A job that has to name a Muse Code model before it has a Meta credential has one offline command for that question. muse model-profile show prints a resolution for the id you pass and for one effort tier. On Muse Code 1.4.24 the command exits 0 for every non-empty id tried on this host, including not-a-model. The field that separates a known string from that fallback is plugin_text_variant. It is stock for three ids. It is exact-minimal-literal-tests for the rest, including documented catalog ids that Muse Code is supposed to run.
Those three ids are muse-spark-1.2, muse-spark-1.3, and muse-spark-1.3-contributor. The comparison used the same binary, the same effort flag, and a config home created for the run. A case change of the first id, Muse-Spark-1.2, left the stock row and joined the fallback. The command echoes the string you typed. It does not fold case, and it does not answer “is this id in the catalog” with a failed exit.
Exit status is the wrong field for that question. measured_tuning_cell is also the wrong field if you read it once. On muse-spark-1.3, --effort high prints the cell as false and --effort max prints it as true, while plugin_text_variant stays stock. At max, an unknown id and the recommended Spark id both show a measured cell. The plugin-text line is what still differs.
The binary and the command
The binary was Muse Code 1.4.24, release 1.4.24-R9131.1, on Linux x86_64, on 20 November 2026. muse --version printed Muse Code 1.4.24 (1.4.24-R9131.1). The versioned file beside the launcher was 471370072 bytes. Its SHA-256 was d752c45dce079ee29e118804b091d81ac2b1b89b67f317b4a162fa1af1206465, which matches the x86_linux entry in the public manifest for that release. MUSE_NO_AUTO_UPDATE=1 was set so the launcher’s update check was not part of the result. Each command used a new config home, data home, state home, and cache home. No muse login was run. No META_API_KEY was set. No prompt was sent to a model.
The help line for the argument says the value is a catalog model id, and the example is muse-spark-1.3. That example is one of the three stock ids. The word catalog in the help sentence is not enforced by the parser. An empty argument is rejected. An unknown effort tier is rejected. A nonsense model id is not.
The empty invocation was muse model-profile show --effort high. It exited 2 and printed error: a model id is required, then the usage text. muse model-profile show muse-spark-1.3 --effort bogus also exited 2. The error named the tier and listed none, minimal, low, medium, high, xhigh, max, and ultra. The tier vocabulary is closed. The model vocabulary is not.
A successful report for the recommended id at the default effort is short. The model line echoes muse-spark-1.3. The effort line says high. base_instructions_variant is full. measured_tuning_cell is false. plugin_text_variant is stock. default_reasoning_effort is high. Under a re-resolution heading labeled ADR 42674 D7, the same report says the base instructions and the measured cell are perRequest, the plugin text is startupFixed, and the default effort selects the session tier so a later re-resolution would fight a /effort change. The plugin-text note says the reminder duty rides the reminder roster, which that rule keeps startup-fixed, and that skill bodies are fixed when the session’s skills load.
The diff from that report to muse-wonder-1.0 at the same effort is three lines: the echoed model, measured_tuning_cell from false to true, and plugin_text_variant from stock to exact-minimal-literal-tests. The diff from muse-wonder-1.0 to not-a-model is the echoed model line only. Wonder, which the models page lists as a catalog id and which the coding-agent guides name for long-running work, resolves here to the same profile as a string that is not a model id.
What stays put
Four lines did not move across this set. base_instructions_variant stayed full. default_reasoning_effort stayed high even when the requested effort was none, max, or ultra. The re-resolution block stayed word for word. That block is why the cell above it can change with the flag you pass. It does not say why an unknown id receives a measured cell, or why its plugin-text name is more specific than stock.
The value exact-minimal-literal-tests is a label in this report. This run did not load a plugin, did not execute those tests, and did not ask a model to follow them. Read the string as a selector the resolver printed. Do not read it as a test suite that ran.
The same fallback, measured cell true and plugin text exact-minimal-literal-tests, was printed for muse-spark-1.1, muse-spark-1.2-contributor, muse-ember-1.0, muse-image-1.0, muse-voice-transcribe-1.0, sam-3.1, muse-spark, muse-spark-1.4, Muse-Spark-1.2, and not-a-model, each at --effort high. The stock row was only the three strings named above. muse-spark-1.2 stayed stock, with a false measured cell, at high, max, and ultra.
