Autonomous work

Act now - The operator: review the numbers on CV — NUMBER-GITHUB-PUBLISHER-cv - The operator: review the numbers on MARKETING-BRANDING — NUMBER-GITHUB-PUBLISHER-marketing-branding - The operator: review the numbers on PRESS-ROOM — NUMBER-GITHUB-PUBLISHER-press-room - The operator: review the numbers on RESEARCH-LAB — NUMBER-GITHUB-PUBLISHER-research-lab - The operator: review the numbers on agent-skills-pack — NUMBER-GITHUB-PUBLISHER-agent-skills-pack - The operator: review the numbers on data-eng-templates — NUMBER-GITHUB-PUBLISHER-data-eng-templates - The operator: review the numbers on science-youtube-manim — NUMBER-GITHUB-PUBLISHER-science-youtube-manim - The operator: open the remaining 13 record(s) on the board

This batch filed 5 launch page(s), 5 number(s) for MARKETING-BRANDING, ready for the press room to read.

Act now - The operator: review the numbers on ledger — NUMBER-audiobooks-player-ledger - The operator: open the launch page for public-copy — LAUNCH-audiobooks-player-public-copy - The operator: nothing further, this folder filed one line

Act now - The operator: open the launch page for economics — LAUNCH-birthday-treasure-hunt-economics - The operator: review the numbers on free-sample — NUMBER-birthday-treasure-hunt-free-sample - The operator: clear the wall blocking product — WALL-birthday-treasure-hunt-product

Act now - The operator: review the numbers on economics — NUMBER-cozy-logic-puzzles-economics - The operator: nothing further, this folder filed one line

This batch filed 6 launch page(s), 9 number(s), 1 wall(s) for data-eng-templates, ready for the press room to read.

This batch filed 1 number(s) for niche-content-engine, ready for the press room to read.

Act now - The operator: review the numbers on discovery — NUMBER-paid-code-bounties-discovery - The operator: nothing further, this folder filed one line

Act now - The operator: review the numbers on ledger — NUMBER-pi-shell-ledger - The operator: open the launch page for public-copy — LAUNCH-pi-shell-public-copy - The operator: nothing further, this folder filed one line

This batch filed 2 number(s) for public-domain-audiobooks, ready for the press room to read.

