Emerald fleet

Generated 2026-09-24 22:20:42 UTC · next refresh in 60s ·
Money
$24.12estimated spent today of $100.5 allowed
12Grok turns today, limit 50
0turns in the last hour, limit 15
0helpers running now, limit 8
A “turn” is one message sent into a Grok session — each one costs money. “Helpers” are extra workers a session runs alongside itself; each is charged separately, which is why they are capped. Dollar figures assume $2.01 per turn, worked out from the 22 Sep invoices. An estimate, not a reading from xAI.
Can finished work go live right now?
✅ PASS
The safety check passes. All 3 book questions return the right source and nothing has regressed. Finished work is allowed to go live.
Checked by gate_status.py, self-test 7/7 passed · read at 2026-09-24T22:07:46Z

How close are we to done

Fixes actually live on your real system2%
15 of 899 · This is the only one that means DONE. A fix that works on a test copy but was never put onto the real server does not count here.
Alarms proven to actually go off7%
43 of 586 · You have hundreds of alarms watching your system. Most have never been tested, so nobody knows if they would fire when something breaks. Each one gets deliberately broken to check it screams. This is the bulk of the work.
Finished work we opened up and double-checked40%
55 of 137 · When a session says 'done, here is the proof', someone opens the proof and confirms it says what was claimed. This is how five false 'done's were caught.
Jobs the sessions have finished76%
100 of 132 · CAREFUL: this is just how many jobs got done, NOT how fixed your system is. This bar reads high while the top one reads low. Adding these two together is exactly how you were once told 77% when the truth was 2%.
15running in production, of 899
543monitors not yet trustworthy
55claims with checked evidence
100finished cleanly
25finished BADLY
5still to do

Who is doing the work right now

Claude — 11 sessions: 6 working, 2 waiting on your decision (parkerup, shipyard)
Grok — 0 sessions paused, out of credit until Friday (none). One more moves to Claude every 2 hours if usage allows.
Your machines (local AI) — running the model tests, not real work yet. They start taking real jobs after the tests finish and the Director hookup is tested and approved (due Sep 30). Tests running now: simple3-studio: 8 passed, 0 failed, 0 never started so far
Live status as of 2026-09-24T22:20:39Z UTC.

Your personal work (MacBook terminals)

What you are doing yourself in Claude Code on the MacBook in the last 3 hours. The fleet sessions read this too, so they don't collide with your work.
Remove image ad from Megan's account 6 min ago · ~/prod-tracks/robot/watch
You: ok, should any of this be added or upadted for how the emerald fleet stuff is being worked on?
Where each claim comes from: - **"$0.00 off on all 7":** my own live check this session. I pulled each account from AdSignal's server and compared it with Meta. - **Ceilings 10x over, and Autopilot off:** from the AdSignal worker's own report (`NEEDS-SEAN.md`), not something I checked myself. I'll read the live values before running anything. The
LeadConnector authentication update 7 min ago · ~/.claude/projects/-Users-emeraldstudios/memory
You: how about this, if someone asks a question that the bot doesnt have an answer to, it may respond with something vague like I'd have to get back to you on that since it depends on a few things that I'd
Your idea is live on every client's bot. When a bot can't answer something, it now says: > "I'd have to get back to you on that since it depends on a few things. Happy to chat about it though, everyone's coming from a different spot" It then asks the next question from that owner's own script, word for word. The old "I'll get the owner to follow
Voice notes for GHL client messaging 26 min ago · ~
You: feel free to send it and then wrap up this session
Done. The email went out to Megan (conversationalenglishpro@gmail.com), and the session is wrapped up. **Live now:** - **Voice Note app:** Megan records on her phone or computer and it goes to her client as a text from the Conversational English Pro number. You heard the test play on your phone, and she's tested it herself. - **Her payment choice:
Run Claude audit handoff stage 2 38 min ago · ~/govlayer-build
You: Read _kb_stage2/claude-audit-handoff.md and run that audit. Report evidence. Do not fix. Do not start stage 3.
Skill: none — read-only evidence audit; `quality-eval` builds a golden set + LLM judge, and the handoff explicitly bars building anything (I only *ran* the existing golden harness). KB RECALL: REFUSE — code/ops audit of retrieval infrastructure, not a books question. No bulk mutation occurred: the only deletions were three scratch files I created t
Kenneithia sub account setup GHL 46 min ago · ~
You: <pasted_content id="01e6"> whats the issue with leads getting 3-4 unsolicited follow-ups? We're using manychat to start the conversation with her new followers AND THEN opening up the conversation. Li
A correction on the demo-page price: I haven't opened the live demo page. The $3,750 figure comes from the page text you pasted in your first message. That text shows $3,750 tuition, a $1,000 deposit, a $45 application fee and parent packages at $1,750 and $3,500. I don't have the page's URL, so I can't confirm what it shows right now. Send me the
Trey affiliate partnership and JV launch strategy 138 min ago · ~/.claude/projects/-Users-emeraldstudios/memory
You: ok, so we can wrap this session then?
Yes, it's safe to close this session. - **Nothing technical to clean up.** This session only sent email and set up the reply watcher, so no code, deploys, file locks or background jobs are left open. - **Memory is saved.** The deal terms, the rule that your 10% is never mentioned, the one-on-one-calls-first rule, and the watcher's ID are all in me

