi-embedded-subscribe.handlers.lifecycle.ts
Line: 58-86

Comment:
**Redundant computation overrides correctly-classified `providerRuntimeFailureKind`**

`providerRuntimeFailureKind` is computed here without an explicit `status` (lines 58–61), then immediately overrides `observedError.providerRuntimeFailureKind` (the spread at line 85), which was computed with the parsed `httpCode` from the JSON payload.

For Codex JSON error payloads like `{"code":401,"message":"Missing scope: api.responses.write"}`, `inferSignalStatus` in the direct call returns `undefined` (no leading HTTP status in the JSON string), so `isAuthScopeErrorMessage(raw, undefined, provider)` early-returns `false` (requires status 401 or 403), and the function falls through to `"unknown"`. Meanwhile `observedError.providerRuntimeFailureKind` correctly returns `"auth_scope"` because it passes `status: 401` from the parsed `httpCode`. The explicit assignment on line 86 clobbers the correct value.

Remove the separate computation and rely on the value already present in `observedError`:

```suggestion
    const observedError = buildApiErrorObservationFields(rawError, {
      provider: lastAssistant.provider,
    });
    const providerRuntimeFailureKind = observedError.providerRuntimeFailureKind;
```

Then remove the `classifyProviderRuntimeFailureKind` import from this file and drop the now-redundant explicit `providerRuntimeFailureKind` key at line 86 (it's already part of the spread).

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

---

This is a comment left during a code review.
Path: src/auto-reply/reply/commands-system-prompt.ts
Line: 110-132

Comment:
**`fullAccessBlockedReason` is discarded — system prompt always shows "runtime constraints"**

`resolveEmbeddedFullAccessState` returns `{ available, blockedReason? }`, but only `.available` is stored in the `elevated` object. `fullAccessBlockedReason` is silently dropped, so when `fullAccessAvailable === false`, `system-prompt.ts` calls `formatFullAccessBlockedReason(undefined)` and always renders "runtime constraints", even when the real reason is `"host-policy"` (which should render "host policy").

Store and forward the blocked reason:

```suggestion
    const fullAccessResult = resolveEmbeddedFullAccessState({
      sandboxEnabled: true,
      execElevated: {
        enabled: params.elevated.enabled,
        allowed: params.elevated.allowed,
        defaultLevel: (params.resolvedElevatedLevel ?? "off") as
          | "on"
          | "off"
          | "ask"
          | "full",
      },
    });
    const sandboxInfo = sandboxRuntime.sandboxed
      ? {
          enabled: true,
          workspaceDir,
          workspaceAccess: "rw" as const,
          elevated: {
            allowed: params.elevated.allowed,
            defaultLevel: (params.resolvedElevatedLevel ?? "off") as "on" | "off" | "ask" | "full",
            fullAccessAvailable: fullAccessResult.available,
            ...(fullAccessResult.blockedReason
              ? { fullAccessBlockedReason: fullAccessResult.blockedReason }
              : {}),
          },
        }
      : { enabled: false };
```

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

_{Reviews (1): Last reviewed commit: "test: align full-access sandbox info exp..." | Re-trigger Greptile}

Addressed the current review comments on runtime truthfulness.

This push:

• stops recomputing providerRuntimeFailureKind in lifecycle and uses the observed error as the source of truth
• threads fullAccessBlockedReason through the commands/system-prompt path instead of dropping it
• resolves sandbox/full-access state from the target session key in runPreparedReply

Focused validation passed locally:

• CI=1 pnpm exec vitest run src/commands/openai-codex-oauth.test.ts src/agents/pi-embedded-helpers.formatassistanterrortext.test.ts src/agents/pi-embedded-helpers.isbillingerrormessage.test.ts src/agents/failover-error.test.ts src/agents/pi-embedded-error-observation.test.ts src/agents/pi-embedded-subscribe.handlers.lifecycle.test.ts src/agents/pi-embedded-runner.buildembeddedsandboxinfo.test.ts src/agents/system-prompt.test.ts src/auto-reply/reply/get-reply-run.exec-hint.test.ts
• pnpm build

Addressed the concrete checks-node-test failure from the last run and pushed follow-up commit e09eb03df4.

The failing job was not generic CI noise: src/auto-reply/reply/get-reply-run.media-only.test.ts was mocking ../../routing/session-key.js without exporting DEFAULT_AGENT_ID, and the newer agent-scope path now reads that constant while resolving sandbox runtime status.

What changed:

• added DEFAULT_AGENT_ID and DEFAULT_MAIN_KEY to the mocked session-key module in that test

This keeps the test fixture aligned with current runtime exports without widening PR scope beyond runtime truthfulness.