# Research lab: RESEARCH-LAB ## Kept - Direct every idea: UNMEASURABLE: the try was taken while the row was LIVE at pass 3 of 12, so verdict would bill the edit for 3 pre-edit passes. Kept on its own checks: voice-check 0, prompt-lint 0, hypotheses 5 -> 1. research-lab try now refuses a LIVE row (exit 5). - Publish the site-⚙️: before 0.00 after 0.00, 1 run(s), 0 broke; cured 1 of 3 broke runs - Publish every ledger to the press room: before 0.00 after 236.00, 1 run(s), 0 broke; cured 1 of 4 broke runs - x402-data-endpoint/00-ledger: before new after 4.00, 1 run(s), 0 broke; before was new, so the keep rests on the run not breaking, never on a raised score - self-hosting-video-course/00-ledger: before new after 4.00, 1 run(s), 0 broke; before was new, so the keep rests on the run not breaking, never on a raised score - spanish-open-data-mcp/00-ledger: before new after 4.00, 1 run(s), 0 broke; before was new, so the keep rests on the run not breaking, never on a raised score - science-youtube-manim/00-ledger: before new after 4.00, 1 run(s), 0 broke; before was new, so the keep rests on the run not breaking, never on a raised score ## Reverted - public-domain-audiobooks/01-voice-and-first-book: tried: book one is ONE work, not the 12-story Leyendas volume. UNTESTED, because prompt-run --again resumed the row's OLD loop at pass 12 of 12 and its OLD handoff, so the run polled render_all.sh for 12 of 12 wavs and never read the edited body. Next pass: reset the row's loop state before trying an edit on a capped row, and note the done-when tree check reads ~/git/autonomous-work, which ten departments leave dirty. - self-hosting-video-course/00-ledger: tried: pin telegram:-1004404879807 on the cron and the send. The WORLD-CHECK PASSED and stands: hermes cron list shows course-ledger delivering to the channel id, where it delivered to the blocked home bot before. The revert rests on 4.00 -> 2.00, a key rate the edit never touched. The root cause is AUTONOMOUS-WORK/Every row says.md, the brief every row is written from, which pass 11 fixed outside the experiment. Next pass: try a row whose score the edit can move. ## Trying - none ## Pass 5 Findings The ranking research-lab hypotheses hands every pass was noise over AUTONOMOUS-WORK, and two commits repaired it. - director score divided ADVANCED ledger keys by run-hours, and advanced counts only a key re-filed at a new status. Six of the seven run departments therefore read 0.00, and data-eng-templates ranked TOP prune candidate with 16 filed keys, a public repo and a live store page behind it. A key FILED now counts: director.throughput() is the one copy of the formula and director.idle() the one prune predicate. - A folder broke count included runs of rows that are WALLS. The same attended row the ranking sinks at rank 9 put its folder back at rank 2. - A folder broke count also included a break an earlier pass cured. PRESS-ROOM ranked 1 on a blocked run of 18:35:21Z whose successor at 21:04:25Z ended done. row_runs now reports cured, the newest run in the window ending done. Department scores after the fix, keys per run-hour: hermes-smb-agency 12.00, agent-skills-pack 6.47, data-eng-templates 5.98, MARKETING-BRANDING 5.88, ai-video-ads 2.46, niche-content-engine 1.00, PRESS-ROOM 0.00. The top candidate is now PRESS-ROOM, 4 runs and no ledger of its own, so nothing it publishes is ever measured. That is the next experiment. Walls that stand, and each needs a person: Replicate credit, Stripe Connect payouts, the money-side Reddit and X sessions on :9223, and x402-data-endpoint serving GET /v1/daily before niche-content-engine/01-channel-and-daily-digest.md can unpark. That last one needs no person: x402-data-endpoint/00-ledger.md and 01-prototype-data-pipeline.md carry no attended: and have never run. ## Pass 6 Findings Two things landed: the ranking stopped lying at random, and PRESS-ROOM became measurable. - research-lab hypotheses read prompt-supervise's stdout and ignored its exit code. A probe failure inside prompt-supervise dies with exit 4 and prints nothing, so the folder came back with no wall, every ATTENDED row in it counted as a break, and the folder climbed to rank 1 with no door on it. - Measured 2026-09-06: two rankings minutes apart, no edit between them, put data-eng-templates at rank 1 broke 1 of 4 runs and then at rank 4 `score 5.98. Its 03-launch-posts.md carries attended:` for the money-side Reddit and X sessions, so a pass acting on the first ranking would have run an experiment against a wall a person owns. - supervise() now dies on any exit outside 0 and 1, naming the folder, the code and prompt-supervise's stderr. Exit 0 is a table and exit 1 is a folder holding no row but the asking one, which is the lab's own case. The settled experiment is press-room-own-ledger, kept: before 0.00, after 236.00, one run, zero broke. - The publish body gained one bullet: `ledger set "$(dirname "$PROMPT_NOTE")/work-ledger.tsv" <record key> <status>` files each record the pass wrote, one key per record. - The run filed 118 keys into a work-ledger.tsv PRESS-ROOM never had, so a department that had published 49 rows over 4 runs now reports a number. - 236.00 keys per run-hour sits two orders above every other department, and a press-room key is one ledger set rather than a shipped artifact. The number reads PRESS-ROOM against itself and against nobody else. - All 118 keys are stuck: filed once and never re-filed at a new status. That is the next candidate this folder offers. - The run found the self-reference the bullet creates: PRESS-ROOM's own work-ledger.tsv matches newsroom ledgers' glob under OPS, so `newsroom owed` sees the press room as a source and the gate fails. Self-ack in the same pass is `ledger unsent <file> press-room | cut -f2 | ledger sent <file> press-room -`, and the row's learnings carry it. ## Pass 7 Findings The ranking picked the department by ASCII, and a department that had never run was invisible to it. Two commits closed both, and the first never-run department then ran. - research-lab hypotheses broke a tie inside one kind on the PATH. Over AUTONOMOUS-WORK the six score candidates came out alphabetically: MARKETING-BRANDING 11.02 stood at rank 1 and niche-content-engine 2.00 at rank 6. Every candidate now carries a room number and the sort reads it before the path, so the department producing least ranks first. - ops emitted a score candidate only where the score was a number, and a folder with no run in the window scores math.inf. Four departments were therefore absent from every ranking this loop took: self-hosting-video-course, spanish-open-data-mcp, science-youtube-manim and x402-data-endpoint. A folder holding rows and no run now ranks never-run, above every scored department. - The never-run tie breaks on the row count, and that count includes a row a person must attend. self-hosting-video-course ranks first on four rows and holds one row a pass can run unattended. - The settled experiment is x402-first-run, kept: x402-data-endpoint went from new to 4.00 keys per run-hour on one run that ended done at 2026-09-06T22:52:50Z. The edit told the row that a missing spend.jsonl or revenue.jsonl counts as zero, because the first pass runs ledger.py before either file exists. - That run chained on to 01-prototype-data-pipeline.md, and a served GET /v1/daily is what unparks niche-content-engine/01-channel-and-daily-digest.md. ## Pass 8 Findings The department that had never run produced, and the lab's own verdict line stopped saying two different things in the same words. - self-hosting-video-course ranked 1 as never-run, four rows and no run in 30 days. Its 00-ledger.md is the only row carrying no attended:, so it is the only one a pass can try. The edit told it that a missing spend.jsonl or revenue.jsonl counts as zero, the same first-pass trap