The published lists do not match the stock row
The models page, fetched the same evening, lists muse-wonder-1.0, muse-spark-1.3, muse-spark-1.2, muse-spark-1.1, muse-ember-1.0, muse-image-1.0, muse-voice-transcribe-1.0, and sam-3.1 as model ids. The Spark page lists contributor variants muse-spark-1.3-contributor and muse-spark-1.2-contributor. It says every Spark version except 1.1 is available in Muse Code. It also says muse-spark-1.3 is recommended for new work. The Muse Code configuration page still says the default model is muse-spark-1.2.
Both of those defaults sit in the stock row, so a job that pins either documented Muse Code default will see stock on this binary. A job that follows the Spark page to muse-spark-1.2-contributor, or the coding-agent guides to muse-wonder-1.0, will see the same plugin text as not-a-model. A job that still names muse-spark-1.1 will see that fallback too. The Spark page says 1.1 is not available in Muse Code. model-profile does not say unavailable. It exits 0 and prints a profile. Absence from the stock row is therefore not a Muse Code availability bit. Wonder is documented as a Muse Code model, and Wonder is also absent from the stock row.
The help example and the configuration default can both be true at once. muse-spark-1.3 and muse-spark-1.2 are both stock. The command does not tell you which of those the product page currently recommends. You still read the page for that. What the command can tell you, on this build, is whether the string is one of the three that select stock plugin text.
A launcher that maps “documented Muse Code id” onto “stock profile” will mis-file Wonder and the 1.2 contributor id. A launcher that maps “not stock” onto “do not call” will refuse Wonder, which the models page and the Muse Code docs both treat as a real model, and will also refuse a typo, which is the one refusal that happens to be right. The profile is a local selector. It is not the availability table.
Effort max is a cell flip, not a refusal
The Spark page says muse-spark-1.3 supports max, and that max is not available with muse-spark-1.3-contributor. The configuration page says the Meta provider does not accept none, and that ultra maps to the provider’s highest supported level when a provider does not accept ultra as its own tier.
model-profile did not apply those availability rules. For muse-spark-1.3 and for muse-spark-1.3-contributor, the diff from --effort high to --effort max was two lines. The effort line changed from high to max. measured_tuning_cell changed from false to true. plugin_text_variant stayed stock. The contributor id was not refused. The two max reports differ only in the echoed model line.
--effort ultra did not perform that flip. On both 1.3 ids it left the measured cell at false and printed effort: ultra. If ultra were only an alias of max inside this command, the cell would have matched the max report. It did not. A wrapper that asks for ultra and then reads the measured cell is not reading the max resolution.
--effort none was accepted for muse-spark-1.3 and printed effort: none with the stock, false-cell profile. The provider rule that Meta rejects none is not this command’s rule. The command’s rule is the enum in the error for bogus. none is in that enum. Whether a later Meta request would reject it was not tested. This host had no Meta credential, and this command does not send one.
At max, not-a-model still printed plugin_text_variant exact-minimal-literal-tests and measured_tuning_cell true. The measured cell is then shared by the stock id and the unknown id. A gate that requires measured_tuning_cell true will accept the typo at every effort tested, and will accept muse-spark-1.3 only when the effort is max. A gate that requires the cell to be false will reject the recommended id at max and will also reject Wonder at high. Neither reading is catalog membership.
The re-resolution block says the measured cell is per request. That matches the high-versus-max diff: the value above the block is the resolution for the tier you passed, not a second binary. It does not match a reading in which true means “this id was measured” and false means “this id is unknown.” On this build the unknown id is the one that stays true, and the recommended id is true only at max.
The run path does not sort the id either
muse exec --list-models is the other command that sounds like a catalog check. With --provider echo it exited 1. It printed a line that local session messaging was disabled because session logging is required, then provider echo has no model catalog. Echo is the provider that can answer a prompt without credentials. It is also the provider with nothing to list. The list-models invocation did not pass --no-session-log. The messaging line was still printed. Do not treat that line as proof that you passed the flag.
With --provider meta and no credential, muse exec --list-models exited 1 before any model list. The error told you to run muse login, set META_API_KEY, or save credentials in auth.json under the config home for that run. The same error, not a model-id error, was printed by muse exec --provider meta --model muse-spark-1.3 --no-session-log --max-model-steps 1 and by the same command with --model not-a-model. A missing credential hides the id check. You cannot tell from that exit whether the id would have been accepted.
--provider local did not distinguish them either. Both ids exited 1 with Local models are disabled. Enable the local_models experiment and restart TBH. That sentence is the gate on this host. It is not a catalog result. The experiment was left disabled. Enabling it would have been a different test, and it was not required to see that the two ids take the same path.
muse exec --provider echo --model muse-spark-1.3 --no-session-log --max-model-steps 1 never reached a model call. It exited 2 with --model requires --provider meta or local. The same line was printed for not-a-model. Echo can print a prompt back when you do not pass --model. It can refuse a model flag. It cannot confirm the id.
