# Delegated completion acceptance — fresh evidence

For GitHub [#25666](https://github.com/pingdotgg/t3code/pull/25666), GitCafe [#25667](https://git.cafe/pingdotgg/t3code/pull/25667). **3/3 mandatory rows captured.** This is server-only evidence: no render path changed, no visual differences evaluated, no UI sheets required.

Environment: **fs-19**, Linux x86_64 (kernel 6.1.102, glibc 2.39), Node 26.12.1, Vite+ 1.4.0 / Vitest 5.0.5, Codex CLI **0.163.0**, model **gpt-6.1-sol** through the configured hub. Browser/device/DPR/font scaling/viewport/theme/frozen UI clock: N/A. R1/R2 are labelled in-memory test harness recordings; R3 is the real app and real Codex, driven through its typed Effect RPC transport.

Before implementation: [401b01747c](https://git.cafe/pingdotgg/t3code/commit/401b01747c1f7b331fb00b573e405a38f646fde8). After: [fdcf43b781](https://git.cafe/pingdotgg/t3code/commit/fdcf43b781a36813afeab8547a25adbde0c65b17), branch `fix/issue-25000`. The head checkout was clean before and after capture. Its source was not changed. Both suite runs use the exact head test blob `78bf5dcce18d7325ed9a150de8f07ff6066cc5a5`.

| Matrix row | Status | Capture / observable result |
| --- | --- | --- |
| R1 — base race | Captured, expected red | [Full base stdout/stderr](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/342cb04eb8a8eda1/r1-base.log): **1 failed, 30 passed**, exit 1. Only accept/checkpoint race fails. Both acknowledge race controls pass. |
| R2 — head race + guards | Captured, green | [Full head stdout/stderr](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c968602c3b2701fc/r2-head-verbose.log): **31 passed**, exit 0, including the accept row and existing delivery guards. |
| R3 — real acceptance + progress | Captured, green | [Chronological event transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6019de96a109f292/r3-transcript.txt), [full driver stdout/stderr](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/cdb7ee011274f78c/runtime.log): real child finishes during active parent, real steering succeeds, accept receipt persists, delivery is delivered, parent and follow-up complete with real Git checkpoints; exit 0. |
| U1 — optional UI video | Not captured, optional | Skipped: mandatory server transcripts fully cover the changed surface; no render changes. |
| X1 — Claude variant | Excluded by matrix | Unnecessary supplemental provider coverage; first-party Claude OAuth is unavailable on this VM per brief. Not attempted. |

No mandatory row is unverified. R3 proves production-route integration, **not the internal checkpoint/accept read-to-commit race**. R1/R2 deterministically force that interleaving. A base-side real-provider run was optional and was not attempted.

## Commands and isolation

```sh
git -C /home/ubuntu/t3code worktree add --detach /home/ubuntu/wt/pr-25666-base 401b01747c1f7b331fb00b573e405a38f646fde8
# In the disposable base checkout:
vp install
git -C /home/ubuntu/wt/pr-25666 show fdcf43b781a36813afeab8547a25adbde0c65b17:apps/server/src/orchestration-v2/DelegatedCompletionDelivery.test.ts > /home/ubuntu/wt/pr-25666-base/apps/server/src/orchestration-v2/DelegatedCompletionDelivery.test.ts
# Once per side, from that side's checkout; full verbose output and exit retained:
vp test run apps/server/src/orchestration-v2/DelegatedCompletionDelivery.test.ts --reporter=verbose
# Head runtime (spawn PID/process group 1667321; actual server 14000, web 5960):
T3CODE_DEV_SEED=0 vp run dev --home-dir /tmp/lead-25666-r1/evidence/20261027T062557Z/home
vp exec node /tmp/lead-25666-r1/evidence/rpc-driver.mjs /tmp/lead-25666-r1/evidence/20261027T062557Z
# Driver/collector inspect only this fresh database:
sqlite3 -readonly -json /tmp/lead-25666-r1/evidence/20261027T062557Z/home/userdata/statev2.sqlite 'select * from orchestration_events order by sequence'
```

The driver creates a synthetic Git repository, project and parent via normal app RPCs. The parent calls `orchestrator_capabilities`, then real `delegate_task` with `mode: async` (`completionWake: always`). Two external FIFOs hold real parent/child shell commands. The driver releases the child only after both runs are running, waits for the real accept receipt, then releases the parent. It never sends `notification.delivery.accept`, inserts fixture rows, or modifies SQLite. The manual follow-up uses normal `message.dispatch` RPC. All timestamps below are UTC on 2026-10-27.

## Key excerpts

R1 at 06:18:23: `actual: { status: 'waiting', completedAt: null, checkpointId: null }`; expected `completed`, `1970-01-01T00:00:00.000Z`, `checkpoint:accept-settle-race-completed`. R2 (full verbose output): `keeps a completed run that settled between accept's read and commit` passes.

