agents: strengthen GPT-5.4 execution bias and close the one-action-then-narrative loophole by 100yenadmin · Pull Request #65257 · openclaw/openclaw

100yenadmin · 2026-04-12T08:28:33Z

Summary

Four targeted fixes that close the behavioral gap between GPT-5.4 and Claude agents in OpenClaw. Based on root-cause analysis of the actual code paths in extensions/openai/prompt-overlay.ts, src/agents/pi-embedded-runner/run.ts, and src/agents/pi-embedded-runner/run/incomplete-turn.ts.

Part of #64227. See the umbrella for how this fits with #65219 (runtime activation) and #65224 (test proof).

What changed (3 commits)

Commit 1: Prompt + detection layer

extensions/openai/prompt-overlay.ts

OPENAI_GPT5_EXECUTION_BIAS rewritten from passive ("Start the real work") to imperative ("Use a real tool call FIRST. Commentary-only turns are incomplete. Do not stop after one step to ask permission."). Removed "Before the final answer, quickly verify" which GPT-5.4 used as a pause point.
New OPENAI_GPT5_TOOL_CALL_STYLE section override: "Call tools directly without narrating. If multiple tool calls are needed, call them in sequence without stopping." Includes all three approval safety lines from the default section (never exec /approve, allow-once is single-command, preserve full command for approval).
resolveOpenAISystemPromptContribution now returns execution_bias + tool_call_style + interaction_style overrides.

src/agents/pi-embedded-runner/run/incomplete-turn.ts

New isSingleActionThenNarrativePattern(): detects turns where exactly 1 non-plan tool call + planning prose ("I'll", "let me") = retry eligible.
Changed the exemption at line 355: hasNonPlanToolActivity && !isSingleActionThenNarrativePattern so the one-action loophole is closed.

Commit 2: Auto-continuation loop

Config surface (src/config/types.agent-defaults.ts + zod schemas + labels):

agents.defaults.embeddedPi.continuationMode: "auto" / "prompt" / "off" (default: "auto" for strict-agentic GPT-5, "prompt" for everything else)
agents.defaults.embeddedPi.continuationBudget: 1-20 (default: 5)

Continuation intent detection (incomplete-turn.ts):

hasContinuationIntent() — detects promise patterns in the model's visible text
hasCompletionLanguage() — detects "done", "finished", "implemented"
AUTO_CONTINUATION_INSTRUCTION — "Continue. Take the next concrete action immediately."

The loop (run.ts, ~line 1720):
When GPT-5.4's turn ends with tool results + continuation intent + no completion language, the runner injects the continuation instruction and continues back to the while loop instead of returning. 11-condition safety guard includes budget, abort, timeout, tool errors, messaging, side effects, approval prompts, and completion language.

Commit 3: Review fixes

Started-count bypass — both the hasNonPlanToolActivity AND startedCount > planOnlyToolMetaCount guards now have the isSingleActionThenNarrativePattern exception (Codex-connector P1 finding: the started-count guard was neutralizing the fix).
Approval safety — OPENAI_GPT5_TOOL_CALL_STYLE now includes the three approval safety lines from the default section (Copilot finding: the override was dropping safety guidance).
Side-effect guard — auto-continuation now checks !attempt.replayMetadata.hadPotentialSideEffects (Codex-connector P1 finding: mutating work could be duplicated on retry).
JSDoc — 700-char blind spot documented on isSingleActionThenNarrativePattern.

Safety rails on auto-continuation

Condition	Prevents
`autoContinuationCount < continuationBudget`	Runaway loops (max 5 turns)
`!aborted && !timedOut`	Continuing after cancel/timeout
`!attempt.clientToolCall`	Continuing when client-side tool execution needed
`!attempt.yieldDetected`	Continuing when model explicitly yielded
`!attempt.lastToolError`	Continuing after a tool error
`!attempt.didSendViaMessagingTool`	Continuing after model already sent a message
`!attempt.didSendDeterministicApprovalPrompt`	Continuing when explicit approval needed
`!attempt.replayMetadata.hadPotentialSideEffects`	Continuing after mutating work (writes, sends)
`toolMetas.some(e => e.toolName !== "update_plan")`	Only continuing when real tool work happened
`hasContinuationIntent(...)`	Only continuing when model shows future-work language
`!hasCompletionLanguage(...)`	Not continuing when model says "done" / "finished"

Tests

src/agents/pi-embedded-runner/run.incomplete-turn.test.ts — 16 passed
src/agents/system-prompt.test.ts — 52 passed
TypeScript check: 0 errors from changed files

Claude is unaffected: continuation loop defaults to "prompt" for non-strict-agentic runs; prompt overlay only applies to GPT-5 on openai/openai-codex providers.

