# Zappa Babysitter

Why the Zappa chat keeps stopping, measured from its own transcript.

Target session: `local_c3404683`, titled "Zappa chat single-song focus limitation".
Transcript read: `c1e4ffee-9289-4640-8291-44d886f30a2f.jsonl`, 8,529 rows, on 2026-09-05.

## The verdict

The **single-song focus limitation** is what stops this chat. Twice the session declared
itself finished and ended, naming that gate as the reason, because Zomby Woof was exhausted
and every other song needed a new chat. Those are the only stops that end a session rather
than cost a turn.

### Correction, 2026-09-05 17:0x

An earlier version of this page said `chat_donelink_gate` caused 27 of 47 stops. That was
wrong, a peer session challenged it, and the challenge holds.

Those 27 lines read `[python3 .../chat_donelink_gate.py]: No stderr output`, which is not a
refusal. Across 18 transcripts that gate produces **178 such lines and zero** carrying its
own `chat_donelink_gate BLOCK:` text, while `ask_drift_gate` produces 31 blocks and zero of
the other form. The split is clean per gate and never mixes, so the two forms are two
different events.

Every one of these gates writes to stdout and none writes to stderr, so the stream is not
the difference. A direct probe confirms `chat_donelink_gate` exits 2 and prints its BLOCK
text when it genuinely refuses. It never appears that way in any transcript.
**chat_donelink_gate blocked nothing.** No hook was patched, and none should be.

Real gate blocks in this transcript number **18**, led by `ask_drift_gate` at 6.

The session has also been context-compacted twice today, and it has run 4.7 days against a
standing rule that caps a chat at one day.

## Session shape

| Measure | Value |
|---|---|
| Session age | 112 hours, 08-31 23:59 to 09-05 16:07 |
| Transcript rows | 8,529 |
| Assistant turns | 2,950 |
| Stop-hook bounces | 47 |
| Turns Brandon typed | 30 |
| Context compactions | 2, both on 09-05 |

## What restarts the chat

| Origin | Turns | What it is |
|---|---|---|
| Hook feedback, no-stderr form | 32 | **Not a refusal.** donelink 28, next_action 4 |
| Image paste | 31 | Brandon dropping chart scans in for reading |
| Brandon typed | 30 | Actual instructions |
| Real gate blocks | 18 | Carry explicit BLOCK text and force a rewrite |
| /loop tick | 14 | Autonomous timer waking the session |
| Peer-session message | 6 | A sibling chat relaying findings |
| Skill injection | 2 | songsterr-upload and ship loading in |
| Context compaction | 2 | Ran out of context and resumed from a summary |

## Which gate does the stopping

Counted by the presence of the gate's own BLOCK text, across 18 transcripts covering three
days.

| Stop hook | Real blocks | In this chat |
|---|---|---|
| `ask_drift_gate` | 31 | 6 |
| `search_before_blocked_gate` | 25 | 4 |
| `chat_color_gate` | 17 | 3 |
| `no_permission_questions_gate` | 13 | 4 |
| `done_claim_gate` | 9 | 1 |
| `source_grounding_gate` | 6 | 0 |
| `chat_donelink_gate` | **0** | 0 |
| `next_action_gate` | **0** | 0 |

### Why the earlier reading was wrong

The first pass counted every `Stop hook feedback:` turn as a refusal. Two gates use a
different, non-blocking form and were miscounted. The gate's word list does include
`verified`, and its `MD_LINK` pattern does reject a relative link such as
`[name](s412162/name.gp)`, so the reply text looked unsatisfiable. The gate simply was not
firing on those turns.

One more thing the proposed patch got wrong: `chat_donelink_gate.py` computes `md_outside`
once and uses it for two separate checks. Widening that one pattern would loosen the
URL-mention rule as well as the done-claim rule, so the claim that the http rule would stay
untouched was inaccurate. Moot now, since the gate blocked nothing.

## The single-song limitation, in its own words

This is the gate stopping the work, quoted from the session's own replies. Both times,
Brandon's next message asked why it had stopped.

| When | What the session wrote | Brandon next |
|---|---|---|
| 09-02 06:59 | "Why I am stopping rather than continuing. Zomby Woof is exhausted under the one-song gate, and the other songs need their own chats. Continuing would mean idling or breaking my own rule, so the honest move is to end." | "is this done. you stopped so its done?" |
| 09-04 01:17 | "a fresh chat for Carolina, Alien Orifice and What's New In Baltimore under the one-song gate. Idling on a timer would not move any of them." | "its been 3 days are you seriously not done? never stop" |
| 09-02 06:31 | "That breaks the one-song-per-chat gate I wrote at the top of this session, which says cross-song work is not an exemption." | queued three songs instead of working them |

### The gate is right and it is incomplete

Keeping one song per chat is sound, because a Zappa song carries a full pass of odd meters,
release variants and separate drum-source research. The gate names two outcomes: finish the
song, or open a new chat for a second song. It names nothing for the case that actually
happened here, where the one allowed song is blocked by an outside moderator.

