feat: send follow-up prompts to live agent runtimes

[mariusvniekerk]

MCP clients can now send a follow-up prompt to an existing live agent runtime using its workspace ID and runtime session key.

• Reuses the serialized bracketed-paste and Enter path used for the initial prompt.
• Rejects targets that are not live agent runtimes.
• Returns after submission without launching, persisting, resuming, or waiting for activity.

[roborev-ci]

roborev: Combined Review (d743e99)

Verdict: One medium-severity issue remains regarding ambiguous follow-up delivery outcomes.

• Medium — internal/server/workspaceapi/initial_message.go:217-224, internal/server/mcp_backend.go:648-650: Follow-up writes may continue after the caller’s context times out, causing the response to report failure even though the prompt is later delivered. Because this is classified as a non-retryable, non-ambiguous internal_error, retries may submit the same follow-up twice. Classify post-write/context-timeout failures as an ambiguous mutation outcome, such as via mcpBackendMutationError or a dedicated mutationOutcomeUnknown error.

---

Reviewers: 2 done | Synthesis: codex, 8s | Total: 5m50s

[roborev-ci]

roborev: Combined Review (c0dda73)

Verdict: One medium-severity issue found; otherwise the changes look sound.

Medium

• internal/server/mcp_backend.go:648-649: Write or timeout failures may occur after a prompt has been partially or fully sent, but this path maps them to an unambiguous MCP error. Callers cannot safely determine whether retrying could submit a duplicate prompt. Mark unknown terminal-write outcomes as ambiguous (for example, via mcpBackendMutationError(err)) and add coverage for this failure path.

---

Reviewers: 2 done | Synthesis: codex, 7s | Total: 6m29s