Projected scorecard

Objective	Before	After
Follow instructions	7/10	9/10
Use tools proactively	7/10	9/10
Multi-step execution	6/10	10/10
No over-confirmation	6/10	9/10
Overall	6.5/10	9.25/10

How this PR relates to the others

agents: GPT-5.4 runtime completion rollup #65219 auto-activates the contract this PR's behavioral fixes build on
agents: GPT-5.4 parity proof rollup #65224 measures the behavior this PR improves — parity scenarios should pass more consistently after this lands
This PR is the behavioral fix — it addresses what users actually experience

…ative loophole, add tool-first reinforcement

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9490e00d38

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Copilot

Pull request overview

Strengthens GPT-5.x agent “act-first” behavior in OpenClaw by (1) tightening the OpenAI prompt overlay to more aggressively reinforce tool execution, and (2) extending the planning-only retry guard to catch “one tool call then planning narration” turns that previously escaped retries.

Changes:

Updated GPT-5 execution-bias overlay text to explicitly require tool/action-first and continued multi-step tool use until completion or a real blocker.
Added a new heuristic (isSingleActionThenNarrativePattern) so a single non-plan tool call followed by planning prose can still be treated as “planning-only” for retry purposes.
Added an OpenAI GPT-5 tool_call_style section override to reinforce direct tool calling without interleaved narration.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
src/agents/pi-embedded-runner/run/incomplete-turn.ts	Adds detection to close the “single tool call then plan/permission ask” retry loophole.
extensions/openai/prompt-overlay.ts	Strengthens GPT-5 execution bias and introduces a GPT-5 tool-call style override.

greptile-apps · 2026-04-12T08:35:53Z

Greptile Summary

This PR closes the "one action then narrative" loophole in the GPT-5.4 planning-only retry detector and adds a new prompt-overlay.ts module with stronger execution-bias and tool-call-style overrides. The detection logic in isSingleActionThenNarrativePattern is sound — 2+ non-plan tool calls and mutating single calls remain exempt, while 1 non-mutating tool call followed by planning prose is now correctly eligible for retry.

Two P2 items to consider before merge:

The existing test "does not retry planning-only detection after tool activity" continues to pass but now via the hadPotentialSideEffects bail path (because bash is mutating), not the hasNonPlanToolActivity bail. The new primary behavior — 1 non-mutating read-only tool call + planning prose → retry — has no dedicated test.
isSingleActionThenNarrativePattern silently falls back to old "no retry" behavior for planning texts longer than 700 characters; this edge case is not documented in the function's JSDoc.

Confidence Score: 5/5

Safe to merge; no P0/P1 defects found. Both findings are non-blocking P2 suggestions around test coverage and JSDoc clarity.

All changes are logically sound. The loophole fix correctly threads the needle between 0-tool, 1-non-mutating-tool, 1-mutating-tool, and 2+-tool scenarios. Existing tests still pass. Remaining findings are purely test-gap and documentation quality concerns.

No files require special attention for correctness; consider adding a test for the 1-non-mutating-tool-call + planning-prose path in run.incomplete-turn.test.ts.

Comments Outside Diff (1)

src/agents/pi-embedded-runner/run.incomplete-turn.test.ts, line 165-178 (link)

Test coverage gap for new core behavior

This test still passes after the change, but now for a different reason: bash is in MUTATING_TOOL_NAMES, so hadPotentialSideEffects = true and the function bails at the replayMetadata.hadPotentialSideEffects guard rather than at the old hasNonPlanToolActivity guard. The test title "does not retry planning-only detection after tool activity" no longer accurately describes the new invariant, which is "does not retry after mutating tool activity."

