feat: native interrupt + Task-subagent goldens (pre-metering)#14
Merged
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b905b020be
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".
| "type": "user", | ||
| "content": "Use the Task tool to launch a subagent that answers this question: what is 7 times 6? Report only the subagent's answer." | ||
| }, | ||
| { "type": "wait_control_request", "timeoutMs": 60000 }, |
There was a problem hiding this comment.
In the recorded task-subagent run, native emits no control_request for this prompt, so this wait_control_request delays the sequence until its 60s timeout; by then the result.success has already been seen, and the following wait_result only waits for a new result, causing the harness to hit its 240s timeout and kill the process (the added report records exit code 143). This makes this golden both slow and contaminated with a SIGTERM exit whenever the ground-truth behavior is actually being reproduced; remove the permission wait/response or start waiting for the result immediately.
Useful? React with 👍 / 👎.
The last irreplaceable native ground truth before claude -p bills at metered API pricing on 2026-06-15. Both sets recorded against native 2.1.170 (sonnet). - 2026-06-10-interrupt/: the 7-case interrupt matrix (cases-interrupt.json): control-protocol interrupt and SIGINT during a long task, double interrupt, and interrupt-then-prompt pipelines. Native answers an interrupted turn with result.error_during_execution; follow-up prompts produce result.success. - 2026-06-10-task-subagent/: new case file for a Task subagent round-trip (tool name "Agent" in 2.1.170). Locks in two facts clarp must match: native emits ZERO sidechain lines on -p stdout, and emits system.task_started / system.task_notification events for the subagent lifecycle. The case waits directly for the result — the Agent tool needs no permission, so a wait_control_request would idle 60s and poison the recorded exit code with the harness's timeout SIGTERM (caught by review on the first capture; re-captured cleanly: exit 0, result at ~36s).
dn00
force-pushed
the
feat/golden-interrupt-task
branch
from
b905b02 to
d3cc6fb
Compare
dn00
added a commit
that referenced
this pull request
Jun 10, 2026
…pe, thinking signature Closes the consumed-field parity gaps confirmed by replay against the 2026-06-09 goldens (PARITY-GAPS #2/#3/#14/#15/#16), and adds the parity-policy tiers to PARITY-GAPS.md (contract / consumed fields / on-demand / won't-fix). - --max-turns now counts assistant API rounds (SSE message_start) like native, not user prompts — a single prompt chaining tools can exhaust it. On exceeding: interrupt, then report result.error_max_turns (is_error, NO result field, num_turns = rounds) and exit 1, even in stream-json mode, whether or not the interrupt beat the turn's natural completion. The old prompt-count check remains as a fallback for backends without SSE. - Ack initialize (truthful payload: commands from transcript init, output_style, pid) and set_permission_mode (echo mode) using native's nested control_response shape {type, response: {subtype, request_id, response}}; migrated the interrupt ack from clarp's old flat shape. - Turn-level backend API errors report subtype "success" with is_error: true (invalid-model golden), not subtype "error"; emitResult decouples is_error from subtype and can omit the result field (error_max_turns has none). - Assembler accumulates signature_delta; thinking blocks are {type, thinking, signature} like native. - Exit code now reflects the last result's is_error on stdin-drain shutdown (native exits 1 after an invalid-model turn even though the stream ends normally).
dn00
force-pushed
the
feat/golden-interrupt-task
branch
from
d3cc6fb to
3f0d23a
Compare
dn00
force-pushed
the
feat/golden-interrupt-task
branch
from
3f0d23a to
5fe0331
Compare
dn00
added a commit
that referenced
this pull request
Jun 10, 2026
…pe, thinking signature Closes the consumed-field parity gaps confirmed by replay against the 2026-06-09 goldens (PARITY-GAPS #2/#3/#14/#15/#16), and adds the parity-policy tiers to PARITY-GAPS.md (contract / consumed fields / on-demand / won't-fix). - --max-turns now counts assistant API rounds (SSE message_start) like native, not user prompts — a single prompt chaining tools can exhaust it. On exceeding: interrupt, then report result.error_max_turns (is_error, NO result field, num_turns = rounds) and exit 1, even in stream-json mode, whether or not the interrupt beat the turn's natural completion. The old prompt-count check remains as a fallback for backends without SSE. - Ack initialize (truthful payload: commands from transcript init, output_style, pid) and set_permission_mode (echo mode) using native's nested control_response shape {type, response: {subtype, request_id, response}}; migrated the interrupt ack from clarp's old flat shape. - Turn-level backend API errors report subtype "success" with is_error: true (invalid-model golden), not subtype "error"; emitResult decouples is_error from subtype and can omit the result field (error_max_turns has none). - Assembler accumulates signature_delta; thinking blocks are {type, thinking, signature} like native. - Exit code now reflects the last result's is_error on stdin-drain shutdown (native exits 1 after an invalid-model turn even though the stream ends normally).
