refactor: Avoid synthetic prompt for approval continuations by daryllimyt · Pull Request #2598 · TracecatHQ/tracecat

daryllimyt · 2026-05-01T15:50:17Z

Summary

replace visible synthetic approval continuation prompts with a hidden Claude Code isMeta continuation tick
preseed/reconcile Claude-format user/tool_result history rows so the model sees the real approved or denied tool result
preserve legitimate thinking from the resumed assistant turn while hiding only SDK continuation metadata and synthetic artifacts
replace legacy approval interrupt placeholders before duplicate detection so existing pending approvals can reconcile correctly
update workflow, runtime, session activity, socket, adapter, unit, Temporal, and integration coverage for approval continuation ordering

Existing Pending Approvals

This should be safe for workflows that are currently waiting on approval. Those workflows are paused before post-approval execution at await self.approvals.wait(). After the approval signal arrives, the new code runs the post-approval path: execute approved/denied tools, reconcile the session history by inserting the real user/tool_result row, reload the session history, then resume Claude Code with the hidden meta continuation tick.

Existing pending approvals should already have the assistant tool_use row and Claude Code interrupt artifacts in session history. The reconciliation step is designed for that state: find the matching assistant tool_use, remove interrupt artifacts including legacy "interrupted" placeholders, and insert the real tool result immediately after the tool use.

Caveats:

Workflows that already received approval and are mid-continuation under the old behavior may still be in a bad/stuck state and are not guaranteed to self-heal.
If session history is malformed or missing the matching assistant tool_use, reconciliation logs and returns without inserting a tool_result.
The workflow worker and agent executor should both be running this version; mixed old/new workers are risky for this path.

Attempted Alternative and Limitation

We also tried a cleaner continuation shape: keep session service as the canonical owner of the approved user/tool_result row, seed Claude Code with history only through the original assistant tool_use, then send that final tool_result as the first resumed SDK stdin message instead of using a hidden Continue. tick.

That does not work correctly with the current Claude Code SDK/CLI behavior. In live cluster testing, the DB showed the original approved tool result persisted correctly, but Claude Code inserted a synthetic assistant row (<synthetic> / No response requested.) before the replayed stdin tool_result. That broke the clean assistant tool_use -> user tool_result adjacency and the model requested the same tool again, creating a second pending approval for the same user request.

TODO: revisit this if Claude Code exposes a supported “resume and consume pending tool_result” API, or if the SDK/CLI allows passing a tool result without inserting the synthetic assistant placeholder. Until then, this PR keeps the more reliable design: preseed the reconciled tool_result into session history and use a hidden metadata continuation tick, while filtering those control-plane rows from normal chat history.

Tests

uv run pytest tests/unit/test_agent_runtime.py tests/unit/test_agent_session_activities.py tests/unit/test_agent_session_approval_sink.py::test_replace_interrupt_with_tool_results_replaces_legacy_interrupted_row tests/temporal/test_durable_agent_workflow.py::test_agent_workflow_routes_approved_tools_to_executor_and_reconciles_history tests/temporal/test_durable_agent_workflow.py::test_agent_workflow_does_not_retry_approved_tool_failures -q
uv run ruff check packages/tracecat-ee/tracecat_ee/agent/workflows/durable.py tests/integration/test_agent_worker.py tests/temporal/test_durable_agent_workflow.py tests/unit/test_agent_runtime.py tests/unit/test_agent_session_activities.py tests/unit/test_agent_session_approval_sink.py tests/unit/test_agent_session_messages.py tests/unit/test_agent_socket_io.py tests/unit/test_vercel_adapter.py tracecat/agent/adapter/vercel.py tracecat/agent/common/protocol.py tracecat/agent/executor/activity.py tracecat/agent/executor/schemas.py tracecat/agent/runtime/claude_code/runtime.py tracecat/agent/session/activities.py tracecat/agent/session/service.py
uv run ruff format --check packages/tracecat-ee/tracecat_ee/agent/workflows/durable.py tests/integration/test_agent_worker.py tests/temporal/test_durable_agent_workflow.py tests/unit/test_agent_runtime.py tests/unit/test_agent_session_activities.py tests/unit/test_agent_session_approval_sink.py tests/unit/test_agent_session_messages.py tests/unit/test_agent_socket_io.py tests/unit/test_vercel_adapter.py tracecat/agent/adapter/vercel.py tracecat/agent/common/protocol.py tracecat/agent/executor/activity.py tracecat/agent/executor/schemas.py tracecat/agent/runtime/claude_code/runtime.py tracecat/agent/session/activities.py tracecat/agent/session/service.py
uv run basedpyright --warnings packages/tracecat-ee/tracecat_ee/agent/workflows/durable.py tests/integration/test_agent_worker.py tests/temporal/test_durable_agent_workflow.py tests/unit/test_agent_runtime.py tests/unit/test_agent_session_activities.py tests/unit/test_agent_session_approval_sink.py tests/unit/test_agent_session_messages.py tests/unit/test_agent_socket_io.py tests/unit/test_vercel_adapter.py tracecat/agent/adapter/vercel.py tracecat/agent/common/protocol.py tracecat/agent/executor/activity.py tracecat/agent/executor/schemas.py tracecat/agent/runtime/claude_code/runtime.py tracecat/agent/session/activities.py tracecat/agent/session/service.py

