Repository navigation
Add an optional integrator control socket - #185
Open
t-kalinowski wants to merge 13 commits into
Open
t-kalinowski wants to merge 13 commits into
t-kalinowski wants to merge 13 commits into
Conversation
Keep integrator evaluation admission, results, and recording projections separate from MCP delivery while sharing the existing worker and sandbox. Expose strict JSONL status, preview, requirements, and evaluation requests.
Apply the existing test-only DuckDB snapshot to direct and MXC SQL scenarios. Assert the selected version before the long query so cold caches cannot silently introduce the known Arrow cancellation limitation. Preserve production defaults, interruption deadlines, and exact results.
Queue stdin while the replacement cell is held at its checkpoint, then release it and collect the complete result. Preserve the exact output and idle-state assertions without assuming one response completes the cell.
Assign the new guide to the Build sidebar and require its task link in the rendered website checks.
Queue semantic control evaluations behind retained MCP delivery ownership. Hold a complete image response before its transport acknowledgment to reproduce the busy response deterministically, then prove the integrator request waits until delivery is acknowledged. Preserve explicit reject and timeout coverage and the exact image, raw output, and projection assertions.
Preserve the immediate kernel-exit assertion and shutdown deadline while showing which captured identities are still present if it fails again.
Keep the required completion marker when a control-and-cell response is split across polls. Exercise the split with a gated native R startup cell and select the same contract in related restart continuations.
Preserve the launch diagnostic for MCP while rejecting integrator admission before it can consume the failure and report an unexecuted cell as complete.
Keep cancelled MCP replies out of integrator previews and results, including restart settlement. Retain repeated recovered replies within the existing output bounds and restore them before restart notices.
t-kalinowski
marked this pull request as ready for review
October 27, 2026 12:35
Member
Author
|
No findings found by gpt-6-astra with reasoning effort max Reviewed commit: |
This branch has not been deployed
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
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.
Applications sharing an agent's Console session can inspect it and submit code through an optional
serve --control-socket PATHUnix socket. The strict JSON Lines interface exposes status, non-consuming output previews, accepted requirements, and R/Python/SQL evaluation. Evaluations can reject a busy session or wait with an admission deadline. Queued work is cancelled on disconnect or worker replacement.Integrator evaluations retain their own results while sharing the existing session and sandbox. Every request and result is journaled;
record: falseomits an evaluation from the Markdown transcript and QMD source projection. Internally, execution and delivery ownership distinguish integrator calls from MCP calls without changing the worker wire protocol.The channel is available on macOS and Linux. Execution remains serialized; concurrent R top-level evaluation and Python background execution are unsupported. The guide documents the protocol, output-attribution limits, and admission after MCP delivery, and is included in the website navigation.
Public fixtures hold a fully written MCP response before its server acknowledgment to verify that integrator admission waits for retained delivery ownership. They preserve exact image bytes, omission accounting, recording projections, and busy-rejection checks. Retirement fixtures also include captured process identities in failure diagnostics. The shared transcript collector preserves the final
[done]marker across polls for combined control-and-cell calls and their continuations. A native R startup fixture holds the replacement cell until its earlier output is collected, then checks exact completion. Production package defaults and retirement assertions are unchanged.On macOS, all 16 control and image cases passed across 21 direct/Seatbelt executions; the focused run including replacement-startup and process-retirement checks passed 18 cases across 25 executions, with the Linux-only adopted-resource case skipped. All 36 rendered website tests,
scripts/format,git diff --check, andscripts/checkpassed. The response-acknowledgment race and both website failures have local red/green evidence.The controlled-completion repair reproduced the missing marker locally and passed eight focused macOS cases across 13 direct/Seatbelt executions, the final formatted gated case, and all 23 client self-tests. Formatting, diff checks, and
scripts/checkpassed.On native Ubuntu 24.04 (
mule), PR heade5c7874cpassedscripts/check, all 23 client self-tests, and 49 focused cases across 66 executions. The focused run includes direct/bubblewrap control-channel tests, the Linux response-acknowledgment checkpoint, adopted-resource retirement, startup completion, and recovery allocation stress. The macOS-only Seatbelt prelaunch-failure case was skipped. Validation used Python 3.14.8 and uv 0.13.1.The hosted Windows runner does not provide MXC, so native Windows sandbox execution remains unverified. Hosted checks for the current head are pending.