Addressed the remaining branch-owned runtime-truthfulness gap on the latest head 6e36f6db1d.

What changed:

• keep /elevated full truth correct for direct host runs instead of deriving unavailability from sandboxed=false
• preserve explicit runtime blocks when fullAccessAvailable=false is set
• added focused regression coverage for the host-run case

Local validation on this head:

• pnpm build
• CI=1 pnpm exec vitest run src/agents/pi-embedded-runner.buildembeddedsandboxinfo.test.ts src/agents/system-prompt.test.ts src/auto-reply/reply/get-reply-run.exec-hint.test.ts

Scope remains runtime truthfulness only:

• owns auth/runtime failure classification + truthful /elevated full
• does not own tool schema normalization, replay/liveness redesign, or parity gating.

Current head is now 8d24cb3504.

This update fixes the remaining branch-owned checks-node-test break on the latest head: src/auto-reply/reply/get-reply-run.media-only.test.ts now keeps the real session-key exports and only overrides the pieces the test needs, so the new runtime path can still resolve parseAgentSessionKey.

Local validation on this head:

• CI=1 pnpm exec vitest run src/auto-reply/reply/get-reply-run.media-only.test.ts src/agents/system-prompt.test.ts src/auto-reply/reply/get-reply-run.exec-hint.test.ts

The remaining red GitHub signals appear to be shared/base CI noise rather than PR-owned regressions: the same msteams failure (conversationStore.get is not a function in extensions/msteams/src/monitor-handler/message-handler.thread-parent.test.ts) is also hitting PR C and PR D, outside this PR’s diff.

Once the rerun settles, this should be ready for human maintainer review as the runtime-truthfulness 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.

Addressed the remaining runtime-truthfulness review items on head ac7a0660a0.

This update:

• accepts slash-terminated /oauth/authorize/ paths when normalizing required Codex OAuth scopes
• resolves sandbox/full-access state in the command system-prompt path from the target sessionKey
• derives command-path full-access availability from the actual sandbox runtime instead of a hardcoded sandboxed assumption
• keeps fullAccessBlockedReason flowing into sandbox prompt metadata

Focused validation passed locally:

• CI=1 pnpm exec vitest run src/commands/openai-codex-oauth.test.ts src/auto-reply/reply/commands-system-prompt.test.ts

All currently open review threads on this PR have been resolved against the latest head.

Current head 191d7f224c addresses the branch-owned checks-node-test-auto-reply regression: the directive-behavior expectation now includes truthful bashElevated.fullAccessAvailable metadata instead of assuming the older smaller shape.

Local validation on this head:

• CI=1 pnpm exec vitest run src/auto-reply/reply.directive.directive-behavior.defaults-think-low-reasoning-capable-models-no.test.ts src/agents/pi-tools-agent-config.test.ts src/security/temp-path-guard.test.ts

That focused rerun passed 3/3 files and 33/33 tests locally. The other two previously failing shards from GitHub (pi-tools-agent-config and temp-path-guard) did not reproduce in this runtime-truthfulness worktree, so if they recur they should be treated as shared/base noise rather than new scope in this PR.

Scope is still runtime truthfulness only.

Landed. Thanks @100yenadmin.

Follow-up fixes applied before merge:

• Added the missing auto-reply fullAccessAvailable expectation and widened that slow e2e-style case timeout.
• Classified raw OpenAI Codex missing-scope payloads as auth_scope even without a leading HTTP status line.
• Kept temp-path security guard green by avoiding the flagged fixture pattern.
• Replaced regex-based base64 canonicalization with loop validation to avoid stack overflow on oversized image payloads.
• Stabilized the Telegram live QA abort-deadline test after the live-transport move.

Local verification:

• pnpm format
• pnpm check
• pnpm build
• pnpm test src/agents/pi-embedded-helpers.isbillingerrormessage.test.ts src/agents/pi-embedded-helpers.formatassistanterrortext.test.ts extensions/qa-lab/src/live-transports/telegram/telegram-live.runtime.test.ts src/media/base64.ts src/auto-reply/reply/bash-command.stop.test.ts
• pnpm test src/auto-reply/reply.directive.directive-behavior.defaults-think-low-reasoning-capable-models-no.test.ts -t "passes elevated defaults when sender is approved"
• OPENCLAW_VITEST_MAX_WORKERS=1 pnpm test before final rebases; all lanes passed except the QA Telegram timing flake fixed afterward.

Merged as fe6341f7026d0599da4b3eefc5b021205d1a49bf.