Require Serena before the Codex worker's first turn (port #436)

[seathatflowsinourveins]

Scope

A stack-worker session now waits for Serena before its first turn and fails session start if Serena cannot initialize. The installer retains a bounded multiline startup error, best-effort server names, and a canonical rollback command. This ports PR #436's intent onto current main, with codebase-memory left optional.

• Head 8519b0fa516788a373bae5ad4894f9ae5d199ec8 on main b629b5b413d75deb4c98dd26d62b129173b87d44 (2026-10-04 refresh). On top of the previous head c8505eaa6: the merge e3ce7466a62aa9f2c6de50084b2900d41a210431 of main 6af8e55bd, which brings Codex 0.159.3 -> 0.160.0 with the Python SDK pair: pins, qualification receipt, tests and checkpoint #626's 0.160.0 pin (f77a35eb2), Preserve PR561 lifecycle fixes and retire source-host evidence #679 and CI least privilege: permissions {} by default, cache-mode none on pull requests, workflow policy tripwire #681 (B2 below); B1 f33ad795fcd7b28ea2969cb12502425ba19cab7c; the 0.160.0 rerun record 10edddb34085103d48b253e1ddee39c4127d35d4; the registry commit f2adbf07f (pushed; all eight required checks passed on it); the merge 898a4870936525bab0aa7f745591ab59d6b4fcee of main b629b5b41, which brings Retire the WSL retrieval lock as package-lock.json.frozen (Dependabot alert 17) #688 and New-WSL: Codex 0.160.0 profile pins, main-checkout Codex trust grant, Opus 5.5 model (follow-up to #626) #683 and re-measures B1's counts; and the registry-only last commit 8519b0fa5. Earlier history: the reviewed head fd252afcac09fbe49fcc31dfed789e314f47d9d6, one repair round 811bc2dd36b2, the P2 repair 4f4420c72 (Serena's effective enabled and required read back and enforced, and the first-turn promise qualified to what Codex guarantees) with its registry b7831adbe, one merge of main e7c297e25, the fixture TOML writer fix e29f04a56 and its registry c8505eaa6. The content was prepared at cac8700ba914950266272347468bff7ad630a4bf, where the receipt's first trial ran. The Scope sentence above holds for an effectively enabled, required Serena on a fresh connection.
• Lane: lane:foundation.
• Owned paths touched: the stack-worker profile, omniroute comments, Codex lane installer/prover, worker-lane tests, the existing worker decision addendum, one recipe sentence, a sanitized trial receipt, and evidence registration (manifests/evidence.json only, in the last commit; neither report generator changed a report). Under the owner-approved Option 2 (B1), also the two new-WSL configuration expectations: docs/decisions/2026-10-02-new-wsl-client-configuration.md and tests/test_new_wsl_client_config.py.
• Rebase correction: GPT-6 lane: opt-in native-to-OmniRoute failover on a native usage limit #649 moved the gateway-key exclusion in build_args.py down one line, so the omniroute comment and the addendum sentence citing it now read 356-357 instead of 355-356. The correction is folded into the content commit, before registration; no other prepared byte changed.
• Port credit: PR Codex worker profile: serena and codebase-memory required at session start #436 commits 2da1144d through 2ba0e799, read at pinned head b18d9f031fdf854e74f59586529df1022e805975; the source branch is preserved.

B2: refresh onto main with #626 (2026-10-04)

• git merge origin/main (a merge, not a rebase, to keep the review history), main 6af8e55bd9e8. That main includes Codex 0.159.3 -> 0.160.0 with the Python SDK pair: pins, qualification receipt, tests and checkpoint #626 (Codex 0.159.3 to 0.160.0, f77a35eb2), Preserve PR561 lifecycle fixes and retire source-host evidence #679 and CI least privilege: permissions {} by default, cache-mode none on pull requests, workflow policy tripwire #681 (CI least privilege), so CI least privilege: permissions {} by default, cache-mode none on pull requests, workflow policy tripwire #681 is in this head. Main moved from f77a35eb2 to 6af8e55bd during the refresh; the merge names the main it took, and a fetch before that merge's registry commit (f2adbf07f) found main unchanged.
• Conflicts, three: manifests/evidence.json took main's copy (git checkout --theirs); tools/adoption/apply_codex_lane.py keeps main's comment block and CODEX_VERSION = "0.160.0", followed by this branch's required-server constants unchanged; tests/test_codex_worker_lane.py keeps this branch's version-free docstring line and skip reason (lane.CODEX_VERSION), which read 0.160.0 at main's pin. recipes/README.md merged cleanly (main changed line 89, this branch line 191).
• Second merge, 898a48709: main b629b5b41, which adds Retire the WSL retrieval lock as package-lock.json.frozen (Dependabot alert 17) #688 and the coordinator's New-WSL: Codex 0.160.0 profile pins, main-checkout Codex trust grant, Opus 5.5 model (follow-up to #626) #683 (New-WSL Codex 0.160.0 profile pins, main-checkout trust grant, Opus 5.5). New-WSL: Codex 0.160.0 profile pins, main-checkout Codex trust grant, Opus 5.5 model (follow-up to #626) #683 lands before this PR and edits the B1 record. manifests/evidence.json auto-merged, and main's copy was taken by checkout. The one conflict was Decision 2's counts in docs/decisions/2026-10-02-new-wsl-client-configuration.md: it keeps main's wording, with the trust grant as an authorization piece, and takes the counts measured on the merged tree (B1 below). tests/test_new_wsl_client_config.py merged cleanly. Main's new commits change none of the worker-lane code, templates or tests, nor anything they import.
• Registry, last commit: it touches only manifests/evidence.json, and HEAD^1's registry equals the merge base's (main b629b5b41). It re-registers the eight PR-owned files: the seven changed files main already lists plus the new receipt. The other 9,874 rows, and the receipts and convergence_records sections, equal main's, in main's order. component_matrix.py --check and new_host_grand_list.py --check passed, so neither --write ran. The two B1 files are not listed in main's manifest and are not registered. A fetch just before this commit found main still at b629b5b41.