Needs you — nobody else can decide these

C57
Sean's own, deliberately deferred until all work is done, by his explicit instruction: rotate the four exposed credentials (Command Center admin cookie, two Telegram bot tokens, HL15 password), the production password already redacted from code, and the age secret exposed in chat || NEW EXPOSURE 2026-09-23T10:07Z, recorded by Claude's hourly ops run (id c71blast-20260923-1015). A read-only diagnostic this run ran `pm
C60
DECISION WAITING ON SEAN since 2026-09-20: Emmanuel's bot has contradicting sales rules and the work cannot proceed until he decides which rule wins. This is separate from tonight's repeat-question bug
C68
Tax: October 15 extension deadline with Steve Hawthorne EA at Belt Accounting. The 2025 statement gaps were resolved and sent; nothing is known to be outstanding, but the deadline is real and this is the only place it is written down
C70
HELD by Sean 2026-09-22: do not rotate the three exposed test credentials until all work is done. kadowcreates (Jason) keeps org read. Revisit with C57.
C95
DECISION WAITING ON SEAN, production ship: instr-b finished its first pass on slice 2 and proved 17 of 189 monitors actually fire when broken. It recommends applying a delete-gate fix to G.215-G.250 (safe-restart.sh sha 959857d0). It correctly did NOT apply it: that is a production change to the deploy path on the VPS, and a reviewer GO is not Sean's GO. RECOMMENDATION: ask instr-b for the exact commands, the backup,
C99
DECISION WAITING ON SEAN, live client bots, and it is the most serious thing on the board: unapproved code is RUNNING RIGHT NOW on /opt/command-center. The tree was left on 6597c0987 (the C71 identity-gate commit) at 2026-09-22T15:49:36Z and the server restarted onto it unattended at 2026-09-23T06:04:12Z. No session shipped it and no GO exists. RECORDED HERE 2026-09-23T12:15Z by Claude's hourly ops run (id c71livedec