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: be1d4381c1

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".

zeropath-ai · 2026-05-01T16:00:17Z

✅ No security or compliance issues detected. Reviewed everything up to 70c4cfe.

Security Overview

🔎 Scanned files: 16 changed file(s)
🔗 Scan Link: https://zeropath.com/app/repositories/00dffd6c-8834-4dc9-b6d8-b44cd1622986?scanId=a51e37d1-8d49-4547-a38c-abf739aae976&codeScanTypes=PrScan&tab=issues

Detected Code Changes

Change Type	Relevant files
Enhancement	► tracecat_ee/tracecat_ee/agent/workflows/durable.py Reconcile SDK transcript with tool results Update executor input for resume with reconciled tool results ► tests/integration/test_agent_worker.py Test approval continuation resumes with seeded tool result ► tests/temporal/test_durable_agent_workflow.py Test approval continuation resumes with seeded tool result Test approval continuation forwards live thinking events Test approval continuation hides SDK meta prompt only ► tests/unit/test_agent_runtime.py Approval continuation sends meta prompt in each sandbox mode ► tests/unit/test_agent_session_activities.py Reconcile tool results removes interrupts and returns tool results ► tests/unit/test_agent_session_approval_sink.py Replace interrupt with tool results replaces legacy interrupted row Replace interrupt with tool results does not duplicate existing result Replace interrupt with tool results preserves interrupt without assistant uuid Replace interrupt with tool results requires same assistant turn
Refactor	► tests/unit/test_agent_runtime.py Update ClaudeSDKClient mock to include connect and disconnect Rename test_approval_continuation_uses_hidden_prompt_in_each_sandbox_mode to test_approval_continuation_sends_meta_prompt_in_each_sandbox_mode

cubic-dev-ai

1 issue found across 14 files

Confidence score: 5/5

Low-risk change overall: the only reported issue is low severity (3/10) and appears confined to test coverage rather than runtime behavior.
The key concern is in tests/temporal/test_durable_agent_workflow.py, where dropping the failure payload assertion weakens retry validation and could let an error-result write regression slip through.
Pay close attention to tests/temporal/test_durable_agent_workflow.py - restore the failure payload assertion so the retry path still verifies the error result is persisted.

Prompt for AI agents (unresolved issues)


Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="tests/temporal/test_durable_agent_workflow.py">

<violation number="1" location="tests/temporal/test_durable_agent_workflow.py:816">
P3: Restore the failure payload assertion here; otherwise the retry test no longer verifies the error result was written.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d255055467

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".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e1994872c0

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".

…AgentRuntime

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 44b691d305

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".

daryllimyt · 2026-05-03T02:21:24Z

feat: Add agent subagent support #2564
refactor: Avoid synthetic prompt for approval continuations #2598 👈 (View in Graphite)
main

This stack of pull requests is managed by Graphite. Learn more about stacking.

blacksmith-sh · 2026-05-03T02:27:45Z

Found 7 test failures on Blacksmith runners:

Failures

Test	View Logs
`test_workflows/test_child_workflow_parallel_loop_with_dispatch_cap`	View Logs
`test_workflows/test_scatter_with_child_workflow`	View Logs
`test_workflows/test_workflow_join_run_if_skip_ok`	View Logs
`test_workflows/test_workflow_scatter_gather[basic-for-loop]`	View Logs
`test_workflows/test_workflow_scatter_gather[nested-for-loop]`	View Logs
`test_workflows/test_workflow_scatter_gather[scatter-gather-with-surrounding-actions]`	View Logs
`test_workflows/test_workflow_table_actions_in_loop`	View Logs

daryllimyt temporarily deployed to internal-registry-ci May 1, 2026 15:50 — with GitHub Actions Inactive

chatgpt-codex-connector Bot reviewed May 1, 2026

View reviewed changes

Comment thread tracecat/agent/session/service.py Outdated

Comment thread tracecat/agent/runtime/claude_code/runtime.py Outdated

This comment has been minimized.