dn00
added a commit
that referenced
this pull request
Jun 10, 2026
…pe, thinking signature Closes the consumed-field parity gaps confirmed by replay against the 2026-06-09 goldens (PARITY-GAPS #2/#3/#14/#15/#16), and adds the parity-policy tiers to PARITY-GAPS.md (contract / consumed fields / on-demand / won't-fix). - --max-turns now counts assistant API rounds (SSE message_start) like native, not user prompts — a single prompt chaining tools can exhaust it. On exceeding: interrupt, then report result.error_max_turns (is_error, NO result field, num_turns = rounds) and exit 1, even in stream-json mode, whether or not the interrupt beat the turn's natural completion. The old prompt-count check remains as a fallback for backends without SSE. - Ack initialize (truthful payload: commands from transcript init, output_style, pid) and set_permission_mode (echo mode) using native's nested control_response shape {type, response: {subtype, request_id, response}}; migrated the interrupt ack from clarp's old flat shape. - Turn-level backend API errors report subtype "success" with is_error: true (invalid-model golden), not subtype "error"; emitResult decouples is_error from subtype and can omit the result field (error_max_turns has none). - Assembler accumulates signature_delta; thinking blocks are {type, thinking, signature} like native. - Exit code now reflects the last result's is_error on stdin-drain shutdown (native exits 1 after an invalid-model turn even though the stream ends normally).
dn00
added a commit
that referenced
this pull request
Jun 10, 2026
…pe, thinking signature Closes the consumed-field parity gaps confirmed by replay against the 2026-06-09 goldens (PARITY-GAPS #2/#3/#14/#15/#16), and adds the parity-policy tiers to PARITY-GAPS.md (contract / consumed fields / on-demand / won't-fix). - --max-turns now counts assistant API rounds (SSE message_start) like native, not user prompts — a single prompt chaining tools can exhaust it. On exceeding: interrupt, then report result.error_max_turns (is_error, NO result field, num_turns = rounds) and exit 1, even in stream-json mode, whether or not the interrupt beat the turn's natural completion. The old prompt-count check remains as a fallback for backends without SSE. - Ack initialize (truthful payload: commands from transcript init, output_style, pid) and set_permission_mode (echo mode) using native's nested control_response shape {type, response: {subtype, request_id, response}}; migrated the interrupt ack from clarp's old flat shape. - Turn-level backend API errors report subtype "success" with is_error: true (invalid-model golden), not subtype "error"; emitResult decouples is_error from subtype and can omit the result field (error_max_turns has none). - Assembler accumulates signature_delta; thinking blocks are {type, thinking, signature} like native. - Exit code now reflects the last result's is_error on stdin-drain shutdown (native exits 1 after an invalid-model turn even though the stream ends normally).
dn00
added a commit
that referenced
this pull request
Jun 10, 2026
…pe, thinking signature Closes the consumed-field parity gaps confirmed by replay against the 2026-06-09 goldens (PARITY-GAPS #2/#3/#14/#15/#16), and adds the parity-policy tiers to PARITY-GAPS.md (contract / consumed fields / on-demand / won't-fix). - --max-turns now counts assistant API rounds (SSE message_start) like native, not user prompts — a single prompt chaining tools can exhaust it. On exceeding: interrupt, then report result.error_max_turns (is_error, NO result field, num_turns = rounds) and exit 1, even in stream-json mode, whether or not the interrupt beat the turn's natural completion. The old prompt-count check remains as a fallback for backends without SSE. - Ack initialize (truthful payload: commands from transcript init, output_style, pid) and set_permission_mode (echo mode) using native's nested control_response shape {type, response: {subtype, request_id, response}}; migrated the interrupt ack from clarp's old flat shape. - Turn-level backend API errors report subtype "success" with is_error: true (invalid-model golden), not subtype "error"; emitResult decouples is_error from subtype and can omit the result field (error_max_turns has none). - Assembler accumulates signature_delta; thinking blocks are {type, thinking, signature} like native. - Exit code now reflects the last result's is_error on stdin-drain shutdown (native exits 1 after an invalid-model turn even though the stream ends normally).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The last irreplaceable native ground truth before
claude -pbills at metered API pricing on 2026-06-15. Both sets recorded against native 2.1.170 (sonnet); pure additions, independent of #12/#13.2026-06-10-interrupt/— the 7-case interrupt matrix: control-protocol interrupt and SIGINT during a long task, double interrupt, interrupt-then-prompt pipelines. Native answers an interrupted turn withresult.error_during_execution; follow-up prompts produceresult.success.2026-06-10-task-subagent/— new case file for a Task subagent round-trip (tool nameAgentin 2.1.170). Locks in two facts clarp must match: native emits zero sidechain lines on-pstdout, and emitssystem.task_started/system.task_notificationfor the subagent lifecycle.