Tried it, and it did not work

These finished, so they stopped showing as 'to do' — but the result was a failure or an undo. This section exists because a rolled-back upgrade once went unreported for exactly that reason.
C10botfleet
Count how many other leads across all nine accounts got a already-answered question re-asked
COULD-NOT-EVALUATE paraphrase-answered-reask cannot be auto-detected. EV:BF.18d4r.jsonl :1 vocal_academy 0/0
C11parkerup
OpenClaw 2026.9.5 upgrade + pm2-clawdbot.service recovery, with per-step checks and rollback
ROLLED BACK after step 1. Reason: the four post-step checks ran on the unchanged box and three failed, so ste
C22-instr-cinstr-c
First ten rows classified with a correct red proven, VPS load recorded (uses instr-a harness)
— G.399 NOT-RED, G.400 RED-CORRECT, G.401 CNE, G.407 NOT-RED, G.408 RED-CORRECT, G.410 RED-CORRECT, G.414 NOT
C24parkerup
Adjudicate 5 rows marked done whose own text says NOT RUN (SIDE pm2, PU.4, PU.5, PU.8, PU.13): verdict + RAN or PLAN-ONLY + evidence path each
SIDE pm2 command FAIL PLAN-ONLY CHECKPOINT.md:62; PU.4 PASS RAN CHECKPOINT.md:13; PU.5 PLAN.md PASS RAN CHECK
C32ops
J9 credential sweep: whether the three ~/live-products secrets ever reached a repo kadowcreates can read. The leak question first, cleanup second
FAIL. Exposed to kadowcreates via org read on those six repos. Passwords are in main history, not on those ti
C35grok
The six receipted items that exist only in a worktree or replica: land each with backup, RED-then-GREEN, UNDO by id and a post-restart pid+sha, or downgrade the status
six receipted wt/replica items DOWNGRADED (not landed). T4.D2 wt-T4.D2 (client replies). T4.C2 wt-T4.C2 (live
C42instruments
Repair the five defects: IN.F.046, 046b (fixture cloned from the wrong job), 047, 050, 051 (name mismatches for jobs that already exist live)
FOLDED-INTO-BATCH BATCH4 (046 FAIL stays). RECEIPT:C41:20260922 FOLDED-INTO-BATCH BATCH1–16. RECEIPT:C29:2026
C45ops
J4 dead-man's switch: the four-case live proof with the real hc-ping URL
DID-NOT-RUN sending the request is the ping, and this dispatch forbids the ping. Cases 1-3 were not run. Case
C46adsignal
Verify the three Autopilot/verdict ships an agent claimed done and Sean never confirmed
— The three claimed ships are not complete on live HEAD 50a412360fe35696d7d7482bf55f0639c2706a08. (a) FAIL. +
C76grok
Group G Flip GO, staged and batched - ANSWERS grok's own open ask at prod-tracks/CHECKPOINT.md:2013, which had sat unanswered ~4h while 38 LAND_NEEDED=yes copies-only rows piled up. Decided by Claude under Sean's standing authority (staging and batching a roll
batch 0 G.021 FAIL (restart proof). Landed wrap then UNDO.
C81instruments
Bookkeeping, no new work and no interruption to the BATCH14/15/16 cadence, which is healthy: C29, C41 and C42 have been open 5.5h with no token in instruments' own CHECKPOINT.md, so from outside they read as never done. For each, write the token with the evide
receipt token with nothing after it - no evidence cited
C86ops
WATCHDOG GAP, proven live today, and it is the reason three sessions sat dead for five and a half hours with no alert. The quota latch in robot/watch/fleet-watch.sh (_quota_latch_holds, id quotalatch-20260922-1315) suppresses frozen-<name> while the session ha
FAIL quota ceiling is present in /Users/emeraldstudios/prod-tracks/robot/watch/fleet-watch.sh:330-366 and dro
C85parkerup
Telegram channel disabled while plugin enabled — is Sean's ONLY alert path silently dead?
FAIL. CHANNEL_ENABLED false at channels.telegram.enabled. PLUGIN_ENABLED true at plugins.entries.telegram.ena
C88parkerup
pm2-clawdbot.service recovery — still PLAN-ONLY, and pm2 list is empty for clawdbot while 8 processes should run
FAIL. Pid file still 8 bytes mode 644 contents 1020288, user clawdbot, ppid 1, zero children. pm2 jlist EMPTY
H01ops
OLDER BACKLOG (Mar): a Parker error-detection fix was written as a prompt and never sent — when Parker's reply is an API error like 'service overloaded', it shows that as his current task instead of an ERROR state. Verify whether this is still true, then fix o
FAIL POST /api/parker-chat/reply stores the reply text as the task detail and logs ok with no lastError at /U
H02botfleet
OLDER BACKLOG (Jun): the 'AFFECTS ALL BOTS' systemic flag was verified in code but has NEVER been seen to render end-to-end. The clean cross-account demo (seed two throwaway accounts, confirm the flag fires, purge) was agreed and never run. Same shape as an un
PARTIAL-CLOSE, sandbox proof PASS, live throwaway-account demo still blocked - no redispatch, reused existing
H07ops
OLDER BACKLOG (Feb): leftover test pages (Paws & Suds) were never cleaned off vocalacademy3.com. Confirm they exist, then propose removal with an export-first path — no deletions without a GO.
COULD-NOT-EVALUATE vocalacademy3.com returned 200, title Vocal Essentials, 1592554 bytes, paws count 0, suds
H08ops
OLDER BACKLOG (Feb): three proofs that the products actually work end-to-end were agreed and never done — a real Instagram DM test, a real client page build via Parker, and confirming the smart-alert rule (only alert after 3 consecutive failed checks) is live.
FAIL smart-alert is not live as specified. /home/clawdbot/inbox-failsafe/health-watchdog.sh:88 sets the thres
C91ops
The cross-account 'AFFECTS ALL BOTS' flag has only ever been verified in code, never seen fire. Run the clean demo: seed two throwaway accounts with one symptom, confirm the Pending Fixes panel renders the systemic flag end to end, then purge them
FAIL live seed was not run and the Pending Fixes panel was not opened. Blocker: .SEAN-GO absent and GRANTS.md
C94ops
Delete the leftover Paws & Suds test pages from vocalacademy3.com - export first, nothing removed without a listing Sean can see
COULD-NOT-EVALUATE no written page on vocalacademy3.com for Paws and Suds. COMMITMENTS.md:76. TODO-next-sessi
C97-instr-binstr-b
Break down the remaining slice rows by instrument family, then extend the shared harness to the largest family it cannot run, with a negative control
receipt token with nothing after it - no evidence cited
A-instruments-20260922-1813instruments
Auto-assigned batch of 40 instrument rows (G.067-adsignal-datamaturity-gate-regressio .. G.106-adsignal-moneycap-audit-regression-j): prove each negative test goes the RIGHT red, sandbox only || BLOCKER RECORDED 2026-09-23T04:28Z by Claude's hourly ops run (id
COULD-NOT-EVALUATE blocker-stands id=instrscope-20260923-0428 COMMITMENTS.md:73 COMMITMENTS.json[118] asked_a
A-instr-c-20260922-2019instr-c
Auto-assigned batch of 40 instrument rows (G.188-backup-to-r2 .. G.268-adstudio-edit-guard-regression-js): prove each negative test goes the RIGHT red, sandbox only
G.188-backup-to-r2 other cannot-be-proven not-run; G.189-dr-staleness other cannot-be-proven not-run; G.190-f
A-instr-c-20260922-2129instr-c
Auto-assigned batch of 40 instrument rows (G.269-adstudio-endcard-fit-regression-js .. G.308-ask-guard-py): prove each negative test goes the RIGHT red, sandbox only
receipt token with nothing after it - no evidence cited
J402ops
VPS disk at 97% (2.7G of 75G free) — EXPORT-FIRST PLAN ONLY, delete nothing. When it fills, the client bots stop taking messages and nothing can be deployed. Measured 2026-09-23T22:10Z (robot/watch/jobs/vadisk.out): /home/clawdbot 34G; reference 5.3G (drive-kb
receipt token with nothing after it - no evidence cited