R3's parent thread is `9a76b386-e88f-44d2-bd48-35f547155a6b`; parent run ordinal 1, follow-up run ordinal 2. Full child task/thread/run IDs, message ownership/generation and effect requests are preserved in the raw extracts.

```text
06:26:35.142 both-gates-ready: parent=running, child=running; child gate released
06:26:37.970 seq=83 child run.updated: completed, checkpoint present
06:26:38.263 seq=85 app_owned subagent.updated: completed, completionDelivery=claimed
06:26:38.263 seq=86 parent run.updated: running, completedAt=null, checkpointId=null
06:26:38.623 Codex outgoing id=6 turn/steer, expectedTurnId=01a17c09-bc52-7dd2-b93e-b8101af60c8d, CHILD_SYNTHETIC_DONE
06:26:38.625 Codex incoming id=6 result.turnId=01a17c09-bc52-7dd2-b93e-b8101af60c8d
06:26:38.646 receipt notification.delivery.accept: accepted, result_sequence=94, error=null
06:26:38.646 seq=93 app_owned subagent.updated: completionDelivery=delivered
06:26:38.646 seq=94 parent run.updated: running, completedAt=null, checkpointId=null (matches seq=86)
06:26:38.671 parent gate released, after accepted receipt was observed
06:26:44.795 seq=110 parent run.updated: completed, checkpoint ordinal 1
06:26:45.315 follow-up submitted through RPC
06:26:47.778 seq=136 follow-up run.updated: completed, checkpoint ordinal 2
06:26:48.283 lifecycle equality and checkpoint Git objects verified
06:26:48.284 R3-complete; EXIT_CODE=0
```

Parent checkpoint `checkpoint:scope:checkpoint-scope%3Athread%3A9a76b386-e88f-44d2-bd48-35f547155a6b%3Aname%3Aroot:name:1` is `ready` and resolves to local Git commit object `1f9e00b6b0dc6a1b356ce58f2318fdde726eda10`. Follow-up checkpoint ends `:name:2`, is `ready`, and resolves to `2ea8ac92bf70d812856795054b7850b434d30e12`. These are synthetic local checkpoint objects, not forge commits. The driver verifies both with `git rev-parse <ref>^{commit}`. `verification.json` records the exact refs, capture times, lifecycle comparison and acceptance receipt.

## Artifacts and reproduction

[Full evidence archive](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e15d512098cb2fb8/full-evidence-logs.tar.gz) includes full test/setup logs; all three fresh runtime attempts, including fixture failures; sanitized server stdout/stderr; complete native Codex protocol logs and server traces; read-only event, receipt, outbox, run, child and checkpoint exports; final RPC projection; prompts; environment; matrix snapshot; scripts; and SHA-256 manifest. Primary successful runtime directory: `20261027T062557Z`. The earlier successful acceptance/follow-up attempt (`20261027T062330Z`) required resuming after a client-side follow-up validation error; it is supplemental, not the primary capture.

Scripts: [RPC/seeding driver](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d4999b03cdeb8fd6/rpc-driver.mjs), [FIFO gate](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1a066591275ed303/gate.py), [test recorder](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/26f45ae04a48a3f3/run-tests.sh), [server start/stop](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/41d5e9ab7108a13c/server-control.py), [read-only export/packaging](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a23bd33b536e7842/collect-evidence.py). Download the scripts together under `/tmp/lead-25666-r1/evidence`, start with `server-control.py start`, wait for the startup pairing line, then run the driver with the printed run directory from the head checkout. Every full retry uses a new timestamped home.

What else changed: **N/A for visual differences**; no images or diff sheets. Test-only overlay is confined to the disposable base checkout. Runtime state, seed repository and checkpoints are synthetic and isolated. The resource-monitor sidecar warning appears at startup but did not block the flow. No app source, PR body, commit or branch was changed.

Cleanup: all three spawned server process groups stopped by their recorded IDs; ports 14000/5960 confirmed clear. Disposable base checkout removed; lead checkout remains clean. Evidence scratch removed after upload. No reads or seeds from `~/.t3/userdata`. Public copies redact startup pairing URLs and matching environment secrets; cookies, pairing-command output, database files and secrets directories are excluded. Credential scan passed before upload.

FRICTION:
- tool | Initial RPC probe used an unavailable direct ws dependency and a constructor class where Effect expects a factory; corrected with the installed ws package and factory | A minimal typed RPC driver example would avoid this.
- other | First gate fixture included sentence punctuation as an extra argument; second attempt omitted required follow-up creation fields; final fresh run passed without recovery | Validate literal gate commands and RPC payload schemas before provider startup.
- runbook | Legacy orchestration_v2_events/receipts tables are empty; current persistence uses unified orchestration_events/orchestration_command_receipts | Name the current tables in the evidence handoff.

Model/harness: GPT-6.1-Sol in Codex. Evidence only; forge checks/reviews were not queried.