More importantly, the primary new behavior — 1 non-mutating tool call (e.g. read_file) + planning prose → should retry — has no test coverage at all. Consider adding a case like:

it("retries after a single read-only tool call followed by planning-only prose", () => {
  const retryInstruction = resolvePlanningOnlyRetryInstruction({
    provider: "openai",
    modelId: "gpt-5.4",
    aborted: false,
    timedOut: false,
    attempt: makeAttemptResult({
      assistantTexts: ["I'll now analyze the structure and make the changes."],
      toolMetas: [{ toolName: "read", meta: "src/foo.ts" }],
    }),
  });

  expect(retryInstruction).toContain("Act now");
});

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/run.incomplete-turn.test.ts
Line: 165-178

Comment:
**Test coverage gap for new core behavior**

This test still passes after the change, but now for a different reason: `bash` is in `MUTATING_TOOL_NAMES`, so `hadPotentialSideEffects = true` and the function bails at the `replayMetadata.hadPotentialSideEffects` guard rather than at the old `hasNonPlanToolActivity` guard. The test title "does not retry planning-only detection after tool activity" no longer accurately describes the new invariant, which is "does not retry after *mutating* tool activity."

More importantly, the primary new behavior — **1 non-mutating tool call (e.g. `read_file`) + planning prose → should retry** — has no test coverage at all. Consider adding a case like:

```ts
it("retries after a single read-only tool call followed by planning-only prose", () => {
  const retryInstruction = resolvePlanningOnlyRetryInstruction({
    provider: "openai",
    modelId: "gpt-5.4",
    aborted: false,
    timedOut: false,
    attempt: makeAttemptResult({
      assistantTexts: ["I'll now analyze the structure and make the changes."],
      toolMetas: [{ toolName: "read", meta: "src/foo.ts" }],
    }),
  });

  expect(retryInstruction).toContain("Act now");
});
```

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

Prompt To Fix All With AI

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/run.incomplete-turn.test.ts
Line: 165-178

Comment:
**Test coverage gap for new core behavior**

This test still passes after the change, but now for a different reason: `bash` is in `MUTATING_TOOL_NAMES`, so `hadPotentialSideEffects = true` and the function bails at the `replayMetadata.hadPotentialSideEffects` guard rather than at the old `hasNonPlanToolActivity` guard. The test title "does not retry planning-only detection after tool activity" no longer accurately describes the new invariant, which is "does not retry after *mutating* tool activity."

More importantly, the primary new behavior — **1 non-mutating tool call (e.g. `read_file`) + planning prose → should retry** — has no test coverage at all. Consider adding a case like:

```ts
it("retries after a single read-only tool call followed by planning-only prose", () => {
  const retryInstruction = resolvePlanningOnlyRetryInstruction({
    provider: "openai",
    modelId: "gpt-5.4",
    aborted: false,
    timedOut: false,
    attempt: makeAttemptResult({
      assistantTexts: ["I'll now analyze the structure and make the changes."],
      toolMetas: [{ toolName: "read", meta: "src/foo.ts" }],
    }),
  });

  expect(retryInstruction).toContain("Act now");
});
```

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/agents/pi-embedded-runner/run/incomplete-turn.ts
Line: 307-320

Comment:
**Silent 700-char blind spot not documented in JSDoc**

`isSingleActionThenNarrativePattern` returns `false` when `text.length > 700`, which means a model that does 1 `read_file` then writes 800 characters of planning prose is silently treated as "real progress" and is **not retried** — the opposite of the loophole this function aims to close. The behavior is consistent with the downstream cap in `resolvePlanningOnlyRetryInstruction`, but nothing in the JSDoc mentions this edge case. Since a long verbose response is actually a common symptom of the "one step then narrate" pattern, it's worth calling out explicitly:

```ts
 * Note: turns where the text exceeds 700 characters are not matched —
 * they fall through to the downstream text-length guard in
 * resolvePlanningOnlyRetryInstruction and are treated as non-retryable,
 * consistent with the zero-tool-call path.
```

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

_{Reviews (1): Last reviewed commit: "agents: strengthen GPT-5.4 execution bia..." | Re-trigger Greptile}

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f725a590ac

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…side-effect guard, JSDoc

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e44b709336

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

… remove as-casts

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a8055323fa

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Copilot

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 2 comments.

…hema help entries

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a64c7af40e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Copilot

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated 3 comments.

…ride

…ent leak on retry

…nguage

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 82b60b8b6c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

… tool_call_style via stablePrefix

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9de4ac9ef6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Copilot

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated 3 comments.

…add single-action unit tests

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b3b717b2de

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

pashpashpash · 2026-04-13T01:53:06Z

Maintainer update:

We split the narrow GPT-5 prompt / one-action-then-narrative retry slice out of this branch and landed it separately in #65597.

Landed commit:
c848ebc

Source head from this PR at the time of follow-up triage:
b3b717b

The remaining auto-continuation / config-surface work in this PR was intentionally not landed as part of that slice. It is a larger product behavior change and should come back, if desired, as a focused follow-up against current main rather than bundled with the prompt / retry fixes.