Broken right now

🚨 A version of your client-bot code that you never approved (9424dda64) is RUNNING RIGHT NOW. It was put in the live folder 2026-09-24 21:56Z and the bot restarted onto it at 2026-09-24 21:58Z. This changes what the bot says to real leads. Ask Claude to look. WHO: the ship-lock journal on the box records session 'deflect' taking the lock at 2026-09-24 21:57:05Z for "owner handoff -> soft deflect +
⛔ Out of Grok Build credit for the week. These sessions have stopped and cannot restart until credits are bought: grok. This is ONE cause, not several separate failures. Buying credits is a spend, so nobody but you can clear it. Waiting is NOT an evidenced plan: no source anywhere records when the weekly allowance resets, and the only renewal date anyone had (09-23) was re-probed at 05:11:30Z, aft
an assistant is running on your Mac with the safety prompts switched off in: adsignal botfleet estate instr-a instr-b instr-c instruments main ops parkerup shipyard. Nothing on file says you approved that. It can change files, including live client-bot code, without asking.
🪫 adsignal has finished everything assigned to it and has NO jobs left. It is not stuck — it is waiting for Claude to give it more work. Eleven idle sessions look identical to eleven busy ones on every other check, which is why this exists.
🪫 estate has finished everything assigned to it and has NO jobs left. It is not stuck — it is waiting for Claude to give it more work. Eleven idle sessions look identical to eleven busy ones on every other check, which is why this exists.
🪫 instr-b has finished everything assigned to it and has NO jobs left. It is not stuck — it is waiting for Claude to give it more work. Eleven idle sessions look identical to eleven busy ones on every other check, which is why this exists.
🪫 instr-c has finished everything assigned to it and has NO jobs left. It is not stuck — it is waiting for Claude to give it more work. Eleven idle sessions look identical to eleven busy ones on every other check, which is why this exists.
🪫 ops has finished everything assigned to it and has NO jobs left. It is not stuck — it is waiting for Claude to give it more work. Eleven idle sessions look identical to eleven busy ones on every other check, which is why this exists.
🪫 parkerup has finished everything assigned to it and has NO jobs left. It is not stuck — it is waiting for Claude to give it more work. Eleven idle sessions look identical to eleven busy ones on every other check, which is why this exists.
🪫 shipyard has finished everything assigned to it and has NO jobs left. It is not stuck — it is waiting for Claude to give it more work. Eleven idle sessions look identical to eleven busy ones on every other check, which is why this exists.
🧊 1 session(s) are stopped on the Grok credit screen and cannot continue until credits are bought AND each pane is cleared. Everything halts until then — this is the single most expensive stall there is, because nothing else can run either.
🟠 7 of your client bots are running but some are not. Not working: voice-agent (restarting over and over). Client impact is UNKNOWN and unverified: a restart loop is all the process list can show, and it cannot prove whether clients are being answered. No reachability check could be run (set BOTSUP_PROBE_URL to give this check one). Do not read this as fine. Ask Claude to look.
🟠 The server running your live client bots is 93% full (5.4G left) and still filling (1321MB less free than 22min ago). Worth clearing before it becomes urgent.

Your machines

These are your own computers. They cost nothing to run work on, unlike the Grok sessions below.
MacBook Pro — ROOM TO SPARE
runs the main Grok session (paused till Fri), the job runner and the checkers
working, with room to spare · 4.0 of 10 cores in use · 6 more test jobs could run here now
Mac Studio — ROOM TO SPARE
runs the 10 work sessions (Claude + paused Grok) and the model tests
working, with room to spare · 11.2 of 24 cores in use · 12 more test jobs could run here now
HL15 — ROOM TO SPARE
runs tests, and has seven AI models
idle - room for more work · 0.37 of 20 cores in use · 15 more test jobs could run here now
VPS — ROOM TO SPARE
runs your live business - not used for testing
working, with room to spare · 2.49 of 8 cores in use · 5 more test jobs could run here now
Skytech PC — ROOM TO SPARE
writes the test scripts - the only machine with a fast GPU
idle - ready to write the next test
Total free test capacity right now: 38 jobs at once, at no cost.