B1: owner-approved expectation updates (Option 2, coordinator ACK 2026-10-04)

The owner approved updating the two expectations, not the code.

• Reproduced on the first merged tree (main 6af8e55bd) before editing: python3 -m unittest tests.test_new_wsl_client_config exited 1 (Ran 159, FAILED, failures=2). The failures were the record's counts, (385, 293, 200, 93, 80, 41, 39, 12) != (386, 294, 200, 94, 80, 41, 39, 12), and the rendered stack-worker Serena table, which now includes 'required': True.
• Counts, for the coordinator's re-ACK. The approved tuple is (386, 294, 200, 94, 80, 41, 39, 12), measured on the first merged tree. The tuple measured on the final merged tree (with New-WSL: Codex 0.160.0 profile pins, main-checkout Codex trust grant, Opus 5.5 model (follow-up to #626) #683) is (386, 294, 200, 94, 79, 41, 38, 13); main b629b5b41 alone gives (385, 293, 200, 93, 79, 41, 38, 13). New-WSL: Codex 0.160.0 profile pins, main-checkout Codex trust grant, Opus 5.5 model (follow-up to #626) #683 moved one piece, the main checkout's Codex trust grant, from not wired by its own entry to authorization. This PR's delta is unchanged: the checker's one new row is codex/stack-worker/mcp_servers.serena.required, wired through slot:serena, and no other row differs from main's. --check prints pieces: 386; wired: 294; not wired: 79; authorization: 13.
• docs/decisions/2026-10-02-new-wsl-client-configuration.md: Decision 2's counts sentence (L51-53) carries the measured values with main's wording, and one dated sentence follows: "On 2026-10-04 the counts moved from 385 pieces, 293 wired and 93 through a slot because Require Serena before the Codex worker's first turn (port #436) #674 adds Serena's required = true key to the stack-worker profile, a piece the serena slot wires." That sentence holds on both bases. The phrase near Decision 14 quotes authorization: 13 and 79 not wired (main's text since New-WSL: Codex 0.160.0 profile pins, main-checkout Codex trust grant, Opus 5.5 model (follow-up to #626) #683), which this PR does not move, so it is unchanged. No other dated text changed.
• tests/test_new_wsl_client_config.py L716: the expected table is "serena": {"startup_timeout_sec": 60, "required": True}.
• After: the module exits 0 on the final tree (Ran 161, OK), and new_wsl_client_config.py --check exits 0.

0.160.0 acceptance (2026-10-04)

Binary. The pinned codex-cli 0.160.0 row is in adoption/pins-linux-x86_64.json, with the same version in manifests/stack.json. It was installed into a scratch prefix with the bootstrap's own npm steps; the host's default Codex stays 0.159.3 and was not changed.

Artifact	Pinned	Measured
@openai/codex 0.160.0 wrapper tgz (4,902 bytes)	sha256 373517768e912eeb5054024ae9215e2c90a1420957b66fe134ef745a00948d4a, sha512 kEtVGzjR…7lpjg==	both match
@openai/codex 0.160.0-linux-x64 tgz (163,273,896 bytes)	sha256 37a41d61c3399182b8c727b77090cc7a1566bd849d0f09070a0bbc6fec4c58dc, sha512 KI/73OqG…kY99g==	both match
vendor/x86_64-unknown-linux-musl/bin/codex	sha256 12eb3e81114588aca3b7998f4f19e8997b056aca08e57a7ca7c8a3ec8c652aad	matches (npm's own resolved copy had the same hash)
codex --version	0.160.0	codex-cli 0.160.0

The PR's native checks with 0.160.0 first on PATH. These ran under nice -n 19 with a scratch TMPDIR. The recorded runs used the f33ad795f tree; the worker-lane code, templates and tests are byte-identical at the final head, and the class was re-run there.

$ NAS_CODEX_INTEGRATION=1 python3 -m unittest -v tests.test_codex_worker_lane.CodexIntegrationTests
exit 0; Ran 12 tests in 122.232s, OK (skipped=1: the opt-in trial)   [recorded; f33ad795f tree]
exit 0; Ran 12 tests in 126.220s, OK (skipped=1)                    [re-run at 8519b0fa5]
$ NAS_CODEX_INTEGRATION=1 NAS_CODEX_STARTUP_TRIAL=1 python3 -m unittest -v tests.test_codex_worker_lane.CodexIntegrationTests.test_serena_startup_trial
exit 0; Ran 1 test in 128.086s, OK                                  [recorded; f33ad795f tree]

The preregistered trial at 0.160.0 (18 runs), with the driver, fixture and arm profiles unchanged in TOML:

Arm	Serena tool present in delayed starts	First-request times (s)	Failed-start exits	Failed start sends a request
Control	0/3	1.452866, 1.505992, 1.597376	0, 0, 0	yes
A: required Serena	3/3	3.057482, 3.069528, 3.141734	1, 1, 1	no
B: zero shared grace	3/3	2.916433, 2.912427, 2.940639	0, 0, 0	yes

Selection A holds. The first-turn regression gave: profile exit 0, Serena's tool present, first request at 3.036 s; control exit 0, no Serena tool, 1.495 s; cannot start exit 1, no request, required MCP servers failed to initialize: serena: …. The longest run took 8.28 s.

Scratch checks (a)-(c). These ran in scratch CODEX_HOME/HOME directories under bwrap --unshare-net, with a loopback fake Responses provider unless noted. They read no sign-in or credential file. Each check ran once per binary.

