Engineering archive

These notes describe work at the date shown. See the usage reference for current behavior.

Phase 5 validation record

Phase 5 adds agent tasks: submit a prompt, run a headless coding agent on a Mac mini, and collect the result as a Git branch. This record is the sanitized template for the three-worker live acceptance in spec section 20.2 items 1 to 5 and 7 to 10. Item 6 (source = origin and publish = push), the Cursor and OpenCode adapter rows, and retention through worker gc are recorded in the Phase 5d section below; this run records their sanitized outcomes.

This record covers the fresh v7 live acceptance against the resumed-turn supervisor revision. Attempts 1 to 5, including the superseded v4, v5, and v6 roots, are not evidence for this run; their tasks were closed or abandoned before the corrected v7 run. The v7 setup gate reached three ready workers with authenticated Codex 0.153.2. No profile was created, uploaded, printed, or used.

Every live row below identifies either observed sanitized evidence or what was not proven. A blocked row is not a passing substitute for the corresponding runbook item.

The record may retain only shortened identifiers, sanitized command categories, terminal states, exit results, durations, before/after fingerprints of mac-worker-owned namespaces and of the isolated clone, and the helper/client revision match. It must omit application output, complete paths, clone origin, credentials, raw environment values, source content, and raw logs.

Automated local and fake-transport evidence

The post-rebase local gate was run against the final source. It is not a substitute for live three-worker acceptance. The ordinary sandbox initially prevented dashboard loopback binds; the escalated all-targets rerun passed.

Command	Revision measured	Recorded result
cargo fmt --check	final source	exit 0
cargo test --locked --all-targets	final source	exit 0 after escalated loopback-enabled rerun
cargo clippy --locked --all-targets -- -D warnings	final source	exit 0
cargo build --locked --release	final source	exit 0
git diff --check	final source	exit 0

Live three-worker acceptance

The v7 run used a fresh isolated root and a copied three-worker inventory. Claude Code is deferred by operator decision; Codex is the only agent for this run. The dashboard tasks view was observed separately. Source origin, push publication, Cursor, OpenCode, and worker gc are Phase 5d rows below with their live or explicitly unproven outcomes.