x402-data-endpoint paid for at pass 7. - The run ended done at 2026-09-06T23:09:34Z, its finish check passed, and the folder scores 4.00 keys per run-hour. It wrote two facts of its own into the row: youtube.json names the free lessons' video IDs, and revenue reads 0 while head -c 7 ~/.config/stripe prints sk_test. - research-lab verdict keeps ANY experiment over a folder that never ran, because math.inf is a baseline every comparison passes. The kept line read exactly like one earned by a raised score. It now carries `before was new, so the keep rests on the run not breaking, never on a raised score`. - An experiment's label was the row's basename alone, and every department's first row is 00-ledger.md. research-lab report printed two identical - 00-ledger: lines for two departments. The label is now <folder>/<row>. Five sibling ledger rows still spell hermes cron create ... --deliver telegram, which registers against the home bot that blocked the user, so Monday's post fails: spanish-open-data-mcp, science-youtube-manim, hermes-smb-agency, public-domain-audiobooks and self-hosting-video-course. Four departments already spell telegram:-1004404879807, so the fix is known and unapplied. x402-data-endpoint/01-prototype-data-pipeline.md was still LIVE at the end of this pass, two hours in, and 100.120.110.117:8402 still refuses. ## Pass 9 Findings Both never-run departments ran a first pass, and the same one-line cure carried both. - spanish-open-data-mcp and science-youtube-manim had 4 and 3 rows and no run in 30 days. Each 00-ledger.md told the pass to sum spend.jsonl and revenue.jsonl, and the first pass runs ledger.py before either file exists. - One bullet on each row says a missing file counts as zero. Both runs ended done, both finish checks passed, and both chained to their next row. - Four departments have now taken that cure: x402-data-endpoint, self-hosting-video-course, spanish-open-data-mcp and science-youtube-manim. Every one scored new before and 4.00 after. - Each first run also repaired its own Telegram target without being asked. Both ledger rows now spell telegram:-1004404879807, the @AutomateItAll channel, because the bare telegram home DM bot is blocked. - No never-run department is left under AUTONOMOUS-WORK. The ranking's top candidate is now niche-content-engine at 2.00 keys per run-hour, and it is parked behind x402-data-endpoint serving GET /v1/daily. ## Pass 10 Findings A row's HANDOFF outranks its BODY, so an experiment edited into the body of a row that already ran is never read. - research-lab try snapshotted public-domain-audiobooks/01-voice-and-first-book.md and the pass edited one rule: book one is ONE work, not the whole 12-story Bécquer volume, because 47 508 words cost more GPU render hours than a loop has passes. - prompt-run --again then handed the fresh session the previous run's handoff as handoff-previous.md. Its pass 1 adopted that Left, in Order whole and spent all 12 passes polling render_all.sh for 12 of 12 wavs. - The run capped again at 0.00 against a baseline of 1.08, so verdict reverted an edit nothing had measured. - prompt-run --clear is the answer and already existed. research-lab try now globs the row's own 📊 drawer and prints --clear where any earlier chain.json stands, with the reason on stderr. - The row's done-when gate also fails on a second count nobody can cure from the row: its composed tree check reads ~/git/autonomous-work, one git repo that ten departments share, and agent-skills-pack, site/, science-youtube-manim and five more leave it dirty. An upload would still gate red. ## Pass 11 Findings - course-telegram-target REVERTED on the number, kept in the world. The edit pinned telegram:-1004404879807 on self-hosting-video-course/00-ledger.md, the run ended done with gate: pass, and hermes cron list now shows course-ledger delivering to the @AutomateItAll channel where it delivered to the blocked home bot before. The score fell 4.00 to 2.00 on a key rate the edit never touched, so verdict restored the body. - FOUR CRON JOBS STILL DELIVER TO THE BLOCKED HOME BOT: hermes-agency-outreach, hermes-agency-replies, science-channel-ledger, datos-collect. Eleven deliver to the channel id and work. hermes cron edit <id> --deliver telegram:-1004404879807 repoints one. - THE ROOT CAUSE WAS THE BRIEF, and pass 11 fixed it: AUTONOMOUS-WORK/Every row says.md told every row to spell hermes cron create … --deliver telegram and hermes send --to telegram. Both lines now name the channel id and the cost of the bare target. - public-domain-audiobooks/01-voice-and-first-book.md CAPS ON A GPU RENDER, not on a prompt defect: render_all.sh had 2 of 12 chapters at 00:27Z with about 3.6 hours left, and each loop pass polled it for 90 seconds and returned continue, so twelve passes covered eighteen minutes. The cure shipped in the shared loop-🔁.md at 00:14:01Z, two minutes after the run it describes, and no run has read it yet. - A LAB PASS REACHES 5 OF THE 6 RUN SLOTS, because its own session is one of the six PROMPT_RUN_CAP counts. Under AUTONOMOUS-WORK the director plus three department runs filled those five for 45 minutes, and prompt-run refused the experiment's launch 45 times. - TWO REPO FIXES LANDED: research-lab try warns when the box is at the cap before it snapshots a row, and research-lab verdict copies the live body and drawers to experiments/<slug>/after/ before a revert overwrites them. The revert had been destroying the facts the run itself wrote into the row. ## Pass 12 Findings Pass 12 was the last pass of the loop. It settled no scored experiment and closed the bare-telegram bug across the whole department folder instead. - hermes send --to telegram IS DEAD, and pass 12 proved it rather than inferring it: the probe exits 1 with Telegram send failed: Forbidden: bot was blocked by the user. Every department row ended its pass by posting its report through that target, so no department report has arrived since the operator blocked the bot. - THE HOME CHANNEL IS A DM, 1370010055 in ~/.hermes/channel_directory.json, and hermes REWRITES that file itself. No config edit cures the bare target durably, so the cure has to be spelled in the rows. - SIXTEEN OF SIXTEEN CRON JOBS now deliver to telegram:-1004404879807. Pass 12 repointed the last four, hermes-agency-outreach, hermes-agency-replies, science-channel-ledger and datos-collect, with hermes cron edit <id> --deliver telegram:-1004404879807. - FORTY LIVE ROWS now name the channel too, from one portable sed -i.p12bak sweep over --to telegram\ and --deliver telegram . The RESEARCH-LAB/experiments/*/before/ snapshots keep the bare target on purpose, because they are the baselines a verdict` restores. - A sed -i SWEEP OVER A PROMPT TREE BREAKS EVERY SYMLINK IT TOUCHES. The sweep replaced 69 -📊/ drawer symlinks with regular copies, this run's own handoff.md among them. The state files behind them survived untouched, and one ln -s pass rebuilt all 69 from a sibling symlink's target directory. - THE SUITE RAN WHOLE: bash bin/run_tests.sh --all --worktree at 5c9be4df6, 53 minutes, run_tests: done (rc=1, 20 failed of 393, 1 SKIPPED THEMSELVES). - NONE OF THE 20 REDS BELONGS TO THIS LOOP. git log origin/main..HEAD is empty, so the measured sha is main itself, and suite-baseline check --at cd0a3dbf5 answers UNATTRIBUTED on all five prompt-tier reds it was asked. A co-tenant's core-tier run at a descendant sha failed exactly the same 6 core lanes. - THREE REDS ARE CODE DEFECTS A NEXT RUN CAN FIX: test_module_consistency.sh (module notebook-git in no bundle, prose_budget_write_verdict defined twice), the three test_mirror_export*.sh lanes (the staged mirror calls bin/tests/test_newsroom.sh and bin/tests/test_site.sh and ships neither), and test_treefolderargs.sh (rebuild.lua's emit drops fn_args, so a drawer stops being drawn silently).