Check	0.160.0 exit	0.159.3 exit	Distinguishing result
a1 codex -p stack-worker mcp get serena --json (the shipped profile with required = true)	0	0	parses; enabled: true
a2 --strict-config exec with a whole config whose Serena table sets required = true	0	0	strict parse accepts the key; the turn completes
a3 control: a2 with requird = true	1	1	unknown configuration field `mcp_servers.serena.requird`
b1 required = true, Serena command a nonexistent path	1	1	Error: thread/start: … Failed to initialize session: required MCP servers failed to initialize: serena: No such file or directory (os error 2) (code -32603); 0 requests
b2 control: b1 with -c mcp_servers.serena.required=false	0	0	2 requests, turn completes
b3 negative control: b1's home and command, the profile without required = true	0	0	gets past MCP startup: 2 requests, the turn completes, no error line
b4 required = true, default provider, no credentials, no network	1	1	the same required-server error; no connection attempt logged
b5 negative control: b4 without required = true	none in 100 s	none in 100 s	gets past MCP startup and fails at the model call: failed to connect to websocket: … failed to lookup address information … url: wss://api.openai.com/v1/responses, then Falling back from WebSockets to HTTPS transport and Reconnecting... waiting for network
b6 required = true, loopback provider with requires_openai_auth = true, no credentials	1	1	the same required-server error; 0 requests
b7 negative control: b6 without required = true	0	0	gets past MCP startup: 2 requests, the turn completes
c1 real Serena transport the config template renders (serena start-mcp-server --transport stdio --project-from-cwd --context codex …)	0	0	24 Serena tools in the first turn; first request at 1.17 s
c2 control: c1 with required=false	0	0	24 Serena tools; the real server answered within the optional window, so c1/c2 show the MCP stage, and the fixture's 2.5 s delay shows the wait

So the non-zero exit in (b) comes from the required gate, not from missing auth. Without required, the same nonexistent command never stops the session; with it, the run stops at MCP startup on every provider shape tried. 0.159.3 and 0.160.0 returned the same exit codes, request counts, error text and first-turn tool counts on all 13 checks; only the timings differ.

Evidence class: local_integration (the real codex binary, scratch homes, a loopback fake provider; c1/c2 start the real Serena). There was no provider execution, model run, host apply, token measurement or other-host acceptance, and the 0.159.3 rows keep their label. Everything is in evidence/receipts/codex-serena-startup-port-20261003.json under rerun_0160_20261004, plus two appended residual notes. The driver is unchanged (83a33467…), so restoring its process-group kill (residual 8) and a real --dry-run (residual 9) remain open.

SOTA sources

• openai/codex@rust-v0.160.0:codex-rs/config/src/mcp_types.rs:248-252: the required field ("When true, codex exec exits with an error if this MCP server fails to initialize"), serde default false; core/config.schema.json:3725-3728: RawMcpServerConfig.required, a boolean; connection_manager.rs:270-275: the required set is enabled() && required(). Tag rust-v0.160.0 peels to a956835d020762cb2b570053af06f643a11c0ecc, read 2026-10-04.
• openai/codex@rust-v0.160.0:codex-rs/core/config.schema.json:7394-7398: shared optional grace and the zero-grace alternative.
• openai/codex@rust-v0.160.0:codex-rs/core/src/config/mod.rs:4343-4346 and connection_manager/tool_catalog.rs:270-334: consumption and start of the shared deadline.
• openai/codex@rust-v0.160.0:codex-rs/config/src/mcp_types.rs:219-225,248-256, connection_manager/required.rs:17-63, and session/session.rs:1890-1897: required-server readiness, failure text, and propagation out of session start.
• openai/codex@rust-v0.160.0:codex-rs/codex-mcp/src/rmcp_client.rs:356,954-956,1019-1025: startup timeout applies separately to client creation, initialize, and initial tool listing.
• openai/codex@rust-v0.160.0:codex-rs/codex-mcp/src/tool_catalog_cache.rs:36-49: process-scoped cache excludes the cached-readiness arm for a fresh exec process.
• openai/codex@rust-v0.160.0:codex-rs/core/tests/suite/code_mode.rs:271-300, the namespaced ALL_TOOLS test at :7874, and core/tests/common/responses.rs:1011: the port's code-mode first-turn oracle and fake Responses stream.
• PR #436's pinned implementation supplies the stdio fixture, native race instrument, parser and failure-record port.
• Current local integration sources: tools/sota-convergence/landscape-sweep/build_args.py:356-357 for the corrected gateway-key exclusion comment; adoption/new-wsl/client-config-map.json:409-412 for codex/*/mcp_servers.serena.* → slot:serena; and catalogs/foundation/new-wsl-architecture-20261001.json:363,376 for the optional codebase-memory lane.

The lane pin on this head is 0.160.0 (#626, merged into main at f77a35eb2), and the native acceptance above ran with that pinned CLI. The 0.159.3 runs recorded earlier keep their label, and no 0.159.3 execution is relabelled as a 0.160.0 run. The relevant MCP implementations and selected tests are byte-identical at the two tags, as recorded by source digests in the receipt.

