fix(*): let raven's own sub-agents pick a model when their menu came back empty

[0xKT]

Summary

On the agent hub, the Model pill of one of Raven's own acp rows (Raven-PPT on the reporting machine) read "follows the main Raven" and could not be opened. The row's model rule was keyed off its transport alone: an acp row could only pick from the choices its handshake advertised, this row's stored handshake advertised none, and subagents.update would have refused any value for the same reason. Nothing re-measures such a menu, because a capability snapshot is invalidated only by its launch-config fingerprint, which had not moved.

Three commits:

1. fix(rpc): the model rule is decided per row, in one helper read by both the listing and the write. An acp row still picks from its advertised menu; one of raven's own that advertised none reports model_source: "raven" and is validated and stored through the existing host-catalogue branch, so the page offers the same provider menu the built-in row gets (no page code change) and stores the canonical provider-qualified id the child's own config.set accepts. A third party that advertised none is still refused. The model_source description says the same on the wire; the enum is unchanged and both generated clients are regenerated.
2. fix(subagents): a ready snapshot of raven's own agent with a measured, empty menu wants another measurement, on connect and in the boot backfill. Scoped to raven's own agents so a third party that really offers none is not relaunched every boot.
3. fix(*): found by an adversarial pass over the served page. The boot re-measure could give the row its menu while the reader had the host picker open, and the write then refused the pick it had just offered ("has no model ...; it offers 55", with a Retry of the same doomed write). The listing's rule is now a menu, not a gate: one of raven's own accepts an advertised choice or a host-qualified id under either rule, because it runs on this host's providers whatever its handshake advertised; a third-party row keeps the membership-only check, and an own row's refusal names both vocabularies. On the page, a host id kept across a re-measure is drawn with its provider label and clear control instead of as a raw string, and a refused model write refetches the listing so the sheet repaints from the current rule. The empty-menu re-measure is bounded: only when the host holds a usable provider credential, and once per row per process, so a child that genuinely advertises no model is not relaunched at every boot and connect.

Not in this PR (follow-up): the task pane's per-instance model picker keeps the old rule (SubagentManager.set_instance_model validates against the advertised menu only), so a host id picked for an own row on the hub is refused by subagents.instance.set_model. Known limit: a product launched on its own key does not inherit the host providers, so a host model picked for it is refused by the child and only logged (acp_agent._say_model_refused).

Type

• Fix
• Feature
• Docs
• CI / tooling
• Refactor
• Other

Verification

• uv run pytest tests/test_rpc_subagents.py tests/test_rpc_subagents_model.py tests/test_rpc_schema_match.py tests/test_agent_schemas.py tests/test_subagent_probe.py tests/test_subagent_acp.py -q -- 817 passed after the first two commits. New: listing rule (own + empty -> raven, third party + empty -> agent, advertised menu kept verbatim), update accepts/refuses through the host branch for an own row, a third-party row with none is still refused; capabilities_wanted and the boot backfill re-measure own + ready + empty exactly once. The rule tests fail against the previous subagents.py; the probe tests fail against the previous probe.py.

• The new tests rerun with an empty HOME -- 84 passed.

• uv run pytest tests/test_rpc_subagents.py tests/test_rpc_subagents_model.py tests/test_subagent_probe.py tests/test_rpc_schema_match.py tests/test_subagent_acp.py -q -- 804 passed after the third commit. New: an own row WITH a menu takes a host id and still takes an advertised value; an id neither vocabulary holds is refused with a message naming both; a third-party row with a menu still refuses a host id; a row told twice it has no menu is not re-measured a third time in the same process; a key-less host re-measures nothing at boot or connect; capabilities_wanted answers false on a key-less home and true once a provider key is written. The update cases fail against the second commit's subagents.py; each bound of the probe change was shown by a targeted mutation to red exactly its own test.

• npm test --prefix ui-web -- src/features/extAgents -- 95 passed (new: an agent row storing a host id outside its choices shows the split id under its provider label and still clears; a refused model pick repaints from the listing).

• npm test --prefix ui-web -- 188 files, 2467 passed.

• npm run gen:check --prefix ui-web, npm run lint:rpc --prefix ui-tui, uv run ruff check raven tests, npm run type-check --prefix ui-web, npm run lint --prefix ui-web -- clean.

• Served page with an isolated agent home carrying a copy of the real capability snapshot (Raven-PPT measured with zero choices): before, subagents.list reports model_source: "agent" and the pill is disabled; after, it reports "raven", the pill opens the host picker, a pick is stored on the row as a provider-qualified id and shown with a clear control, the clear puts it back; a sibling with a measured menu and a third-party row with none are unchanged. Second pass, Raven-Code's stored menu emptied so the boot re-measure fires: the re-measure runs once and restores 55 choices; a pick made from the host picker opened before the re-measure and sent after it lands (stored openrouter/z-ai/glm-5.3-flash, pill shows the model and provider with the clear control, no Retry); a host id stored before the re-measure reads "anthropic/claude-opus-5 . OpenRouter" after it; refusals carry the combined sentence; the third-party row keeps its disabled pill.