Local AI model tests

Test plan progress 86%
30.8 of 36 clean runs done (includes the run in progress). Not finished: studio-qwen3.8-27b simple 2/3, studio-qwen3.6-35b-a3b hard 1/3, studio-qwen3.6-35b-a3b simple 0/3
Every run is listed. Numbers are passed / failed / never started. Hard set = 9 real past jobs per run, simple set = 10. Each model needs 3 clean runs per set before it is trusted.
model and machinesetruns (passed/failed/never started)clean totalmin per pass
hl15-qwen3.6-35b-a3bhardagent1-hl15: 7/2/0
agent10-hl15: 6/3/0
agent2-hl15: 7/2/0
agent3-hl15: 6/3/0 EXCLUDED: 1 job then restarts; replaced by agent10-hl15
agent4-hl15-ctx32k: 5/4/0
25 of 36 (4 clean runs)45.4
skytech-qwen3.6-27bhardagent4-sky: 5/4/05 of 9 (1 clean run)51.0
skytech-qwen3.6-35b-a3bhardagent1-sky-b: 5/2/2
agent1-sky: 0/0/9 64k memory crash (setting) - fixed at 32k
agent5-sky: 4/5/0
agent6-sky: 5/4/0
14 of 36 (4 clean runs)12.1
skytech-spark-4bhardagent7-sky: 0/9/0
agent8-sky: 1/8/0
agent9-sky: 0/9/0
1 of 27 (3 clean runs)87.3
skytech-vortex-9bhardagent7-sky: 2/7/0
agent8-sky: 1/8/0
agent9-sky: 1/8/0
4 of 27 (3 clean runs)36.9
studio-qwen3-coder-nexthardagent1: 1/8/0 early run, before one-at-a-time fix1 of 9 (1 clean run)204.1
studio-qwen3.6-27bhardagent2-studio: 7/2/0 EXCLUDED: settings changed mid-run
agent4-studio: 4/5/0
agent6-studio: 1/3/0 stopped by us (27B ruled out)
4 of 9 (1 clean run)109.2
studio-qwen3.6-35b-a3bhardagent-smoke: 0/1/0 setup check, not a scored run
agent1: 4/5/0 early run, before one-at-a-time fix
agent9-studio-35b-64k: 5/4/0
9 of 18 (2 clean runs)23.5
studio-qwen3.8-27bhardagent11-studio: 2/7/0
agent3-studio: 3/6/0 first start only counts (2/3 + 4 stopped by us)
agent5-studio: 4/5/0
9 of 27 (3 clean runs)31.3
hl15-qwen3.6-35b-a3bsimplesimple-hl15: 9/1/0
simple2-hl15: 10/0/0
simple3-hl15: 9/1/0
28 of 30 (3 clean runs)8.6
skytech-mimo-9bsimpleagent6-simple-sky: 0/10/0 MiMo = not compatible (never used the tools)0 of 10 (1 clean run)—
skytech-qwen3.6-35b-a3bsimpleagent6-simple-sky: 8/2/0 MiMo = not compatible (never used the tools)
simple2-sky: 10/0/0 Qwen: 6 never started (server crash at load) - counts
simple3-sky: 10/0/0
simple4-sky: 10/0/0
38 of 40 (4 clean runs)1.2
skytech-spark-4bsimpleagent6-simple-sky: 7/3/0 MiMo = not compatible (never used the tools)
simple2-sky: 4/6/0 Qwen: 6 never started (server crash at load) - counts
simple3-sky: 7/3/0
18 of 30 (3 clean runs)11.9
skytech-vortex-9bsimpleagent6-simple-sky: 6/4/0 MiMo = not compatible (never used the tools)
simple2-sky: 7/3/0 Qwen: 6 never started (server crash at load) - counts
simple3-sky: 6/4/0
19 of 30 (3 clean runs)2.1
studio-qwen3.6-27bsimplesimple-studio: 10/0/010 of 10 (1 clean run)7.8
studio-qwen3.8-27bsimplesimple-studio: 9/1/0
simple2-studio: 9/1/0
simple3-studio: 8/0/0
26 of 28 (3 clean runs)4.9

What each worker is doing right now

Green = moved in the last 30 minutes. Amber = over 30. Red = over 90 minutes, which usually means stuck.
workerstatelast movedhelperswhat it is doing
grokREPLACED - its list is now worked by main (Claude)——Grok (MacBook) is out of credit until Fri
botfleetWORKING (Claude)now0EN/test-fixtures/ai-identity-deflect.js
instrumentsIDLE (Claude)now0the next ANSWER.md update.
opsWORKING (Claude)now1◯ general-purpose Reading command-center package.… 2m 30s · ↓ 107.7k tokens
adsignalIDLE (Claude)now0real per-account goldens and re-run the battery. Full detail in CHECKPOINT.md.
shipyardWAITING ON YOUR DECISION (Claude)now0appended to CHECKPOINT.md.
parkerupWAITING ON YOUR DECISION (Claude)now0replied to any of the three queued items and you want me to act on his answer.
estateWORKING (Claude)now0✶ Sock-hopping… (6m 6s · ↓ 17.9k tokens · thinking more)
instr-aWORKING (Claude)now1◯ general-purpose Copying sandbox files, launching… 5m 13s · ↓ 86.9k tokens
instr-bWORKING (Claude)now0✽ Calculating… (2m 22s · ↓ 15.3k tokens)
instr-cIDLE (Claude)now0Enter to select · ↑/↓ to navigate · Esc to cancel
mainWORKING (Claude)now1◯ general-purpose Reading CHECKPOINT.md lines 757… 10m 0s · ↓ 293.8k tokens

Still to do — 5

A-instr-a-20260923-2240instr-aAuto-assigned batch of 30 instrument rows (G.557-meta-safe-symlink-guard .. G.586-webhook-auth-parity-js): prove each negative test goes the RIGHT red, sandbox only
C90 oldermain (Claude)Parker error detection: when Parker's reply is an API error like 'AI service temporarily overloaded', stop showing that as his current task - show ERROR and keep the real task visible. Written as a prompt in March 2026 a
C92 oldermain (Claude)Director empty-deliverable check: a patch was claimed applied in August but never confirmed. Verify the helper exists, is called from the review path, and both supervisor regressions pass - or finish it
C93 oldermain (Claude)Director board legibility: the 'N things need you' counter counts cards that are auto-shipping or already finished, and the board does not separate what needs Sean from what is handling itself. Never started
IN.F.STEP2-AinstrumentsIN-SCOPE WORK FOR instruments, recorded because the fleet kept reporting it as having no jobs left while the only row under its name was a Group G batch its own RULES forbid (A-instruments-20260922-1813, now closed COULD