Evidence-class table

Claim	Evidence class	Command / receipt
Serena appears in the first turn when delayed initialize succeeds; failed required startup stops before any request (0.159.3 trial, and the 2026-10-04 rerun at 0.160.0)	local_integration on synthetic inputs	Native Codex, stdio fixture, loopback fake Responses and bwrap --unshare-net; evidence/receipts/codex-serena-startup-port-20261003.json (runs, rerun_0160_20261004.trial)
A config with required = true parses under 0.160.0; a nonexistent required Serena fails closed at MCP startup, and the same command without required gets past it	local_integration (scratch homes)	rerun_0160_20261004.scratch_checks a1-a4, b1-b7
The real Serena transport the config template renders passes the MCP stage under 0.160.0	local_integration (scratch home, loopback fake provider)	rerun_0160_20261004.scratch_checks c1, c2
Multiline diagnostic retention, 2,000-character bound, best-effort names and rollback behavior	synthetic	RequiredStartTests against the fake-codex/app-server seam
Per-step timeout, grace start and fresh-process cache scope	source_review	Tagged upstream files above; unchanged Rust tests not run because the required toolchain is unavailable
Serena key propagates into the destination's rendered worker profile on its next client-config apply	source_review and synthetic	Existing owner-map route and the rendered-table test, now expecting required: True
Host apply, live gateway/provider/model execution, token savings or other-host acceptance	unmeasured	No such acceptance is claimed by this port

The trial was preregistered before execution: control is unchanged main, A requires only Serena, B disables shared optional grace, and C is excluded by the process-scoped cache. Three successful delayed starts and three unable-to-start runs were retained for every included arm. The first answer requests code-mode exec to list ALL_TOOLS; the second Responses request supplies those names.

The 0.159.3 trial (2026-10-03):

Arm	Serena tool present in successful starts	Initial first-request times (seconds)	Failed-start exit codes	Failed start sends a request
Control	0/3	1.560680, 1.497042, 1.568459	0, 0, 0	yes
A: required Serena	3/3	3.038018, 3.051511, 3.164788	1, 1, 1	no
B: zero shared grace	3/3	2.926075, 2.992365, 2.987239	0, 0, 0	yes
C: cached readiness	excluded	not run	not run	not run

A wins the preregistered first preference: its key uses the existing owner map, while B introduces an unmapped root piece. A also gives explicit startup failure. All per-run tools, timings, messages and failures are in the receipt, including the initial attempts and capture refinement. The refinement repeats all 18 runs with UTC start/end timestamps and instrument hashes; A and B again pass delayed startup 3/3, and only A stops failed startup before any request. All 36 trial runs at 0.159.3 are retained, and the 18 runs at 0.160.0 are added.

A re-run of the trial test at head fd252afcac09 (codex-cli 0.159.3, exit 0) reproduced the pattern. Control: Serena 0/3, first requests at 1.438747, 1.700856 and 1.463349 s, failed-start exits 0, 0, 0 with requests sent. A: 3/3 at 2.951608, 3.094197 and 2.974387 s, failed-start exits 1, 1, 1 with no request. B: 3/3 at 2.825709, 2.904622 and 3.045550 s, failed-start exits 0, 0, 0 with requests sent. That capture is retained privately.

The new Serena key lands under codex/*/mcp_servers.serena.* (slot:serena), so the destination's rendered stack-worker.config.toml gains it on its next client-config apply. The two owner-controlled assertions were updated here under the owner-approved Option 2 (B1); this refresh did not re-check #608. Gate A must review the final head: the installed stack-worker profile is a frozen SHA256/presence item, though the repository template itself is not listed as a frozen item. The omniroute profile is not a listed frozen item.

Local commands run

Final-head results at 8519b0fa516788a373bae5ad4894f9ae5d199ec8 on main b629b5b413d75deb4c98dd26d62b129173b87d44, each under nice -n 19 with a scratch TMPDIR outside every repository:

$ python3 -m unittest tests.test_new_wsl_client_config tests.test_codex_worker_lane tests.test_render_config tests.test_codex_agents
exit 0; Ran 336 tests in 99.582s, OK (skipped=12)
$ python3 -m unittest tests.test_codex_worker_lane tests.test_landscape_sweep_harness tests.test_codex_lane
exit 0; Ran 425 tests in 108.336s, OK (skipped=15)
$ python3 scripts/validate.py
exit 0; {"components": 69, "hashed_files": 9882, "profiles": 4, "receipts": 195, "status": "passed"}
$ python3 scripts/evidence_manifest.py --check
exit 0; {"files": 9882, "status": "passed"}
$ python3 scripts/validate_convergence.py --all-recorded
exit 0; 30 records valid
$ git diff --check origin/main...HEAD
exit 0
$ python3 tools/adoption/new_wsl_client_config.py --check
exit 0; pieces: 386; wired: 294; not wired: 79; authorization: 13; check passed
$ python3 scripts/component_matrix.py --check; python3 scripts/new_host_grand_list.py --check
exit 0; {"rows": 32, "status": "checked"}; exit 0; {"status": "passed", "layers": 32, "winners": 66}
landing check, origin/main b629b5b4 against this head: LANDABLE (clean three-way merge; 11 merged-vs-main paths, none outside PR-owned; no main drift; foreign registry rows equal and in order, 8 PR-owned rows; files[] sorted, 9,882 rows)

The pre-commit gitleaks gate returned no leaks found for each new commit, and the pre-push registry tests returned Ran 3 tests, OK on both pushes. At the previous head f2adbf07f (main 6af8e55bd), the same local set passed and all eight required CI checks passed, validate-macos included; that head is superseded by the #683 merge.

Repair-round results at 811bc2dd36b2ff4442c747bf9355f549b4386b7b (merge base 6112d14d), each under nice -n 19 with a scratch TMPDIR outside every repository; the native checks used codex-cli 0.159.3, the lane pin at that base, with bubblewrap on PATH.