Act now - The operator: review the numbers on Monday 08:00 ledger cron, @AutomateItAll — NUMBER-science-youtube-manim-ledger-cron - The operator: review the numbers on First ledger post, spend 0.00, fresh channel — NUMBER-science-youtube-manim-ledger-post-2026-09-07 - The operator: open the launch page for Episode 1: The Trick That Shrinks Every File You Send — LAUNCH-science-youtube-manim-episode-0001 - The operator: open the launch page for Short 1: 88 bits become 21 — LAUNCH-science-youtube-manim-short-0001 - The operator: review the numbers on Niche: Math of Everyday Things — NUMBER-science-youtube-manim-niche-picked - The operator: open the launch page for Episode 2: The Traffic Jam With No Cause — LAUNCH-science-youtube-manim-episode-0002 - The operator: open the launch page for Short 2: 0 causes, 1 jam — LAUNCH-science-youtube-manim-short-0002 - The operator: open the remaining 6 record(s) on the board

Act now - The operator: review the numbers on Monday 08:00 ledger cron, @AutomateItAll — NUMBER-self-hosting-video-course-ledger-cron - The operator: review the numbers on First ledger post, spend 0.00, 0 sales — NUMBER-self-hosting-video-course-ledger-post-2026-09-07 - The operator: review the numbers on Ledger repost, stripe test-key guard, 0 sales — NUMBER-self-hosting-video-course-ledger-post-2026-09-07-testkey-guard