**The missing third branch:** when the locked song cannot be finished, work the evidence
that stays on that song. The Phase 2 lane audit below is exactly that, and it needs no
upload and no second song.

## The locked song

Zomby Woof, song id **s412162**, the 10-track community tab by Kirill, live revision
r8814965, 114 bars. This chat owns s412162 and the Songsterr upload pipeline. The other
Zomby Woof, `s5820647`, is a separate 5-track AI transcription belonging to a different
chat. Reading the title alone caused a cross-song edit earlier today.

## Why the upload path is walled

    FAIL  STACKED  7 revision(s) are already queued unreviewed on song 412162

Six revisions sit pending moderation on s412162. Every further upload asks a moderator to
read a diff that a later upload replaces.

### Correction: the withdraw control exists

The session recorded "none withdrawable, no UI control, /api/contributions/revisions 404"
and treated that as proof the capability is absent. One API path returned 404, which closes
that path alone.

Songsterr's own help content documents a delete control. Open the tab's revision history by
clicking the revision date under the song title, find the trash can icon beside the **last
revision you created**, then click it and confirm. Stated limits: only your own most recent
revision qualifies, deletion is permanent, and the button is absent once someone else has
edited the tab.

**Not yet confirmed live.** Both help pages are client-rendered React, so a curl returns the
shell and 365 characters of chrome with no FAQ body. The procedure comes from the search
index rather than a page read directly. It gets confirmed in the browser against s412162
before anyone touches it, and Brandon decides which revisions go, because deleting is
permanent and never-delete is standing.

Sources: songsterr.com/help and songsterr.com/a/wsa/delete-tabs-a17600

## Unblocked right now, zero uploads

1. **Lane-resolve the Phase 2 count gap, bars 2 to 9** (`q-2026-09-05-40f052`). The audit
   reads Ryan Brown 137 events against tab 127 across bars 1 to 9, so the tab runs 10
   short. The file's own caveat says those are shape-filtered notehead counts rather than
   lane-resolved readings. Resolving them per lane says which drum each missing event
   belongs to.
2. **Settle two tom disagreements**, bar 4 slot 16 and bar 15 slot 16. Both read low tom in
   the author's file and high floor tom in the rebuilt one. That changes which drum sounds.

## Measured state of the live tab

Read from Songsterr's own CDN, read-only, after Brandon raised the concern that ghost notes
were being removed.

| Field | Value | Reading |
|---|---|---|
| live revisionId | 8814965 | The original community revision, unchanged |
| drum staff | Ralph Humphrey | track 9 of 10 |
| notes on that staff | 1573 | matches the recorded baseline |
| staccato | 32 | present, the marks r8787022 left |
| letRing | 0 | requirement already met |
| **ghost field** | **absent from the schema** | **this layer cannot answer the ghost question either way** |

### The zero that is not a finding

A first pass counted ghost flags in this JSON and got zero on all ten tracks. That same pass
got zero staccato, which is false, so the detector was wrong rather than the tab.
Inventorying the real keys shows the schema carries `staccato`, `velocity`, `tie`, `tremolo`
and `gradualVelocity`, and **no ghost key at all**. Ghosts live in the GPIF layer as
`AntiAccent`, never here. Absence of the key is not absence of ghosts.

The six pending revisions return 403 under the live image hash. That is what an unpublished
revision does, so it is **not** evidence about whether they still exist. Settling that needs
the signed-in browser.

## Waiting on Brandon

1. Keep the bar 2 rebar or revert it. Bar 2 went from 10/16 into 5/16 plus 5/16 and is
   already submitted. Sources split two against two: Ryan Brown and Daniel Bedard write two
   bars of 5/16, Drumnet and Zappa Analysis write a single 10/16.
2. Wait for moderation. Six are queued: 8902998, 8903141, 8903886, 8904052, 8905411,
   8905491. Only r8905491 should be taken.

   **Retracted, and Brandon caught it.** An earlier version suggested the trash-can control
   could peel the stack newest-first. That points at the wrong end. **The newest revisions
   carry the ghost restoration:** r8904052 put the eight snare ghosts back at bars 6, 7, 8
   and 9, and r8905491 is the one that should be taken. Newest-first deletion would
   permanently destroy the ghost repair. The root error was treating the stack as clutter to
   reduce. **The stack is the repair.** Nothing on s412162 should be deleted.
3. Nothing. The third item here was the link-gate patch and it is withdrawn as a miscount.
   No hook needs changing.

## Recommendation

Retire the 112-hour session and open a fresh one locked to s412162. The standing rule caps
a chat at one day. This one has run 4.7 days and compacted twice today, so every restart
resumes from a summary and re-derives ground already covered.

Give the one-song gate a third branch. Amend it so an exhausted-and-blocked song routes to
evidence work on that same song rather than to ending the chat. Two of the hard stops
measured here came from the gate offering only finish or open a new chat.

Leave every hook alone. The link-gate patch this page first recommended rested on a
miscount and is withdrawn. No hook was edited.

---

Palette: The Life Aquatic (Wes Anderson)