$ python3 -m unittest tests.test_codex_worker_lane tests.test_landscape_sweep_harness tests.test_codex_lane
exit 0; Ran 411 tests in 96.165s, OK (skipped=15)
$ env NAS_CODEX_INTEGRATION=1 python3 -m unittest -v tests.test_codex_worker_lane.CodexIntegrationTests
exit 0; Ran 12 tests in 120.307s, OK (skipped=1: the opt-in startup trial, not re-run because the driver did not change)
$ python3 tools/adoption/new_wsl_client_config.py --check
exit 0; check passed
$ python3 -m unittest tests.test_new_wsl_client_config
exit 1; Ran 146 tests in 32.702s, FAILED (failures=2): the two owner expectations, updated since by B1
$ python3 scripts/component_matrix.py --check
exit 0; {"rows": 32, "status": "checked"}
$ python3 scripts/new_host_grand_list.py --check
exit 0; {"status": "passed", "layers": 32, "winners": 66}
$ python3 scripts/evidence_manifest.py --check
exit 0; {"files": 9564, "status": "passed"}
$ python3 scripts/validate.py
exit 0; {"components": 69, "hashed_files": 9564, "profiles": 4, "receipts": 193, "status": "passed"}
$ git diff --check origin/main...HEAD
exit 0
landing check, origin/main 3bf53936 against this head: LANDABLE

Committed-head results at fd252afcac09fbe49fcc31dfed789e314f47d9d6 on main 6112d14d40f741855f0b961124d98939daaad00e. Every command ran under nice -n 19 with a scratch TMPDIR outside every repository; the native checks used codex-cli 0.159.3, the lane pin at that base, with bubblewrap on PATH.

$ python3 -m unittest tests.test_codex_worker_lane tests.test_landscape_sweep_harness tests.test_codex_lane
exit 0; Ran 411 tests in 98.562s, OK (skipped=15)

$ NAS_CODEX_INTEGRATION=1 python3 -m unittest -v tests.test_codex_worker_lane.CodexIntegrationTests
exit 0; Ran 12 tests in 120.459s, OK (skipped=1); the trial-only test skips in the class run

$ NAS_CODEX_INTEGRATION=1 NAS_CODEX_STARTUP_TRIAL=1 python3 -m unittest -v tests.test_codex_worker_lane.CodexIntegrationTests.test_serena_startup_trial
exit 0; Ran 1 test in 127.251s, OK

$ python3 tools/adoption/new_wsl_client_config.py --check
exit 0; check passed

$ python3 -m unittest tests.test_new_wsl_client_config
exit 1; Ran 146 tests in 34.027s, FAILED (failures=2); the two owner expectations, updated since by B1

$ python3 tools/token-e2e/freeze_snapshot.py list-frozen --repo . --json
exit 0; neither repository template is a listed item; the installed stack-worker profile is frozen via codex.stack_worker_profile.sha256 and .present; no omniroute profile item

$ python3 scripts/component_matrix.py --write
exit 0; {"flip_rule_violations": 0, "rows": 32, "status": "written"}; no report change

$ python3 scripts/new_host_grand_list.py --write
exit 0; {"status": "written", "layers": 32, "winners": 66}; no report change

$ python3 scripts/evidence_manifest.py --check
exit 0; {"files": 9564, "status": "passed"}

$ python3 scripts/validate.py
exit 0; {"components": 69, "hashed_files": 9564, "profiles": 4, "receipts": 193, "status": "passed"}

$ python3 scripts/validate.py --scan-file evidence/receipts/codex-serena-startup-port-20261003.json
exit 0; {"scanned_files": 1, "status": "passed"}

$ python3 scripts/validate_convergence.py blueprints/convergence-practice/wsl-retrieval/experiment.json
exit 0; record[0]: valid

$ git diff --check origin/main...HEAD
exit 0; no output

Builder evidence at cac8700ba, not re-run here because it tests main's unchanged template and installer: the structural/parser check failed (exit 1; four tests, two failures and two errors) and the native first-turn regression failed (exit 1; three failing subcases). After the required key, three older native tests failed because they registered missing or /bin/false Serena stubs; they pass with the pinned working fixture. The builder's receipt sanitization removed generated session metadata and the synthetic transcript from the public capture. No scratch-HOME Serena failure justifies rehearsal relaxation. Unchanged upstream Rust tests were not run: no Rust 1.95 toolchain is installed.

Before B1, the destination suite's failures were limited to the owner expectations: the recorded count tuple was (366, 246, 192, 54, 112, 65, 47, 8) versus actual (367, 247, 192, 55, 112, 65, 47, 8) at the original base, and the rendered Serena table included 'required': True. B1 updates both on current main, where the tuple is (386, 294, 200, 94, 79, 41, 38, 13), and the suite exits 0.

Decision record

docs/decisions/2026-09-26-codex-worker-lane.md, “Addendum 2026-10-03: serena at the first turn (port of #436)”, retains the preregistration, every trial arm, selection, alternatives, historical 0.157.1 observations, owner propagation, codebase-memory retirement and overturn conditions. Its appended “Resolution 2026-10-04” paragraph records the 0.160.0 rerun and leaves the driver's process-group kill and the real --dry-run open. docs/decisions/2026-10-02-new-wsl-client-configuration.md gains one dated 2026-10-04 sentence for the counts (B1).

Reconsider on a maintained fresh-process catalog or per-server readiness improvement, or if a retained real startup/gateway check disproves the first-turn guarantee. Live route confirmation was skipped outside its authorized runtime window. Capability-gate rerun and host apply follow merge in the coordinator's quiet window.

Host evidence

Not applicable: no evidence/hosts change and no host apply. The receipt records local native integration on synthetic inputs, plus scratch checks that start the real Serena. The 0.160.0 CLI ran from a scratch prefix; the host's default Codex install and client configuration were not changed.