Closing this PR as superseded by the landed split.

agents: strengthen GPT-5.4 execution bias, close one-action-then-narr…

9490e00

…ative loophole, add tool-first reinforcement

Copilot AI review requested due to automatic review settings April 12, 2026 08:28

openclaw-barnacle Bot added agents Agent runtime and tooling extensions: openai size: S labels Apr 12, 2026

Copilot started reviewing on behalf of 100yenadmin April 12, 2026 08:29 View session

chatgpt-codex-connector Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts Outdated

Copilot AI reviewed Apr 12, 2026

View reviewed changes

Comment thread extensions/openai/prompt-overlay.ts Outdated

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts Outdated

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts

100yenadmin mentioned this pull request Apr 12, 2026

GPT-5.4 / Codex agentic runtime parity in OpenClaw #64227

Closed

greptile-apps Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts

agents: add auto-continuation loop for GPT-5.4 multi-step execution

f725a59

openclaw-barnacle Bot added docs Improvements or additions to documentation size: M and removed size: S labels Apr 12, 2026

chatgpt-codex-connector Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts Outdated

Comment thread src/agents/pi-embedded-runner/run.ts

agents: fix review findings — started-count bypass, approval safety, …

e44b709

…side-effect guard, JSDoc

This was referenced Apr 12, 2026

agents: GPT-5.4 runtime completion rollup #65219

Merged

agents: GPT-5.4 parity proof rollup #65224

Closed

chatgpt-codex-connector Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run.ts

Comment thread src/agents/agent-scope.ts Outdated

agents: fix completion-language false positive, typed config surface,…

a805532

… remove as-casts

chatgpt-codex-connector Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts Outdated

100yenadmin requested a review from Copilot April 12, 2026 10:00

Copilot started reviewing on behalf of 100yenadmin April 12, 2026 10:01 View session

Copilot AI reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/agent-scope.ts Outdated

Comment thread src/config/schema.labels.ts

agents: tighten continuation-intent regex, narrow config type, add sc…

a64c7af

…hema help entries

chatgpt-codex-connector Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts

Comment thread src/config/zod-schema.agent-runtime.ts

100yenadmin requested a review from Copilot April 12, 2026 10:32

Copilot started reviewing on behalf of 100yenadmin April 12, 2026 10:32 View session

Copilot AI reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run.ts

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts

Eva added 3 commits April 12, 2026 18:00

test: update OpenAI extension test snapshots for tool_call_style over…

80af5a2

…ride

agents: clear auto-continuation instruction after consumption to prev…

ac4aa5a

…ent leak on retry

test: add unit coverage for hasContinuationIntent and hasCompletionLa…

82b60b8

…nguage

chatgpt-codex-connector Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts Outdated

Comment thread extensions/openai/prompt-overlay.ts Outdated

agents: use tighter regex in single-action detector, preserve dynamic…

9de4ac9

… tool_call_style via stablePrefix

chatgpt-codex-connector Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts

100yenadmin requested a review from Copilot April 12, 2026 21:09

Copilot started reviewing on behalf of 100yenadmin April 12, 2026 21:10 View session

Copilot AI reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/agent-scope.ts Outdated

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts Outdated

Comment thread src/agents/pi-embedded-runner/run.incomplete-turn.test.ts

agents: sync per-agent embeddedPi type, extract single-action check, …

b3b717b

…add single-action unit tests

chatgpt-codex-connector Bot reviewed Apr 12, 2026

View reviewed changes

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts

Comment thread src/agents/pi-embedded-runner/run/incomplete-turn.ts

pashpashpash mentioned this pull request Apr 12, 2026

agents: split GPT-5 prompt and retry behavior #65597

Merged

pashpashpash closed this Apr 13, 2026

gugu91 mentioned this pull request Apr 15, 2026

research: investigate OpenClaw changes that improved OAI model behavior gugu91/extensions#419

Closed

Uh oh!

Conversation

100yenadmin commented Apr 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What changed (3 commits)

Commit 1: Prompt + detection layer

Commit 2: Auto-continuation loop

Commit 3: Review fixes

Safety rails on auto-continuation

Tests

Projected scorecard

How this PR relates to the others

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

greptile-apps Bot commented Apr 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 5/5

Comments Outside Diff (1)

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

100yenadmin commented Apr 12, 2026 •

edited

Loading

greptile-apps Bot commented Apr 12, 2026 •

edited

Loading