[philazar]

Why

The runner can now dispatch a Claude agent's turn through the build's agent gateway (#2632), and the host can forward Claude Code's model traffic natively (#2916). Nothing yet connects the two for Claude Code. A wake that falls back from the gateway also has no request to carry its fallback notice, so the user never sees why it fell back.

What

• A Claude turn that the runner routes through the gateway runs on its own native forwarding generation and its own Claude process.
• Compaction follows the same route as the agent's turns.
• A wake's gateway fallback notice now shows in the Activity transcript, right after the prompt it belongs to.

How

• The turn opens a generation at the route revision it was dispatched with, with the turn's frozen headers. Admission is sealed at Claude's qualifying idle, and the generation closes after the process exits. No other turn runs on it or overlaps it.
• A route that moves on revokes the generation natively and ends the process.
• Nothing waits on gateway readiness or retries. A moved route, a refusal, an unreachable native side or a protected agent fails visibly before anything spawns. It never falls back to Claude's own configuration on its own. A turn the runner already sent to Claude's own configuration (gatewayFallback) runs there unchanged.
• Compaction reads the route once. Where turns go through the gateway, so does compaction, on its own generation. Where they would fall back, compaction is refused. Where the route doesn't forward Claude, it runs on Claude's own configuration.
• Fallback notices are kept in memory by the id of the message Claude took, at most 100 per agent. A message Claude never took has no id, so its notice shows nowhere rather than on a later turn.
• The harness declaration stays off until its wire proof.

Risk

Low today. The Claude harness declaration is off, so no build forwards Claude turns yet. The change touches the Claude Code runtime and the Activity transcript, so a regression there would affect every Claude agent.

Testing

Local pre-push lane at 017c8dbb2 (bin/node scripts/check-push.mjs, clean tree): 986 files passed, 9 skipped; 18,612 tests passed, 23 skipped. An earlier run of the same lane failed under heavy machine load and did not reproduce.

Not run yet:

• Native app acceptance with a live gateway, for turns, interrupt, compaction and the fallback notice. No live run is approved yet.
• A choice on where a fallback notice should show for a wake whose message Claude never took. It shows nowhere today.

Bigger picture

Review order: #2632 (shared gateway client) → #2479 (Pi) → #2916 (native forwarding) → this PR → goose caller. The stack conflicts with current main and needs a main integration before merge. Ready for review does not mean ready to merge.

🤖 Generated with Claude Code

[tellaho]

🤖 Design review at 017c8dbb2019ae6fa83c7a5b3a2d35ec12143f64. Intent: connect Claude turns to the gateway and place a wake’s fallback explanation beside its prompt. No additional design blocker found in source review. Claude’s gateway opt-in remains off; this is not acceptance of the live forwarding flow. Reviewed with #2632 and #2479; Pi’s recovery findings are on #2479.

Source review only. No screenshots supplied; I did not run tests, the app or a live gateway, or verify rendered themes, widths, enlarged text, keyboard/focus, interrupt or compaction behavior.

[tellaho]

🤖 Could you add a screenshot or short video of the change to the PR? Design uses these for faster reviews.