Required evidence	Sanitized record
Live-task source commit	a88a4cd. The v7 task executions used the isolated source clone at this revision.
Final post-rebase build/setup	Release build and worker setup were rerun after main advanced; helper/client digests matched on all three workers.
Worker protocol version	4 on setup and final ready-worker refresh.
Helper/client revision match	Match on all three workers during v7 and again after the final post-rebase setup. The installed helper digest equaled the corresponding release client digest; installation locations are omitted.
Selected worker names	mini-1, mini-2, mini-3.
Agent capabilities	agent:codex on all three; Codex 0.153.2 authenticated; no env profiles; worker Git identity absent on all three. Claude remains deferred. Cursor and OpenCode are Phase 5d rows below.
Sanitized command categories	Isolated setup; five-task batch; needs_input then say; cancel then say; pinned follow-up; --wip identity; disconnect/reconnect and killed runner; dashboard observe-and-shutdown; retained-metadata scan; Git-identity task.
1. Five-task run, default max_parallel, submitting shell exits	Run 040843ed. Positions 1–3 were active on distinct workers and positions 4–5 queued. All five ended closed/done; new-shell wait exited 0. The kit’s inherited origin first caused INVALID_ORIGIN (exit 64), with no task created; removing that origin and retrying produced the recorded run.
2. needs_input then say	Task d6e98229 on mini-2: the first turn ended needs_input with a session present; the prescribed resume hit the stale commit instruction’s read-only .git gate, then a corrective same-session follow-up completed done. Session continuity was present across all three turns.
3. cancel during a turn, then say	Task 73cf7888 on mini-1: the first turn was cancelled, the resumed turn ran the required test but hit the read-only .git gate, and a corrective same-session follow-up completed done. Cancel-then-resume completion was proven.
4. Follow-up pinned to a busy worker	Busy holder fa39e6f5 stayed on mini-3; its pinned follow-up did not migrate. Independent idle-first task c65fedde was admitted on mini-1 and completed. Live task projection exposed no blocking-code field.
5. --wip base	Task bc81e32a preserved the source clone’s HEAD, index, intentional dirty state, and diff-check result. After the marker-only untracked input was excluded from preflight, the turn ran; the read-only task sandbox blocked the agent’s attempted commit, but the publisher produced temporary result head d3f58fb4, which fetched successfully. --publish push was refused at preflight with TASK_CONFIG_INVALID (exit 64), with no task or remote mutation.
6. source = origin and publish = push	Proven in the phase 5d section below (rows "Source origin and push publication" and the push rows): source = origin completed against an SSH origin inside the pool, BASE_NOT_ON_ORIGIN and BASE_UNAVAILABLE were reported before any task or workspace was created, publish = push with --publish-branch landed the expected head on the origin, an uncommitted --wip base was refused, and a rejecting origin left the task open with PUBLISH_FAILED.
7. Disconnect and killed runner	Two logs -f followers disconnected and reattached to the same turn. A submit --wait disconnect and a separately validated killed runner each reconciled with one runner replacement; status/logs reconnected by the original task/turn. No second turn was accepted. Probe tasks were terminalized during cleanup.
8. Dashboard observe and shutdown	While item 1 was active, the loopback dashboard showed total 5, queued 2, active 3, and task/run counts matching the CLI; its active-job projection was 0. Selected detail and bounded stdout matched the active CLI state. Final snapshot showed all five closed and no active/queued records.
9. Retained metadata privacy	Fixed-script count-only scans found marker 0 and secret 0 everywhere; local records 0; worker records 0/1/0 for mini-1/2/3. The single mini-2 record hit was a tasks/<project>/<task>/status.json; count-only correlation showed its task was created before 2026-09-04 22:00 local, so it predates the redaction fix (an attempt-4/5 task). Application-log counts are transparency-only: local 16, mini-1 66, mini-2 80, mini-3 52.
10. Locked-keychain adapters and Git identity	Task bd6ee54d on mini-3 completed and fetched result head c5362a52; author and committer fields were present, but the worker had no configured Git identity, so attribution was not proven.
Worker namespace fingerprints	Final ownership fingerprints are below. Helper digest matched the client on all three; data namespaces changed as task metadata accumulated.
Unrelated remote entries	Immediate sibling counts and fingerprints were unchanged for each worker.
Original isolated clone	Baseline HEAD ad15b09, intentional README.md modification plus untracked scratch state, and clean diff check remained unchanged after submit and remote execution.

Namespace fingerprints

Local isolated namespaces:

Namespace	Before	After
config	1 / 4e39df499c7ffec1	1 / 4e39df499c7ffec1
state	0 / e3b0c44298fc1c14	131 / a0cb56454c9cbe75
cache	0 / e3b0c44298fc1c14	193 / 060015f422768acb
data	0 / e3b0c44298fc1c14	0 / e3b0c44298fc1c14

Worker-owned namespaces and immediate sibling sets:

Worker	Data before → after	Helper before → after	Data siblings before → after	Helper siblings before → after
mini-1	1784 / 0dafcd5465ce7ac5 → 2294 / acd0675177795e1c	1 / 2289bb5e6c55af5c → 1 / 2289bb5e6c55af5c	8 / be0eadec22acfb36 → 8 / be0eadec22acfb36	11 / 1505472269bf4716 → 11 / 1505472269bf4716
mini-2	800 / 6c53fb6b223cd2fb → 1290 / 27024ae47f578a00	1 / 2289bb5e6c55af5c → 1 / 2289bb5e6c55af5c	8 / 0bb9bc0127fafb03 → 8 / 0bb9bc0127fafb03	18 / 5b3d5df63db065e3 → 18 / 5b3d5df63db065e3
mini-3	753 / 4ea1a6993456a90f → 1167 / 8aaa686f70e81612	1 / 2289bb5e6c55af5c → 1 / 2289bb5e6c55af5c	7 / f07984e28cc77ee3 → 7 / f07984e28cc77ee3	10 / b4920d4ad5d0e033 → 10 / b4920d4ad5d0e033

The live record states that the validated helper and client revisions matched, without recording installation locations or connection details. It records only sanitized command categories and outcomes; application output is not safe evidence by default because it may contain application-emitted secrets.

Phase 5d

This section records the sanitized outcome for spec section 20.2 item 6 plus the Cursor, OpenCode, and retention rows. Live adapter turns were withheld when the refreshed worker facts did not provide authenticated agents or secure profiles; no credentials or profile values were created, uploaded, or printed.

