Episode 09 --- Harmony
Working YouTube title:
I Found a Real Bug in My Own Product, Live
Thumbnail thought:
it said it worked. it lied.
Target runtime: ~7:45
[0:00--0:20] COLD OPEN
[CAMERA --- Episode 08's closing question, on screen: "what happens when nobody's around to run muse resolve?"]
GABRIEL:
That's Harmony's whole job. Let's see how well it actually does that job.
[beat]
Spoiler: this episode doesn't go the way I planned it.
[TITLE CARD --- fast]
HARMONY
[Music enters.]
[0:20--1:00] THE PROBLEM
[CAMERA]
Two agents, working independently, hit the exact same kind of conflict twice. Nobody's available to make the call the second time.
[beat]
Should the system really ask a human the same question twice?
[ON SCREEN]
FOUR TIERS: POLICY → EXACT REPLAY → SEMANTIC → ESCALATE
That's Harmony's answer, straight from the CLI's own description.
[1:00--1:50] RECORDING HAPPENS WHETHER YOU ASK OR NOT
[TERMINAL --- the exact conflict from Episode 08]
$ muse merge feature/bob
CONFLICT (both modified): calc.py::add
$ muse harmony list
{ "patterns": [{ "path": "calc.py::add", "resolution_count": 0, "recorded_by": "auto_apply" }] }
GABRIEL VO:
Nobody ran muse harmony record. The conflict happening at all was
enough --- Harmony already has a pattern for it, with zero
resolutions, waiting to learn.
[1:50--2:30] TEACHING IT
[TERMINAL]
$ muse resolve calc.py::add
$ muse commit -m "resolved"
$ muse harmony show sha256:8b49f9...
{
"strategy": "manual",
"confidence": 1.0,
"human_verified": true,
"rationale": "User manually resolved conflict in 'calc.py::add' via muse resolve"
}
GABRIEL:
The ordinary act of resolving a conflict and committing is teaching Harmony. Full provenance, no extra step.
[2:30--3:10] THE SAME CONFLICT, AGAIN
[TERMINAL --- fresh branches, identical edits, same base]
$ muse merge feature/eve
✔ [harmony] auto-resolved: calc.py::add
Merge made by the three-way strategy.
GABRIEL VO:
No conflict markers. A green checkmark. Exactly what teaching it was supposed to buy us.
[beat --- pause, look uncertain on camera]
Let's actually check the file.
[3:10--4:00] THE MOMENT
[TERMINAL]
$ cat calc.py
def add(a, b):
return a + b + 1
GABRIEL --- direct to camera:
That's wrong.
[beat]
I taught it +101. The resolution I recorded --- confirmed, I checked
the stored object directly, it really does say +101 --- never made
it into this file. This is just +1. Untouched. "Ours."
[CAMERA]
It printed a checkmark. It said auto-resolved. It incremented the resolution's usage count. And it threw away the actual answer.
[4:00--4:45] WHY I'M NOT CUTTING THIS
[CAMERA]
Every episode this season, I told you I'd show real output, including when it broke on me --- Episode 06's plugin bugs, Episode 07 and 08's smaller inconsistencies.
[beat]
Those were rough edges. This is different. This is the exact feature this episode is named after, silently doing the opposite of what it told me it did.
[ON SCREEN]
A CHECKMARK IS A CLAIM. THIS ONE WAS FALSE.
I filed it the way I've filed everything else this season --- for real, in the actual tracker, with the actual repro steps.
[4:45--5:40] WHAT ACTUALLY WORKS
[CAMERA]
I want to be precise here, not just dunk on my own project. The memory half of Harmony is solid. I verified each of these for real:
[TERMINAL --- rapid real commands]
$ muse harmony policy-add --policy-id ... --scope repo --action prefer-ours
$ muse harmony policy-list
$ muse harmony escalate --reason "..." <pattern-id>
$ muse harmony escalations
$ muse harmony audit
GABRIEL VO:
Recording, storage, provenance, policies, escalation tracking, and an audit log that genuinely logs policy and escalation actions --- all real, all verified.
[beat]
The one broken link is the last step: taking a correctly-stored resolution and actually writing it into the merged file. That's it. That's the bug. Not the architecture --- one write.
[5:40--6:15] THE POINT
[CAMERA]
A resolution system is only as trustworthy as its worst silent failure.
[beat]
I'd rather you learn that Harmony has a real bug in it from me, on camera, than discover it yourself in a repo where it actually mattered.
[ON SCREEN]
EVERY CLAIM IN THIS SERIES IS SUPPOSED TO BE CHECKABLE. THIS ONE JUST FAILED THE CHECK.
[6:15--6:50] OUT
[TERMINAL --- back to a clean repo, muse domain-info --domain code]
GABRIEL VO:
We've spent four episodes on diff, merge, and resolution. Let's step back and look at everything the code domain can tell you about a codebase when nothing's even conflicting.
[beat]
Next episode: blast radius, hotspots, gravity, coupling. Muse understanding code, not just versioning it.
[CAMERA.]
[CUT TO BLACK]
musehub.ai
Production Notes
Episode 09 is the season's credibility test. Every previous episode asked the audience to trust real, unedited output; this episode is where that trust either survives contact with a real, serious bug or the whole season's premise looks like marketing the moment something actually breaks. Play it straight. Do not soften the bug, and do not oversell how bad it is either --- it's a real data-integrity issue in one specific code path, not evidence the whole product is broken.
Opening
Foreshadow it immediately: "this episode doesn't go the way I planned it." Don't bury the discovery at the end for a twist --- the audience should watch the whole demo already knowing something's going to go wrong, which makes the "let's actually check the file" beat land as confirmation, not a cheap surprise.
The Bug Must Be Shown, Not Just Described
Every claim in this beat is real and was verified multiple times before scripting: the false success message, the incremented `applied_count`, the stored `outcome_blob` genuinely containing the taught value, and the merged file genuinely not containing it. Show all four pieces of evidence on screen. A viewer who can't verify the claim themselves shouldn't have to take Gabriel's word for a bug this significant --- the same standard the season holds Muse's own claims to.
Filed for Real
`muse hub issue create` against the actual tracker, on camera, with the real repro steps, is the appropriate way to close this beat --- it demonstrates the same "receipts, not promises" standard applied to the show's own findings, not just to Muse's marketing claims.
Don't Skip "What Actually Works"
Without this section the episode reads as a takedown. The point isn't that Harmony is broken --- it's that one specific, serious link in an otherwise sound chain needs fixing, and the season's commitment to accuracy means saying so precisely rather than either hiding it or exaggerating it.
The Seed
The viewer arrives thinking:
Okay, this whole season has been suspiciously smooth. What happens when something actually goes wrong?
They should leave thinking:
He really will tell me when it's broken. So when he says something works... it probably works.
That's not a question Episode 10 needs to answer directly --- it's the trust the rest of the season gets to spend.