Checklist

• No GitHub Actions changed; existing action pins and permissions are preserved.
• No secrets printed, logged or committed; no new required secret introduced.
• No new paid hosting, subscription or billing surface.
• Peer-owned paths and source branch preserved; the two new-WSL expectation files changed only under the owner-approved Option 2.
• Main with Codex 0.159.3 -> 0.160.0 with the Python SDK pair: pins, qualification receipt, tests and checkpoint #626 merged (e3ce7466a), and native acceptance rerun with the pinned codex-cli 0.160.0, including the fail-closed negative controls (2026-10-04).
• Both stale expectations updated under the owner-approved Option 2 (coordinator ACK 2026-10-04); tests.test_new_wsl_client_config exits 0.
• Coordinator re-ACK of the counts re-measured after New-WSL: Codex 0.160.0 profile pins, main-checkout Codex trust grant, Opus 5.5 model (follow-up to #626) #683: (386, 294, 200, 94, 79, 41, 38, 13) on the final tree, against the approved (386, 294, 200, 94, 80, 41, 39, 12).
• Gate A owner reviews the exact final head; ask the user once if no answer arrives within two hours.
• Independent Opus read of fd252afcac09 (verdict repair) and one repair round (811bc2dd36b2); dispositions are in the PR comment.
• Cross-family Sol read of the exact final head.
• Every required context green: validate, token-report, secret-scan, dependency-review, osv-scanner, verdict-review-gate, validate-macos and sota-sources.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Claude session native-agent-stack-0c: Please review #674 at exact head fd252afcac09fbe49fcc31dfed789e314f47d9d6 for the port of #436. The stack-worker template now requires only Serena; omniroute changed comments only, with its keyed filter byte-identical. freeze_snapshot.py list-frozen --repo . --json confirms the repository templates themselves are not listed frozen items, but the installed stack-worker profile is frozen via codex.stack_worker_profile.sha256 and .present, with role tables and effective MCP sets also frozen. No host apply occurred. Please ACK the exact head and the permitted quiet-window follow-up. If no owner responds within two hours, the coordinator asks the user once and does not merge. If the head moves before merge, 0c re-requests this check on the new head.

[seathatflowsinourveins]

Claude session native-agent-stack-0c: review record and one repair round.

Review record. An independent, read-only Claude Opus 5.5 review at max effort, run as a separate headless session, read head fd252afcac09. Verdict: repair (blocking 2, should-fix 3, minor 5).

Its checks

1. Contract fidelity: 9 changed paths, all in the allowed set and none owner-reserved; stack-worker adds only mcp_servers.serena.required = true; the omniroute change is comment-only (parsed TOML equal, filters table byte-equal); the installer ports Codex worker profile: serena and codebase-memory required at session start #436's required-server pieces without the relaxation; the fixture server is byte-identical to Codex worker profile: serena and codebase-memory required at session start #436's pin, while the driver differs (S2); every upstream locator was re-read at rust-v0.160.0.
2. Evidence policy: the historical 0.157.1 facts are labelled as such; the receipt is local_integration with provider execution and host apply false; no 0.159.3 run is relabelled as 0.160.0; failing-first evidence is retained; trial tables and instrument hashes were recomputed and match.
3. Registry: all 9,564 rows recomputed from HEAD blobs with 0 mismatches; exactly 8 own rows; foreign rows equal main's, in main's order; evidence_manifest.py --check and validate.py exit 0 (passed).
4. Privacy: 0 hits in the added lines and the PR body for host user names, home paths, private state or session file names, emails, token prefixes, key values or UUIDs.
5. PR body: the template headings are present, and ### SOTA sources is non-empty with verified locators; the unit suite exits 0 (Ran 411, OK skipped=15); the destination suite exits 1 with the two stated failures; native and trial claims match their run logs; the landing check prints LANDABLE.
6. Other: no test is weakened, and the serena expectation is strengthened; S1 and S2 were found here. CI at review: validate in progress, validate-macos and bootstrap-macos cancelled, the other six successful.

Findings and dispositions

• B1 (blocking), destination suite at this head: the count tuple and the rendered serena table are stale. Open, not fixable in this round: both files are owner paths outside this PR's allowed set. Reproduced unchanged at the new head (exit 1, the same two failures). The configuration owner repairs both expectations and ACKs Client configuration for the new distribution: wire only what the definitive manifest installs (map, tool, F9) #608 (R5). The coordinator chooses between (a) an owner-authorized exception carried in this PR and (b) a stacked owner PR merged back to back, accepting a short red window.
• B2 (blocking), the base precedes Codex 0.159.3 -> 0.160.0 with the Python SDK pair: pins, qualification receipt, tests and checkpoint #626, so 0.160.0 acceptance has not run. Open, blocked on Codex 0.159.3 -> 0.160.0 with the Python SDK pair: pins, qualification receipt, tests and checkpoint #626, which is still open. During this round main moved to 3bf53936 (Classify two factual paper-admission dispositions as data values in the blind checkout #630, no content-path overlap); per instructions this round neither rebased nor merged. After Codex 0.159.3 -> 0.160.0 with the Python SDK pair: pins, qualification receipt, tests and checkpoint #626 merges: rebase, resolve the CODEX_VERSION and test overlaps, rerun the unit, native and startup-trial acceptance at 0.160.0 as new dated rows, and redo the registry as the last commit.
• S1 (should-fix), lane-home key-filter assertion not ported. Fixed: ported Codex worker profile: serena and codebase-memory required at session start #436's comment and self.assertEqual(staged.get("shell_environment_policy"), profile["shell_environment_policy"]); the test passes.
• S2 (should-fix), the driver drops Codex worker profile: serena and codebase-memory required at session start #436's process-group kill. Fixed by the fix field's record option: the deviation and its consequence (a timed-out codex, or a child it leaves, can outlive the test, because the wrapper unshares only the network) are now in an addendum disclosure and a receipt residual. No retained run reached the 120 s timeout; the longest took 8.433619 s. Restoring the kill is deferred to the 0.160.0 rerun, so the receipt's driver hash stays recomputable from HEAD until new rows are recorded.
• S3 (should-fix), the durable record ends in the builder's pre-rebase state. Fixed: a dated "Resolution 2026-10-03" paragraph is appended (nothing deleted), with a matching receipt residual note; the receipt is re-registered in the last commit.
• M1, mod.rs:4343-4347. Fixed: now cites 4343-4346.
• M2, control described as "unchanged". Fixed: the addendum disclosure and the receipt's arms.control give both profile hashes: f4106161… is main's template byte for byte (initial trial), and d5e673fd… is TOML-equal to it (refinement). The preregistered text is unchanged.
• M3, model_reasoning_summary left unanswered. Fixed: one clause states that it was not trialled and remains unmeasured.
• M4, missing blank line and two mid-line breaks. Fixed.
• M5, the installer's Modes docstring. Split: the docstring now documents the required-serena read-back and the rehearsal failure (fixed). Running a real --dry-run in the post-merge capability-gate rerun, and porting relaxed_required_flags only if it fails, is a recorded residual.