Older leftovers — 2

Things left open in conversations from March to August that were never on any list. Each starts as a question — is this still real — because some are certainly dead by now. Being done after this week's work, at your instruction.
H06botfleetOLDER BACKLOG (May): jaclyn_stapp's inactivity-followup was never being registered — /api/inactivity-followup/register was not called, and 9 of 10 recent contacts had zero follow-up touches. Verify on live.
H09grokOLDER BACKLOG (Aug, Jason, P3): OPS-260811-001, a skill to write in ASD-STE100 Simplified Technical English plus the Google developer documentation style guide, for docs and git messages. Jason's own note records it was adopted wi
Done and verified — 100 (click to open)
A-estate-20260923-2240 · estate · Auto-assigned batch of 40 instrument rows (G.477-fc-healthz-3144 .. G.556-meta-blindspot-audit): prove each negative test goes the RIGHT red, sandbox
A-instr-a-20260923-2222 · instr-a · Auto-assigned batch of 40 instrument rows (G.397-instagram-cookie-failure-detection-r .. G.436-repo-hygiene-page-dedup-regression-j): prove each negat
A-parkerup-20260922-2019 · parkerup · Auto-assigned batch of 40 instrument rows (G.147-adsignal-winner-ghl-precedence-regre .. G.187-guard-liveness-check-sh): prove each negative test goes
A-parkerup-20260923-2222 · parkerup · Auto-assigned batch of 40 instrument rows (G.356-coach-serverjs-keynote-regression-js .. G.396-ig-website-fill-regression-js): prove each negative tes
A-parkerup-20260923-2240 · parkerup · Auto-assigned batch of 40 instrument rows (G.437-repo-ownership-regression-js .. G.476-run-fix-queue-dispatch): prove each negative test goes the RIGH
A-shipyard-20260922-1748 · shipyard · Auto-assigned batch of 40 instrument rows (G.025-adsignal-account-history-regression- .. G.066-adsignal-dashboard-field-parity-js): prove each negativ
A-shipyard-20260922-2019 · shipyard · Auto-assigned batch of 40 instrument rows (G.107-adsignal-multi-offer-parity-js .. G.146-adsignal-window-unification-regressi): prove each negative te
A-shipyard-20260923-2222 · shipyard · Auto-assigned batch of 40 instrument rows (G.309-awesome-design-public-list-regressio .. G.355-coach-run-bound-regression-js): prove each negative tes
B-notdeployed-20260923 · instr-b · Itemise the 67 KNOWN NOT DEPLOYED rows PROGRESS.md counts but never names: classify each (on-disk-only / staged-not-landed / not-restarted / committed
C1 · grok · Settle live T4.FC: does unparseable compliance SEND or BLOCK, with live md5 and real-lead count
C12 · adsignal · Adjudicate 7 AS.SEC1* rows marked pass with no receipt cited
C13 · shipyard · Adjudicate 8 SY.1* rows marked proven with no receipt, incl. the 56-item ledger claim
C14 · grok · Adjudicate 5 claim-vs-evidence conflicts; say which are real and which are my checker misreading EXPECT_FAIL
C15 · botfleet · Fix recommendation for the paraphrase-blind already-asked check: exact commands, backup, UNDO by id, RED-then-GREEN. Recommendation only, no live ship
C16 · instruments · Five rows say landed while their own proof text says NOT landed; correct them and audit every other row for the same split
C16-ops · ops · Adjudicate J3 (proven, no receipt cited), name its real owner, and state the J4 dead-man blocker in one line for Sean
C17 · adsignal · Verify all six work/as-int fixes against their receipts BEFORE shipping; name any held back
C18 · adsignal · Ship work/as-int to production with backup, before/after HEADs, per-check results and a proven rollback
C19 · adsignal · Anthony's AdSignal menu link iframe->new tab, or the exact clicks for Sean if UI-only
C2 · grok · Verify vps-t4fc-live-stage is an empty repo and say what that does to receipts citing it
C20 · adsignal · AS.EVAL: graded knowledge-base + Autopilot evaluation - does AdSignal's advice actually follow the KB, and how many Autopilot money-spending decisions
C21 · adsignal · Settle AS.J1a-d (all FAIL) and say whether they affect the advice path or only display
C22-instr-a · instr-a · First ten rows classified with a correct red proven, VPS load recorded (harness owner)
C22-instr-b · instr-b · First ten rows classified with a correct red proven, VPS load recorded (uses instr-a harness)
C23 · grok · Adjudicate the 2 CONFIRMED contradictions (G.007b-nightly-drift-land, T3.3c) with a verdict + evidence path each, and give Sean the one-line Group G l
C25 · ops · EMPTY-SKIP blind spot in grokbot/robot.py: an idle session with an empty report is invisible for 90 min. Fix with backup, RED-then-GREEN, UNDO by id,
C26 · botfleet · AI-denial on rung 2: DEFLECT_FALLBACK reaches a code path that never reads the rule, so whether the bot DENIES being AI is unestablished. Legal exposu
C27 · botfleet · The 50-minute window 19:17-20:07 on 2026-09-19 with work but no stop, no log line and no receipt - establish whether a ship ran in it
C28 · botfleet · BF.13esclock: Meta's behaviour must be read from the recorded response body, never the exit code; confirm which was used
C29 · instruments · Close the 009L Mac-side UNDO gap, and repair the corrupted dispatched-ts column