• Relevant tests pass locally

• Relevant lint / type checks pass locally

• User-facing docs or screenshots are updated when needed

Risk

• Security impact considered
• Backward compatibility considered
• Rollback path is clear for risky changes

model_source: "raven" now means "picks from raven's catalogue" rather than "is the built-in row"; the agent hub page is its only reader. A stored host id on an own acp row is an ordinary provider-qualified model id, readable by older code as a value outside the advertised menu (shown and clearable). The boot re-measure costs one child handshake per process for an own agent that keeps advertising no menu, and none on a host without provider credentials. Rollback: revert the three commits in reverse order.

Related Issues

N/A

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the full diff plus the capability-snapshot, model-dispatch, RPC, and web-picker callers. I also checked the relevant history, backward compatibility of the optional wire field, AGENTS.md and CONTEXT architecture constraints, and whether existing tests were weakened; the new cases preserve the third-party behavior while covering Raven-owned empty menus and remeasurement.

Verification: uv run pytest tests/test_rpc_subagents.py tests/test_rpc_subagents_model.py tests/test_subagent_probe.py -x (193 passed); uv run pytest tests/test_rpc_schema_match.py -x (392 passed); npm test -- --run src/features/extAgents/ExtAgentsPage.test.tsx (36 passed); npm run lint -- --quiet (passed); source-language check and git diff --check passed.

[0xKT]

Third commit pushed, 5cc6b578, after an adversarial pass over the served page with a snapshot whose own row had an emptied menu, so the boot re-measure of the second commit fires.

#	Finding	Disposition
1	The rule flipped under an open sheet: the boot re-measure gave the row its 55 choices while the page still offered the host picker, so the pick was refused ("has no model ...; it offers 55") and the sheet offered a Retry of the same doomed write.	Fixed on the write, not the listing: an own acp row (snapshot.agent_name == "raven") takes an advertised choice OR a host-qualified id under either rule, through the existing _host_pair branch, because it runs on this host's providers whatever it advertised. A third-party row keeps the membership-only check. An own row's refusal names both vocabularies. The page also refetches the listing after a refused model write so the sheet repaints from the current rule.
2	A host id stored before the re-measure rendered as a raw string with no provider chip afterwards (shownModel looked the value up in model_choices only).	Fixed for own rows: a value outside the choices that carries a / falls back to the host split, so it reads "anthropic/claude-opus-5 . OpenRouter" with its clear control. Not applied to third-party rows, whose ids are their own.
3	An own row whose menu is genuinely empty was re-measured at every boot and every Connect, forever.	Bounded: the empty-menu reason fires only when the host itself holds a usable provider credential (credential_status(..., include_external=True), the predicate the model picker's "configured" flag reads; section_has_credentials reports never-signed-in OAuth vendors as usable and would not bind), and once per row per process. Stale, unmeasured-menu and credential-refusal reasons are untouched.
4	subagents.instance.set_model (task pane, per instance) still validates against the advertised menu only, so a host id picked on the hub is refused there.	Not in this PR; already named as the follow-up in the description.

Tests: own row with a menu accepts a host id and still accepts an advertised value; an id neither vocabulary holds is refused with a message naming both; third-party row with a menu still refuses a host id; a row told twice it has no menu is not re-measured a third time in one process; a key-less host re-measures nothing; capabilities_wanted is false on a key-less home and true once a key is written; on the page, an agent row storing a host id outside its choices shows the split id under its provider label and clears; a refused pick repaints from the listing. The update cases fail against the previous subagents.py; each probe bound was shown by a targeted mutation to red exactly its own test; the two vitest cases fail against the previous page files. Python 804 passed on the related set; whole npm test --prefix ui-web 188 files / 2467 passed; gen:check and lint:rpc in sync (the model_source description gained one clause, so openrpc.json and both generated clients moved).

Served page: the boot re-measure still runs exactly once and restores the 55 choices; a pick made from the host picker opened before the re-measure and sent after it lands (stored openrouter/z-ai/glm-5.3-flash, pill shows model and provider with the clear control, no Retry); a host id stored before the re-measure reads with its provider after it; the third-party row keeps its disabled pill. The once-per-process half could not be staged end to end (it needs a child raven that answers ready with no menu, which a credential-less child does not do); it rests on the unit tests and mutations above.

PR description updated (three commits).

[gloryfromca]

No blockers; suggestions only, and they are marked inline.

The new revision closes the listing/write race coherently: Raven-owned ACP rows accept either advertised or host-qualified model IDs, the UI preserves the host provider badge after remeasurement, failed writes refresh the held listing, and empty-menu remeasurement is bounded without changing the stale/auth/old-snapshot paths. I found no new inline defect; the only remaining item is the author's explicitly deferred follow-up that subagents.instance.set_model still validates per-instance choices against the advertised menu rather than also accepting host IDs.

I reviewed the delta and rechecked its capability-snapshot and model-dispatch callers, wire/generated-client compatibility, UI state flow, relevant history, AGENTS.md and CONTEXT architecture constraints, and test changes. Verification: the related Python and RPC/schema set passed 589 tests; the two web suites passed 47 tests; web lint, source-language validation, and git diff --check passed.