Sign in to view

cubic-dev-ai Bot reviewed May 1, 2026

View reviewed changes

Comment thread tests/temporal/test_durable_agent_workflow.py Outdated

daryllimyt temporarily deployed to internal-registry-ci May 1, 2026 16:23 — with GitHub Actions Inactive

daryllimyt added 4 commits May 2, 2026 17:51

Avoid synthetic prompt for approval continuations

cda1def

Pass approval tool result during SDK connect

ba03331

Seed approval tool results before resume

d777755

Update socket IO SDK client mock

d255055

daryllimyt force-pushed the feat/cleanup-approval-continuation branch from 380c6dd to d255055 Compare May 2, 2026 21:51

daryllimyt temporarily deployed to internal-registry-ci May 2, 2026 21:51 — with GitHub Actions Inactive

daryllimyt temporarily deployed to internal-registry-ci May 2, 2026 21:52 — with GitHub Actions Inactive

This comment has been minimized.

Sign in to view

chatgpt-codex-connector Bot reviewed May 2, 2026

View reviewed changes

Comment thread tracecat/agent/session/service.py Outdated

daryllimyt added 3 commits May 2, 2026 18:42

fix: preserve thinking during approval continuation

0b74438

fix: replace legacy approval interrupts

b639830

test: assert approval continuation tool results

dc66265

daryllimyt temporarily deployed to internal-registry-ci May 2, 2026 22:47 — with GitHub Actions Inactive

daryllimyt changed the title ~~Avoid synthetic prompt for approval continuations~~ refactor: Avoid synthetic prompt for approval continuations May 2, 2026

Optimize approval tool result idempotency check

e199487

daryllimyt temporarily deployed to internal-registry-ci May 2, 2026 23:41 — with GitHub Actions Inactive

chatgpt-codex-connector Bot reviewed May 2, 2026

View reviewed changes

Comment thread tracecat/agent/session/service.py

Comment thread tracecat/agent/executor/activity.py Outdated

refactor: tighten claude runtime lifecycle

ccc0888

daryllimyt temporarily deployed to internal-registry-ci May 2, 2026 23:49 — with GitHub Actions Inactive

This comment has been minimized.

Sign in to view

daryllimyt added 3 commits May 2, 2026 20:00

refactor: improve handling of approval continuation prompts in Claude…

a8ba631

…AgentRuntime

refactor: clarify approval interrupt checks

70a3748

refactor: share claude session line helpers

6e18551

daryllimyt temporarily deployed to internal-registry-ci May 3, 2026 00:29 — with GitHub Actions Inactive

refactor: move approval interrupt classifiers

297ffb7

daryllimyt temporarily deployed to internal-registry-ci May 3, 2026 00:33 — with GitHub Actions Inactive

refactor: remove stale approval runtime payload

a2e0459

daryllimyt temporarily deployed to internal-registry-ci May 3, 2026 00:45 — with GitHub Actions Inactive

daryllimyt added 2 commits May 2, 2026 21:55

fix: preserve interrupts without assistant uuid

08e035e

fix: avoid duplicating approval results

44b691d

daryllimyt temporarily deployed to internal-registry-ci May 3, 2026 01:58 — with GitHub Actions Inactive

This comment has been minimized.

Sign in to view

chatgpt-codex-connector Bot reviewed May 3, 2026

View reviewed changes

Comment thread tracecat/agent/session/service.py Outdated

fix: require same assistant for approval reconciliation

70c4cfe

daryllimyt temporarily deployed to internal-registry-ci May 3, 2026 02:18 — with GitHub Actions Inactive

daryllimyt mentioned this pull request May 3, 2026

feat: Add agent subagent support #2564

Open

daryllimyt merged commit 879e829 into main May 3, 2026
20 of 22 checks passed

daryllimyt deleted the feat/cleanup-approval-continuation branch May 3, 2026 02:35

Conversation

daryllimyt commented May 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Existing Pending Approvals

Attempted Alternative and Limitation

Tests

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

zeropath-ai Bot commented May 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

This comment has been minimized.

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

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!

This comment has been minimized.

This comment has been minimized.

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

daryllimyt commented May 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

blacksmith-sh Bot commented May 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Failures

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

daryllimyt commented May 1, 2026 •

edited

Loading

zeropath-ai Bot commented May 1, 2026 •

edited

Loading

daryllimyt commented May 3, 2026 •

edited

Loading

blacksmith-sh Bot commented May 3, 2026 •

edited

Loading