A live capture against claude 2.1.150 (both --print and the interactive
stream-json transport clide uses) confirms --include-partial-messages
emits the in-progress reply as stream_event envelopes wrapping Anthropic
streaming deltas — NOT assistant events with partial:true, which is what
T-168 assumed, so that handler never fired and streaming was inert.
Replace it: accumulate content_block_delta text per message id (tracked
from message_start, since deltas carry no id) and emit a placeholder under
a stable partial-<id> uuid the controller upserts in place; the matching
single-text-block assistant event reuses that uuid to finalize, while
tool_use / thinking blocks keep their own uuids and append in order. Tests
rewritten against the captured shape; spike doc records it.
T-184.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude's tmux team mode let teammates message each other and share a task
list; that mode is undocumented and unavailable headless. clide rebuilds
the same behavior over its own managed sessions, as the broker.
Verified live against claude 2.1.150 that a spawner can host an in-process
("SDK") MCP server entirely over the stream-json control channel — no
subprocess, no --mcp-config, no socket: declare the server name in the
initialize handshake's sdkMcpServers, answer the mcp_message JSON-RPC
round-trips (initialize / tools/list / tools/call) under
response.response.mcp_response. SDK tool calls are permission-gated through
the existing can_use_tool path. Documented in the 2.1.150 spike §6.
StreamJsonSession gains an McpServer hosting seam; TeamBroker + TeamMcpServer
expose send_message / broadcast / list_teammates / inbox / claim_task /
task_status, all routed through one shared broker. The orchestrator owns the
broker, registers each team session, delivers a message into the target's
next turn on its stdin, and injects roster + role via --append-system-prompt.
Solo sessions are unchanged (no MCP server, no initialize handshake).
T-170, D-77.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two spot-check fixes (T-178, T-179), both grounded in a boundary test of
the stream-json wire (findings folded into the spike doc):
- Harness-injected user messages (skill loads, slash-command expansions,
system reminders) carry isSynthetic on the wire (isMeta in the
transcript). They were rendering as blue "you" cards though the user
never typed them; now UserMessage.injected flags them and the view
shows a muted, collapsed "context" card instead.
- Permission prompts now show the command/input being permitted (a
capped, scrollable code block) so you can see what you approve. Instead
of fully hiding a prompted tool-use, once resolved it collapses to a
one-line summary with a green (approved) or red (denied) border; the
session tracks per-tool_use_id outcome and the view colours it. The
result is kept.
Corrects an earlier wrong assumption: the Skill tool is auto-allowed
(no permission prompt); the inject only appears once the Skill tool is
actually invoked, which is why deny-captures missed it.
T-178, T-179, D-78.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Empirical spike against claude 2.1.150 (driving the real CLI over
stdin/stdout + reading the shipped binary's zod schemas) pinned the
wire shapes for the stream-json control protocol: the can_use_tool
permission request, the control_response envelope, the
--permission-prompt-tool stdio enabler (without it "ask" tools silently
auto-deny), the allow-requires-updatedInput quirk, and AskUserQuestion
answered via updatedInput.answers. Captured in a version-pinned spike
note with a resilience section (detection canaries + a ranked fallback
menu) so a future Anthropic change to this undocumented contract doesn't
leave us at a blank slate.
D-78 records the decision: permissions ride the stdio control channel,
not MCP (MCP is reserved for capability/tool provision); refines D-77.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
test / unit + widget + golden + a11y (push) Failing after 31s
test / integration_test (xvfb) (push) Has been skipped
test / bundle smoke (xvfb 5s) (push) Has been skipped
test / daemon subprocess + web WASM smoke (push) Has been skipped
test / dart doc (lib API) (push) Failing after 31s
Findings note for the tmux agent-team work (epic T-132), validated from
real on-disk artifacts (42 past team configs, real team + subagent
transcripts, current-version .meta.json) plus a synthetic control-mode
test and the tmux manual — no live team run / quota needed.
Key results: teammates get tmux panes (config tmuxPaneId) and write
transcripts at <munged>/<sid>/subagents/agent-<hex>.jsonl + .meta.json;
tmux 3.6a has no %pane-died (use %window-close/%layout-change), and
polling list-panes -a is the reliable lifecycle baseline; the real risk
is identity linkage — pane/teammate (config: name@team, agentType,
%pane) joins to the transcript only via agentType (config <-> .meta.json),
ambiguous for same-type teammates, so T-139 needs a joinedAt/mtime or
lead-transcript tiebreaker.
Co-Authored-By: Claude <noreply@anthropic.com>
jpmschweitzerandClaude