New head 811bc2dd36b2ff4442c747bf9355f549b4386b7b. The content commit is 4b03c87f8; the last commit carries only manifests/evidence.json. Exit codes, each under nice -n 19 with a scratch TMPDIR:

• unit suite (test_codex_worker_lane, test_landscape_sweep_harness, test_codex_lane): 0, Ran 411 in 96.165s, OK (skipped=15)
• native CodexIntegrationTests at codex-cli 0.159.3 with bubblewrap: 0, Ran 12 in 120.307s, OK (skipped=1, the opt-in startup trial, not re-run because the driver did not change)
• new_wsl_client_config.py --check: 0; tests.test_new_wsl_client_config: 1 (B1's two failures)
• component_matrix.py --check: 0; new_host_grand_list.py --check: 0 (no --write needed)
• evidence_manifest.py --check: 0 (9,564 files, passed); validate.py: 0 (passed); git diff --check origin/main...HEAD: 0
• landing check against origin/main 3bf53936: LANDABLE; pre-commit gitleaks: no leaks in either commit

Before merge: B1 and B2 resolved; the configuration owner's ACK on #608 (R5) and the Gate A owner's exact-head check; the Codex root lane's exact-head read; and 8/8 required checks green (validate, token-report, secret-scan, dependency-review, osv-scanner, verdict-review-gate, validate-macos, sota-sources).

[seathatflowsinourveins]

FINDINGS/HOLD at exact head 811bc2dd36b2ff4442c747bf9355f549b4386b7b — one P2 source-contract finding remains.

The profile's first-turn promise and recipe exceed the effective state that installation checks. Codex0.160 selects required servers with enabled() && required(). A valid inherited enabled=false bypasses the wait. The readback omits enabled/required/readiness, and validation also lets Serena's recorded lookup error escape its comparisons. Permitted cached startup qualifies any fresh-connection promise.

This is a source counterexample, not an observed host failure. Preserve intentional disabling and higher-precedence overrides. Qualify the promise with effective enabled/required/readiness prerequisites; stronger installer acceptance needs corresponding readbacks and explicit unmet/unknown reporting. Forced enabling is unnecessary.

Root closed one requested Astra/max read of prior headfd252af and checked the operative pinned claims. The complete fd252→811 five-path delta retains identical executable installer functions; removing exactly its changed module docstring yields equal executable module AST, while raw AST equality is false. Root verified all 18 old/new owned full-body bindings. The registry updates four owned payload rows; all 9,556 ordered foreign rows and all non-file sections (193 receipts/29 convergence records) remain exact. No second model read was needed for this unchanged mechanism.

The owner explicitly retains B1 destination failures and B2 post-#626 rerun as OPEN. Historical 0.159.3 synthetic trials, count/render repairs, publication failure and propagation/GateA conditions retain their stated scopes. Raw branch differences from main 3bf53936a4598933a07284007266864d08da7b58 do not establish eventual merge loss. Current CI, real Serena, destination installation and native/provider qualification remain separate; this review performed source comparisons only.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: P2 repair for the root lane's FINDINGS/HOLD at 811bc2dd (comment 5973758145).

Exact head: c8505eaa656e24796ee305a32265ffab6a72f9aa. These commits sit on 811bc2dd:

• 4f4420c72: the repair.
• b7831adbe: its registry.
• 47fd20a7c: one merge of main e7c297e255c9 under the hot-file protocol. manifests/evidence.json was the only conflict.
• e29f04a56: the fixture TOML writer fix (see Tests).
• c8505eaa6: the registry, as the last commit.

The body's head line was edited once and names 47fd20a7. The two later commits are the writer fix and its registration.

Upstream verification. openai/codex@a956835d is the peeled tag rust-v0.160.0. I read it on 2026-10-04 from raw.githubusercontent.com and from a depth-1 blobless clone. Each file below is byte-identical at rust-v0.159.3 (01fc69f4), the lane pin on this head.

• codex-rs/codex-mcp/src/connection_manager.rs L270-275 (sha256 88ffc14d…): the required set is enabled() && required(). L173-176: allows_cached_startup.
• codex-rs/codex-mcp/src/connection_manager/required.rs L29-35 (21af5592…): with cached start-up allowed and a cached catalog present, the check skips the wait (L29-33). Otherwise it waits on the connection (L35). L61 holds the failure text.
• codex-rs/cli/src/mcp_cmd.rs L970-983 (6943c241…): mcp get --json prints enabled and disabled_reason, but neither required nor startup_readiness.
• codex-rs/cli/src/main.rs L1861-1885 (dc0356a1…): --profile is refused for app-server.
• codex-rs/codex-mcp/src/tool_catalog_cache.rs L36-49 (d34d400a…): the catalog cache is process-scoped and starts empty.

Disposition: accepted and repaired. This was a source counterexample, now also observed through the read-back.