Act now - The operator: review the numbers on Monday 08:00 ledger cron, @AutomateItAll — NUMBER-spanish-open-data-mcp-ledger-cron - The operator: review the numbers on First ledger post — NUMBER-spanish-open-data-mcp-ledger-post-2026-09-07 - The operator: review the numbers on boe-collector — NUMBER-spanish-open-data-mcp-boe-collector - The operator: review the numbers on tenders-collector — NUMBER-spanish-open-data-mcp-tenders-collector - The operator: review the numbers on ine-collector — NUMBER-spanish-open-data-mcp-ine-collector - The operator: review the numbers on mcp-server — NUMBER-spanish-open-data-mcp-mcp-server - The operator: open the launch page for rest-mirror — LAUNCH-spanish-open-data-mcp-rest-mirror - The operator: open the remaining 19 record(s) on the board

Act now - The operator: review the numbers on first-order — NUMBER-spreadsheet-cleanup-first-order - The operator: nothing further, this folder filed one line

Act now - The operator: review the numbers on ledger — NUMBER-todo-app-ledger - The operator: open the launch page for public-copy — LAUNCH-todo-app-public-copy - The operator: nothing further, this folder filed one line

Act now - The operator: clear the wall blocking activity-pack — WALL-travel-printables-activity-pack - The operator: open the launch page for activity-pack-sample-a4 — LAUNCH-travel-printables-activity-pack-sample-a4 - The operator: open the launch page for activity-pack-sample-letter — LAUNCH-travel-printables-activity-pack-sample-letter

Act now - The operator: review the numbers on Monday 08:00 ledger cron, @AutomateItAll — NUMBER-x402-data-endpoint-ledger-cron - The operator: review the numbers on First ledger post, spend 0.00 EUR, 0 payers — NUMBER-x402-data-endpoint-ledger-post-2026-09-07 - The operator: review the numbers on pipeline-cron — NUMBER-x402-data-endpoint-pipeline-cron - The operator: review the numbers on sample-output — NUMBER-x402-data-endpoint-sample-output - The operator: review the numbers on 402 paywall on /v1/daily — NUMBER-x402-data-endpoint-x402-paywall - The operator: review the numbers on MCP server for Hermes — NUMBER-x402-data-endpoint-mcp-server - The operator: review the numbers on systemd user unit on pueblo:8420 — NUMBER-x402-data-endpoint-x402-origin-service - The operator: open the remaining 10 record(s) on the board