C3 · grok · Resolve the three self-contradicting rows and add CONTRADICTED as a blocking ledger state
C30 · grok · T3.3f settled against evidence - it currently gates nine T4.1b client rollouts while its own records disagree (table FAIL, detail claims proven, gate
C31 · ops · J6: VPS load average 7.44 sustained, disk 86%, and the token-monitor-agent crash loop that restarted 28 times with no alert
C33 · grok · Report what fraction of items marked proven are actually running in production, generated from the DEPLOYED column, replacing the 77% figure that coun
C34 · grok · List by id every row that is copies-only / worktree-only / not landed, so the dark set is enumerable instead of estimated
C36 · grok · Install the t31 secret-scan drafts
C37 · grok · Jason's four critiques: the loss-guard, two-signal paging, unsourced numbers in displayed output, and the tenant config store
C38 · botfleet · Ratchet root-cause fix - forward-looking only, explicitly NO bulk sweep across client rules
C39 · botfleet · BF.15cg two-rule deactivation on cg_guitar (shrule_f50ba1a4680a, shrule_3d822c95a3af) - reversible by id, needs a GO before touching a live bot
C4 · grok · Confirm no T4.1b client rollout was dispatched by the bad watchdog order
C40 · botfleet · AI-suspicion tracker: build it, and resolve the unresolved 15-vs-31 baseline conflict first
C41 · instruments · Batched landing: noise forecast first (hold out any heartbeat that would page immediately), then batches of five with a 15-minute quiet check between
C43 · ops · J5c: make the session map data-driven and the subagent count trustworthy. Prefixes for the new sessions are already added - verify they took effect
C44 · ops · J7 is landed but unledgered, and J8 has not started - ledger J7 properly and say what J8 is blocked on
C47 · adsignal · The remaining backlog: silent ask-first queue, HOLD visibility, P4.5 hook card, numeric-consistency guard, Anthony's menu link, webhooks on the remain
C48 · shipyard · Confirm the .git state in govlayer-build, and reconcile the 94-vs-7 unpushed-commit conflict
C49 · shipyard · Deploy medic retry, and the agent-factory lock verification
C5 · estate · vps-t4fc-live-stage alone: tree SHA, file list, bytes, commit count, staged-artifact verdict
C50 · estate · ES.2: generate session status from receipts instead of restating it
C51 · estate · ES.3: the eleven single-file repos that are each never-checked
C52 · estate · ES.4: split the source-of-truth repo from the backup mirror
C53 · estate · ES.5: stop committing runtime state into repos
C54 · estate · ES.6: express the size ceiling as a glob
C55 · estate · ES.7: use branches rather than tree copies
C56 · estate · ES.8: one canonical home for the duplicated reference material
C58 · grok · T4.FC test conflict - DECIDED by Claude: update the test, keep fail-closed behaviour, RED-then-GREEN, no live change
C59 · adsignal · autopilot-fire-gate-error - DECIDED by Claude: silent to the customer, internal page every occurrence, escalate after two failed retries
C6 · estate · Full list of repos sharing the lone-.gitignore HEAD tree, with parents:0 confirmed per repo
C61 · botfleet · VERIFY CLOSED OR FIX: the BotLab DM bot fabricated booking links and pricing to real clients, and two real leads received incorrect information (found
C62 · adsignal · VERIFY: AdSignal Autopilot was found on 2026-09-18 to have never made a real change on any client account because tracking and attribution were unwire
C63 · grok · The reel pipeline: server.js was patched on disk but never committed because another session held the file, and the restart was deferred (safe-restart
C64 · grok · The knowledge-base index: 24KB of files were unindexed with no progress between two checks on 2026-09-21. Establish whether indexing is still stalled
C65 · ops · The GitHub share sync had never run as of 2026-09-21. Its heartbeat is now watched, but confirm it has actually run at least once and that Jason is se
C66 · ops · The evidence-gate reporting bug tied to a missing job id, open since 2026-09-21
C67 · grok · The Command Center audit: eight cleanup items (HL15, Stores, AI Hub, Onboarding, DMs, Direction, Coach, Boardroom) are unanswered because Sean could n
C69 · grok · SHIP the fail-closed compliance fix to live (Sean GO 2026-09-22), plus establish whether the two real leads actually received the unchecked replies
C7 · estate · A check that catches gitignore-only repos, proven with a negative control (RED then GREEN)
C71 · botfleet · SHIP the fail-closed identity gate AND neutralise the server.js system-prompt instruction to deny being a bot. BLOCKER CORRECTED 2026-09-22T23:15Z by
C71b · botfleet · HYPHENATED IDENTITY PHRASINGS STILL SLIP THE GATE. Measured 2026-09-23T22:09Z by Claude's hourly ops run (job robot/watch/jobs/c71averi.sh, output job
C72 · grok · ADJUDICATE T3.3c, the only CONFIRMED contradiction on the fleet. Its state column says proven while its own proof cell says 'G other-5 hash+GET not ru
C73 · ops · The GitHub share sync is HUNG, not stale: last completion 05:41:05Z, a run started 05:56:05Z still holds the lock, heartbeat frozen at 1790055665 and