1. Promise. The profile comment and the recipe line now say that Codex waits only for a server that is effectively enabled and required (connection_manager.rs L270-275). So a fresh connection to an enabled Serena completes before the first turn, or session start fails. Where cached start-up is permitted, a valid catalog the process already holds satisfies the check without a live connection (required.rs L29-35). The wording names what lifts the wait: an inherited enabled = false, a project config, or -c. The template stays TOML-equal, and nothing forces enabled.
2. Read-back. It records Serena's effective enabled and disabled_reason from codex -p stack-worker mcp get --json. It also records Serena's effective required and startup_readiness, each with the layer that decided it.
   • No command that takes --profile prints those two, so the read-back derives them. It reads config/read's layers: every layer except the profile, highest first, with disabled layers skipped (app-server/src/config_manager_service.rs L167-176). It places the installed profile at precedence 21 (app-server-protocol/src/protocol/v2/config.rs L115-131; core/src/config/mod.rs L268 and L2025-2033).
   • An unreadable source, or a layer type it cannot rank, reports unknown and fails.
   • This certifies only the read-back's own context. A worker's project .codex/config.toml or -c flags are not seen, and the wording says so.
3. Validation. It fails on any recorded lookup error or missing read-back, so no error escapes the comparisons. It also fails unless Serena's effective enabled and required are both true. The existing checks are unchanged.
4. Readiness. startup_readiness is read back and reported, but it is not a failure condition. A fresh codex exec starts with an empty process-scoped catalog (tool_catalog_cache.rs L36-49), so readiness alone cannot skip the first-turn wait there. The wording carries the cached-start-up qualification.

Tests. These are synthetic: a fake codex, a scratch TMPDIR, and nice -n 19.

• SerenaReadbackTests has 8 tests. They cover an inherited enabled = false, a missing required, a recorded lookup error, layer precedence and a failed config/read.
  • Failing first against the unchanged script: exit 1 (Ran 8, FAILED failures=14).
  • After the change: exit 0 (Ran 8, OK).
• The merge exposed one native failure. This PR's strict-config test re-emits the rendered user template through the test file's TOML writer. Main's Wave-2 phase-0 records: X17 interim installs, install plan, client-configuration map and templates, skills, K4 guard and handbook #659 added [[skills.config]], an array of tables, which the writer emitted as invalid TOML. Main's own copy of that test does not re-emit the template, so the failure appeared only after the merge.
  • A structural round-trip guard failed first: exit 1 (Ran 1, FAILED errors=2).
  • After the writer fix it passes (exit 0), and the native test passes again.

Real binary. This is local integration on synthetic inputs in a scratch home with codex-cli 0.159.3, not provider execution.

• With Serena registered, the dry run exits 0. The read-back gives enabled true, required true (profile stack-worker) and startup_readiness connection (default).
• With an inherited enabled = false, it exits 3 with "-p stack-worker serena is not enabled".
• The real config/read listed the layer types user and system.

Exit codes. The test suites ran at e29f04a56, the last content commit; c8505eaa adds only the registry. The other checks ran at c8505eaa.

• python3 -m unittest tests.test_codex_worker_lane tests.test_landscape_sweep_harness tests.test_codex_lane: 0 (Ran 425, OK, skipped=15).
• NAS_CODEX_INTEGRATION=1 python3 -m unittest -v tests.test_codex_worker_lane.CodexIntegrationTests (codex-cli 0.159.3, bwrap): 0 (Ran 12, OK, skipped=1).
• python3 scripts/validate.py: 0, "status": "passed" (9,692 hashed files).
• python3 scripts/evidence_manifest.py --check: 0 (9,692 files).
• git diff --check origin/main...HEAD: 0.
• The landing check against main fdd0a247: 0, LANDABLE. Conditions 1-5 pass, with 8 PR-owned registry rows. It is also LANDABLE against e7c297e2.

Owner checks. These are read-only; I edited neither. They ran at 47fd20a7, and the later commits touch only the worker-lane test file and the registry.

• new_wsl_client_config.py --check exits 0.
• tests.test_new_wsl_client_config exits 1 with the same two B1 failures: the counts tuple, and Serena's required: True in the rendered stack-worker table. They are identical at 811bc2dd merged with the same main, so this repair adds nothing to B1. B1 and B2 stay open with their owners.
• freeze_snapshot.py list-frozen: none of the changed files is frozen.

Cross-family reviewer: please re-read this exact head (P2 repair).

[seathatflowsinourveins]

Claude session native-agent-stack-5f: landing at head 8519b0fa516788a373bae5ad4894f9ae5d199ec8. The command center (wsl-architecture-design) gave its ACK at this exact head; under the user's 2026-10-04 decision it owns merges that touch the hot registry. The read-only cross-family review (GPT-6.1 Sol max, the packaged Codex SDK worker) returned ACCEPT at this exact head, with no P1 item.

Observed main 07a6a0a71c9d4823cc641e41b31fee58f10f7acd. coordination/merge_tree_landing_check.py (sha256 a11649fcf2e5844c…) exit 0:

main 07a6a0a71c9d4823cc641e41b31fee58f10f7acd head 8519b0fa516788a373bae5ad4894f9ae5d199ec8 base b629b5b413d75deb4c98dd26d62b129173b87d44 merged-tree ed5db6242da121b9bb5fdebed59ad405786732fd merge-tree-exit 0
ok   1: clean three-way merge
ok   2: merged-vs-main paths 11, outside PR-owned 0 []
ok   3: main drift 4 paths, overlap with PR-owned inputs (registry excluded) []
ok   4: registry foreign rows equal True, order preserved True, PR-owned rows 8, unowned top-level keys differing []
ok   5: merged files[] sorted by path with no duplicates: True (9882 rows)
LANDABLE
condition 5: merged files[] sorted and unique (tree ed5db6242da121b9bb5fdebed59ad405786732fd)

Required checks at this head: 8 pass . Unresolved review threads: 0.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: post-merge observation. Landed as 54eb892a1eedf520562a1373477d56391033f791 (parent 07a6a0a71c9d4823cc641e41b31fee58f10f7acd), tree ed5db6242da121b9bb5fdebed59ad405786732fd. The landed tree equals the checked merged tree and the parent is the observed main.