fix: resolve inherited cross-runtime models against runtime defaults - #259

The run_agent path inherits the parent session's whole runtime, including
reasoning effort, through NestedBaseRuntime. The `compozy exec --agent`
fallback is a fresh CLI invocation, so it reads workspace defaults instead.
The contract told it to pass --ide but said nothing about reasoning effort,
so advisors could silently run at a different effort than the session that
dispatched them.

State the requirement and show it in the example.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

Review of the fallback surfaced two gaps.

The fallback substituted a model with no durable trace. Its only signal was a
warning, and `exec` silences operational logs without --verbose, while
SessionStartedPayload carried no model at all. A run could execute on a model
nobody chose and leave nothing in the artifacts to show it. configureSession
now reports the model the runtime accepted, sessions record it in their
identity, and session.started carries it. This also closes a plainer gap: the
resolved model was never observable, even on the happy path.

The wiring that feeds the decision had no test. ModelExplicit is hand-copied
through a 40-field struct literal into ClientConfig, and every existing test
set ClientConfig directly, so dropping either assignment would silently
downgrade an explicitly pinned model with a green suite. Both hops are now
asserted; mutating either one fails a test.

Also note on ClientConfig.ModelExplicit that it governs SessionRequest.Model
too, which is safe only while both carry the same value.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

Two things a reader cannot infer from the code.

RuntimeForTask gives type and id task rules authority over --model, while
ModelExplicit only distinguishes the flag from everything else. Such a rule is
therefore correctable here even though it outranks the flag upstream. Encoding
that faithfully would mean giving the flag three states and touching the
precedence other resolution paths read, so record the limit instead.

The current-value guard exists because a value the runtime omits from its own
option list is no safer than the request.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

`compozy setup` installs extension skills into <cwd>/.agents/skills in project
scope (internal/setup/install.go canonicalSkillsRoot), so the tracked copy at
.agents/skills/cy-idea-factory is a build output that happens to be committed.
Every project-scope setup rewrote files nobody edited, and the copy had already
drifted from extensions/cy-idea-factory/skills/cy-idea-factory by a stray blank
line.

The copy predates the extension: compozy#35 created it when the skill was a plain repo
skill, and compozy#93 migrated the source without removing the old path.

Untracking rather than deleting keeps existing installs working; the directory
stays on disk and link-skills.sh skips it when absent, so a fresh clone simply
has no local link until the extension is installed.

This removes the duplicate the surrounding commits had to edit in tandem.
Issue #23 covers the wider question of separating tracked skill sources from
install destinations under .agents/skills.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

…llers

Adding Model to the runtime SessionIdentity changed the
agent.post_session_create payload without updating its SDK mirrors, so
TestPublicHookAndHostTypesStayAlignedWithRuntime failed on field drift and
typed extensions could not read the accepted model. Both SDKs now carry the
optional field.

Two direct ClientConfig callers pass a user-chosen model but left
ModelExplicit false, so the fallback would substitute it. That is worst in
TestLiveCodexModelAvailability, whose whole purpose is to prove a requested
model is available: it would have passed against a different one.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

RuntimeForTask lets type and id rules override --model, unlike complexity
rules, which yield to it. Those pins nonetheless reached the ACP client as
inherited values, because ModelExplicit only reflected the run-wide flag. A
`--task-runtime id=task_01,model=...` selection the runtime did not advertise
was therefore replaced by the runtime default instead of failing.

Marking the pin at the point that already grants it authority keeps the two
consistent: a rule that outranks the flag is at least as deliberate as one.
Complexity rules stay inherited, matching their weaker precedence.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>