C74 · ops · SHARE-SYNC RECOVERY (rescoped by Claude 2026-09-22T13:15Z; ops is out of Grok Build credit and cannot file this itself). The original text asserted th
C75 · estate · ADJUDICATE ES.1c gates, the only CONFIRMED contradiction on the fleet. State column says PASS, its own proof text in the same cell says COULD-NOT-EVAL
C77 · botfleet · botfleet was idling with 22 runnable rows and 0 subagents, holding ALL of them on 'GO tg 221 outstanding' - which gates only the C26fix LIVE LAND. Hol
C78 · grok · THE SHIP PATH IS BLOCKED, and it blocks every Group G row. C76 batch 0 closed correctly (abort on first failure, UNDO g021-c76-20260922 ran, live rest
C79 · ops · (1) File the share-sync recovery under its own id: the C73c work IS the answer to C74 but was written as RECEIPT:C73b/C73c, so C74 can never close and
C8 · estate · Generate the repo total with the command printed beside it; stop restating 218/231/238
C80 · botfleet · Self-imposed freeze VOIDED BY NAME (second time today; C77 was the first): the excuse 'GO outstanding (tg 221); every remaining row is Sean-gated' and
C82 · ops · Carried forward out of C74 so it is not lost when C74 closes on the C73c receipt. Establish whether the C73b change - a heartbeat-write failure exits
C83 · parkerup · SHIP_LOCK PU-LIVE-UPGRADE-3 LAPSED at its 2026-09-22T12:00Z deadline with nobody present: parkerup hit the Grok Build weekly limit at 10:45:35Z and th
C84 · SEAN · DECISION WAITING ON SEAN, and it outranks the others because it is the one that could cause an unauthorised change to a LIVE client bot. IT IS NOW TWO
C85 · estate · Bookkeeping, no new work. Your C75 adjudication is CORRECT and accepted - CHECKPOINT.md:119 settles ES.1c as COULD-NOT-EVALUATE by way (b) and :120 ca
C86 · parkerup · The sixth OpenClaw agent: missing, or is the check wrong about six?
C87 · parkerup · The five cron lines: which exist, which are disabled, which are genuinely missing
C89 · parkerup · PU-LIVE-UPGRADE-3 lock lapsed at 12:00Z — confirm live is still baseline and write the lapse line
C9 · botfleet · Emmanuel repeat-question defect: is it paraphrase-blind matching (A) or a slot-fill failure (B)
C96-adsignal · adsignal · Instrument slice of 114 rows (G.135-adsignal-strategy-chat-regression-js .. G.248-stress-tester-regression-js): classify, then prove each negative tes
C96-estate · estate · Instrument slice of 110 rows (G.477-fc-healthz-3144 .. G.586-webhook-auth-parity-js): classify, then prove each negative test goes the RIGHT red, sand
C96-instruments · instruments · Instrument slice of 114 rows (G.021-rule-guardrail-regression-js .. G.134-adsignal-strategist-isolation-js): classify, then prove each negative test g
C96-parkerup · parkerup · Instrument slice of 114 rows (G.363-contactid-doubling-regression-js .. G.476-run-fix-queue-dispatch): classify, then prove each negative test goes th
C96-shipyard · shipyard · Instrument slice of 114 rows (G.249-structural-rule-drift-regression-js .. G.362-contact-stage-regression-js): classify, then prove each negative test
C97-instr-a · instr-a · Break down the remaining slice rows by instrument family, then extend the shared harness to the largest family it cannot run, with a negative control
C97-instr-b2 · instr-b · RE-ASK of C97-instr-b, which closed BAD/NO-EVIDENCE: the checkpoint carries the bare marker RECEIPT:C97-instr-b:20260923 with nothing cited under it,
C97-instr-c · instr-c · Break down the remaining slice rows by instrument family, then extend the shared harness to the largest family it cannot run, with a negative control
C97-parkerup · parkerup · parkerup was alive and idle for two hours behind a STALE NEEDS-SEAN.md from 18:11 - pane_read 20:10:30Z shows a free input box and 'Worked for 34s', n
C98 · ops · VPS DISK CEILING. Measured first-hand 2026-09-22T20:15:37Z: /dev/sda1 75G, 70G used, 2.2G free, 98%. Was 11G free at 04:02:30Z and 2.7G at 16:48Z (est
H03 · ops · OLDER BACKLOG (Aug): the Director empty-deliverable check was applied but never verified — it is supposed to send a job back when its summary contains
H04 · ops · OLDER BACKLOG (Aug): Director board legibility was agreed and never started — the 'N things need you' counter counts cards that are auto-shipping or a
H05 · botfleet · OLDER BACKLOG (May): a merge-field empty contact_id leak affecting conversational_english, jaclyn_stapp and need_more_chops. Verify whether it was eve
J401 · ops · voice agent pm2 crash loop — DIAGNOSIS ONLY, no restart. pm2 reports voice-agent 'online' with restart_time 23937 at 2026-09-23T21:49Z and 24422 at 22

What this page covers

Everything. This week's work between Sean and Claude, audited against the actual conversation transcripts rather than anyone's memory, plus the older leftovers from March to August that were never on any list. The only things deliberately left out are Jaclyn's exercise player and anything to do with Balance & Becoming, both at Sean's instruction.

If something is missing from this page, it is because it was never written down anywhere — said out loud, or on a phone call. That is the one gap no file can close.