BUG: Carry the composer's live model selection onto the bug session

[gennadiryan]

Important

Problem

A bug session runs the workspace's configured default model, not the model the user was actually running when the bug appeared. Model-specific defects are therefore filed against the wrong model, may not reproduce for whoever triages them, and the mis-attribution is silent — nothing errors, the recorded model is simply wrong.

Approach

Carry the composer's live model selection (provider, model, and variant) across the app→extension command bridge and onto the report-a-bug arming turn, with precedence over the configured default. Key reason: every piece already exists except the plumbing — the bridge already has a validated extractor for this exact payload, and the session command API already accepts an optional model and variant.

Approaches Considered

• Carry the selection over the existing command bridge (chosen) — the extractor and its validation already exist and are unit-covered; the only new surface is passing the value through two call sites.
• Session-level configuration pin only — already shipped, but it cannot express a variant and ignores the user's live choice entirely.
• Resolve the model server-side from the originating session — needs a reverse lookup the bug-report orchestration deliberately avoids; the extension owns the session end-to-end by design.
• Record the model in the context envelope without running it — fixes attribution in the report but the bug session still runs the wrong model, so behavioural repros stay wrong.

Scope

• In: validating the model payload at the bridge; accepting it on the report-a-bug command; passing provider/model and variant to the arming turn; precedence over the configured default; tests for each branch.
• Out: the configured-default behaviour itself — unpinned installs still let the server resolve, and that is correct. Other bridge commands — only report-a-bug carries a payload today, and the allowlist must not widen. Model-selection UI — unchanged; this only reads what the composer already knows.

Assumptions / Open Qs

• Assumes the composer can supply its live selection at click time. If it cannot, this degrades to the configured default — i.e. current behaviour — so the change is safe but inert.
• Open: should a selected-but-unavailable model fail loudly, or fall back silently? Current validation strips malformed payloads and falls back; that is the assumed answer unless reviewers disagree.

---

Acceptance Criteria

1. Invoking report-a-bug with a model selected in the composer arms the bug session with that provider and model.
2. A selected variant travels with the selection and is applied to the same arming turn.
3. With no live selection, the configured default is used; with neither, the model field is omitted entirely and the server resolves its own.
4. A malformed, oversized, or absent model payload never blocks the command — it falls back to the next precedence level.
5. The payload is shape-validated and length-bounded at the bridge, before it reaches the orchestrator.
6. Tests cover: live selection honoured; variant honoured; precedence over the configured default; malformed payload stripped; absent payload tolerated.
7. No additional command gains a payload channel, and the bridge command allowlist is unchanged in size.

Key Decisions

• Precedence is live selection → configured default → server default. Each level is a strict fallback; none of them guesses.
• Omit rather than send empty. An empty model string pins the session to a nonexistent model, which is worse than sending nothing.
• Validation stays at the bridge, which is the trust boundary — the framed app renders model output, so its messages are untrusted input. The orchestrator receives an already-validated value.
• The model rides the arming turn, not session creation. Session creation carries identity and the context envelope; the model is a property of the turn.

Constraints & Invariants

• The bridge handler is deliberately paranoid: strict command allowlist, bounded payloads. Any new field must keep that posture.
• Sanitization invariants for bug sessions are unchanged — run context stays pointer-only, no absolute paths, no log payloads.
• The single-open bug-session invariant and the machine-managed lifecycle are untouched by this change.

Prior Art / Patterns

• The bridge module already exports a validated extractor for exactly this payload shape (provider, model, optional variant), with bounded-string checks and undefined-on-malformed semantics. It has unit coverage and no production caller — wiring it is the bulk of this task.
• The session command API accepts optional model and variant alongside the required command and arguments.
• The bug-report orchestrator already resolves a session-level model pin from configuration and omits the field when unset; the live selection slots in ahead of it.
• The bridge's existing bounded-string validation for the default-model mirror is the pattern to follow for the new field.

Source

• Parent: PRD: Add a Report-a-Bug entry point with machine-managed bug sessions #249 — PRD: Add a Report-a-Bug entry point with machine-managed bug sessions.
• Surfaced by a harness bug sweep of the extension's vendored-opencode seam. The sweep found the extractor and its test asserting a handoff that was never wired at either level.
• The session-level half (configuration pin → arming turn) is implemented separately; this issue is only the composer→bridge half.
• Related: Harness-swappable: the vendored-opencode seam becomes an adapter boundary #243 — harness-swappable: the vendored-opencode seam becomes an adapter boundary.

Notes

Until this lands, bug sessions run the configured or server default and any model attribution in a filed report is unreliable. That is worth knowing during triage of every bug filed through the feature, not just model-specific ones.

There is a mild recursion here: a bug report about the model handoff, filed through the report-a-bug feature, is itself authored under the wrong model. Filing this one conventionally avoids the confusion.