Runtime fallback session.status pattern drift QA

Expected impact and scope

This change affects the OpenCode-connected runtime-fallback hook when OpenCode emits a session.status event whose status type is retry. The intended behavior is immediate fallback classification and dispatch for Free usage exceeded, subscribe to Go. Session error handling, fallback chain selection, prompt replay, and the first-prompt watchdog remain otherwise unchanged.

The drift fix shares the existing model-core runtime-fallback retryable regex set with both status-path classifiers. The auto-retry extractor retains its additional retrying in, usage limit, and limit reached signals.

What was tested

1. Focused regression suite:
   • Command: bun test packages/omo-opencode/src/hooks/runtime-fallback packages/model-core
   • Artifact: focused-tests-summary.txt
   • Purpose: cover the exact signal extraction, direct session-status handler dispatch, shared-pattern identity, and existing runtime-fallback/model-core behavior.
2. Repository typecheck:
   • Command: bun run typecheck
   • Passing artifact after dependency refresh: typecheck-after-install.txt
   • Initial environment failure: initial-verification-failure.txt
3. Full Bun suite:
   • Command: bun test
   • Passing artifact after dependency refresh: full-tests-summary.txt
   • Initial environment failure: initial-verification-failure.txt
4. Worktree dependency refresh:
   • Command: bun install
   • Artifact: bun-install-summary.txt
   • Purpose: restore missing local Senpi and workspace package links in the dedicated worktree before rerunning repository-wide gates.
5. OpenCode QA harness self-check:
   • Command: bash scripts/lib/common.sh --self-check from /Users/yeongyu/local-workspaces/omo/.agents/skills/opencode-qa
   • Artifact: opencode-qa-common-self-check.txt
6. OpenCode SSE hook plumbing probe:
   • Command: bash scripts/sse-hook-probe.sh --self-test from the opencode-qa skill directory
   • Artifact: opencode-sse-hook-probe.txt
   • Purpose: prove an isolated OpenCode server publishes lifecycle events through the same event stream observed by plugin hooks.
7. Runtime-fallback hook module probe:
   • Command: bun .omo/evidence/20260801-runtime-fallback-status-drift/runtime-fallback-hook-probe.ts
   • Script: runtime-fallback-hook-probe.ts
   • Output: runtime-fallback-hook-probe.txt
   • Purpose: create the production runtime-fallback hook, drive session.created followed by the exact mocked session.status retry event through its public event handler, and assert one immediate abort plus one fallback prompt dispatch to openai/gpt-5.4.
8. OpenCode isolation proof:
   • Before: opencode-before.txt
   • After: opencode-after.txt
   • Both report the live DB session count as 21932.
9. Language-server diagnostics:
   • All changed TypeScript files were checked. No diagnostics remained after adding the required git_master fields to the touched test fixture.

What was observed

• The focused suite passed: 580 tests, 0 failures.
• Typecheck passed after bun install restored the worktree package links.
• After rebasing onto current origin/dev, the full suite passed: 12,665 tests, 3 expected skips, 0 failures.
• The OpenCode QA self-check passed all dependency, sandbox, HOME, DB path, SQL escaping, and cleanup checks.
• The isolated SSE probe received server.connected from /event.
• The production hook module probe handled Free usage exceeded, subscribe to Go immediately, recorded one abort, and dispatched one prompt using openai/gpt-5.4 with the persisted user text.
• The real OpenCode DB session count was unchanged before and after QA, proving the spawned OpenCode server stayed isolated.
• Installed OpenCode version was 1.18.7.

Why this is enough

The deterministic tests cover both classifier seams named in issue #6538 and the consumer behavior that previously returned through the non-matching diagnostic path. The shared-array identity test prevents the OpenCode status fallback list from becoming a separate copy again. The model-core extractor now consumes the same canonical array directly, so future additions to that runtime-fallback set automatically reach both status checks.

The public hook probe exercises the real runtime-fallback module boundary and dispatch flow, not only a standalone regex. The SSE probe separately proves OpenCode lifecycle event plumbing in an isolated harness. The unchanged live DB count demonstrates that QA did not pollute the user's OpenCode state.

What was omitted

• No real provider quota was exhausted. Forcing a live account into quota exhaustion is impractical and could affect user billing or availability. The exact provider message was injected deterministically through the production hook module instead.
• No TUI smoke was run because the changed behavior is an event-hook classification and prompt dispatch path with no visual interaction.
• Full per-test output from the 12,668-test repository suite was summarized instead of committed because it was over 1 MB. The reviewer-readable counts and result are in full-tests-summary.txt.
• No raw environment dump, credentials, auth headers, provider tokens, or private session contents were captured.

Residual risk

Low. The change broadens the auto-retry status extractor to every pattern already accepted by the runtime-fallback session-error classifier. This is intentional drift prevention, and focused plus full-suite coverage found no behavior regressions. A provider can still introduce a new retry phrase not represented in the canonical pattern set, which would continue to use the existing non-matching diagnostic until classified.