Evidence captured or explicitly not proven, without recording secrets or complete local paths:

• Source origin and push publication: exact base preflight; origin:<host> routing; branch appears as --publish-branch; exit/state
• Cursor env-profile turn: prebind_session before first turn; profile-keyed capability; same chat on say; exit/state
• OpenCode env-profile turn: JSON session binding; --auto and pointer prompt; same session on say; exit/state
• Retention and discard: worker gc preview/apply; branch/mirror/transfer candidates; native delete if available; exit/state

Acceptance item	Exact-object / session	Capability / policy	Publication / continuity	Exit / state
Source origin and push publication	Prepared and seeded an SSH origin inside the pool; source=origin completed with accept/origin-note.txt; an unadvertised base was rejected as BASE_NOT_ON_ORIGIN before task creation, and an advertised-but-unavailable base returned BASE_UNAVAILABLE before workspace creation.	All three refreshed workers were ready on protocol 4 with authenticated Codex and the configured origin capability; the origin advertisement check succeeded from a worker.	publish=push completed with --publish-branch accept/phase-5d; refs showed that branch at the fetched result head; an uncommitted --wip push was refused as PUBLISH_REQUIRES_COMMITTED_BASE; the rejecting origin left the failure task open with PUBLISH_FAILED.	Source-origin preflight, worker fetch/verify, push publication, committed-base refusal, and rejection recovery were observed live; the failed publication task was closed during final cleanup after its open-state evidence was recorded.
Cursor env-profile turn	Local fake transport exercised prebind_session before the first turn and retained its binding for say.	Live refresh reported Cursor unknown and no secure env profile; no credential creation, upload, or printing.	Live first turn and say were withheld; worker-side chat continuity was not proven.	Local Cursor/OpenCode adapter suite: 13 passed; live Cursor state not run.
OpenCode env-profile turn	Local fake transport exercised JSON session binding, --auto, the pointer prompt, and resume.	Live refresh reported OpenCode unknown and no secure env profile; no credential creation, upload, or printing.	Live first turn and say were withheld; worker-side session continuity was not proven.	Local Cursor/OpenCode adapter suite: 13 passed; live OpenCode state not run.
Retention and discard	Complete release-binary fleet preview (protocol 4, apply false) listed 26 job candidates, all reason legacy protocol (15/6/5 across the three workers), plus two unreadable-record warnings; transfer candidates were 0.	The owner-only locked apply was authorized; legacy candidates are listed but protected from deletion, and no non-legacy candidate was selected. No task close/discard was invoked, so native session deletion was not exercised.	Apply exited 0 with 26 candidates and 0 applied; the post-apply preview exited 0 with the same 26 legacy candidates, 0 applied, and 0 transfer candidates. The retained legacy candidates survived.	Before/after counts were unchanged: local config/state/cache/data 1/4/0/0; worker data files 1837/1039/847 and installed helper present on all three. Preview/apply/post-preview warnings were 2/28/2.

Operator checklist

1. Use an isolated clone and isolated XDG roots. Install the release helper only after core Task 10. Do not record hostnames, complete paths, tokens, session identifiers, or transcript text.
2. Run spec section 20.2 item 6: submit with source = origin and publish = push; confirm only origin-capable workers are selected; confirm the pushed branch name matches --publish-branch; pin a worker that lacks origin:<host> and confirm CAPABILITY_MISSING without reroute.
3. Run a Cursor env-profile turn from a locked-keychain SSH session through CURSOR_API_KEY. Confirm prebind_session before the first turn, the profile-keyed capability, and the same chat on say.
4. Run an OpenCode env-profile turn through its secure profile. Confirm JSON session binding, --auto and the pointer prompt, and the same session on say.
5. Confirm a push failure leaves the task open with PUBLISH_FAILED. Preview then apply worker gc; confirm protected branches remain and only expired candidates are removed. Scan retained records for planted profile values and complete local paths; record hit counts only.
6. Record shortened IDs, aliases, states, exit codes, durations, helper/client revision match, and before/after fingerprints only.

Delivery boundary

The v7 Codex record above is supplemented by these Phase 5d outcomes. Item 6 and live Cursor/OpenCode turns remain unproven because worker-to-origin SSH authentication and authenticated adapter profiles were unavailable. Retention preview/apply completed with legacy candidates preserved and no candidate applied. Claude Code remains deferred on the workers by operator decision unless the operator records a later decision here.