The offline distinguisher is model-profile. On this binary it distinguishes three literal strings, plus one effort-dependent cell on two of those strings. It does not fail closed on an unknown id. A launcher that treats exit 0 as “this id is served” will accept a typo. A launcher that prefers the report whose measured cell is true will prefer the typo over muse-spark-1.3 at the default effort, because the typo is measured and the recommended id is not, until you pass max.
A check that matches this binary
Decide which question you are asking before you read a field.
If the question is whether this binary selects stock plugin text for the id, compare plugin_text_variant to stock and require the echoed model line to be exactly muse-spark-1.2, muse-spark-1.3, or muse-spark-1.3-contributor. Case is part of the id. Muse-Spark-1.2 is not the stock id. A prefix, a suffix, or a version you hoped existed, such as muse-spark-1.4, is not the stock id either. Those strings exited 0 and took the fallback.
If the question is whether max changes the measured cell for a stock id, run the same id at high and at max and require the plugin line to stay stock while the cell moves from false to true. On this binary that happened for muse-spark-1.3 and muse-spark-1.3-contributor. It did not happen for muse-spark-1.2, which stayed false at max. Do not use ultra as a stand-in for that pair. Do not treat the contributor id’s acceptance of max as a contradiction you can file against a live request. It is a contradiction of the Spark page’s availability sentence by this offline command. A served request might still reject the tier. This command will not show that rejection.
If the question is whether the id is in the published catalog, do not use this command’s exit code. Read the models page. Use muse exec --provider meta --list-models only after a credential is present. This run stopped at the credential error, so it does not show what a logged-in list contains. It does show that the logged-out error is the same shape for a stock id and for not-a-model.
A practical diff is three commands, run from a config home that is not the operator home:
muse model-profile show muse-spark-1.3 --effort high
muse model-profile show muse-spark-1.3 --effort max
muse model-profile show not-a-model --effort high
On 1.4.24 the first and third differ in the measured cell and the plugin text. The first and second differ in the effort line and the measured cell, and they do not differ in the plugin text. Add muse-wonder-1.0 and muse-spark-1.2-contributor if the job might follow those documented ids. On this host both joined not-a-model, not the stock row.
The command does not need a login, and it should not be given one if the only question is the local resolution table. A login changes the question into a request the credential can authorize. That is a different measurement.
What this run did not prove
This was not a model call. It does not show latency, tool use, or whether Meta would reject muse-spark-1.1 or muse-wonder-1.0 at request time. It does not show the contents of a logged-in model list. It does not show what exact-minimal-literal-tests loads in a session. The local-models experiment was not enabled. No session was named, and no session message was delivered. The credential file named in the meta error was not created.
The stock row is an observation on 1.4.24-R9131.1, not a promise about the next release. If a later binary adds Wonder to the stock row, or starts refusing unknown ids, the three-command diff above will show it. An exit-code check will not.
Verification
Confirm muse --version prints Muse Code 1.4.24 (1.4.24-R9131.1). On x86_64 Linux, confirm the versioned binary’s SHA-256 against the x86_linux checksum in the release manifest before you trust the row. Run the three commands above on a config home that is not your operator home. Require plugin_text_variant stock and measured_tuning_cell false for muse-spark-1.3 at high. Require stock and true at max. Require exact-minimal-literal-tests and true for not-a-model at high. Then run muse-wonder-1.0 at high and require it to match the unknown id in those two fields, not the Spark id. If Wonder matches Spark, you are not on the resolution this article measured.
Sources
muse --versionandmuse model-profile showon 20 November 2026, Muse Code 1.4.24 (1.4.24-R9131.1), Linux x86_64. The versioned binary SHA-256 matched the public manifest. No login and no model call.- Public manifest for release
1.4.24-R9131.1,x86_linuxchecksumd752c45dce079ee29e118804b091d81ac2b1b89b67f317b4a162fa1af1206465, size 471370072. - Models, fetched 20 November 2026. Lists the catalog ids compared above.
- Muse Spark, fetched 20 November 2026. Says 1.1 is not available in Muse Code, lists both contributor ids, and says
maxis not available onmuse-spark-1.3-contributor. - Configuration and context, fetched 20 November 2026. Says the default model is
muse-spark-1.2, that the Meta provider does not acceptnone, and thatultramaps to a provider’s highest supported level. muse model-profile show --helpon the same binary. The argument line calls the value a catalog model id and givesmuse-spark-1.3as the example.