ExitPlanMode arrives as a can_use_tool permission prompt and was approved
like any other tool — the control_response was sent but the tracked
SessionStatus.permissionMode never changed, so the mode indicator and
composer stayed on "plan" after the plan was accepted.
On approving an ExitPlanMode prompt, sync the tracked mode to 'default'
(the CLI performs the transition itself, so no set_permission_mode control
request is sent — we only mirror it). The change rides the existing
statusStream → claude_pane._status plumbing, so the permission-mode
control and status indicator update with no extra wiring. Deny, and any
non-ExitPlanMode tool, leave the mode untouched.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Raise the declared minimums in pubspec.yaml to what our deps already
require: Flutter >=3.35.0 / Dart >=3.9.0 (was 3.19.0 / 3.5.0). alchemist
0.12 needs Flutter 3.32; Dart 3.9 first ships in Flutter 3.35, so 3.35 is
the binding floor. Pin the exact build toolchain in .fvmrc (Flutter
3.44.1).
Moving to the Dart 3.9 language level switches `dart format` to the new
"tall" style and enables two new lints. This commit is the resulting
mechanical churn, isolated from any behaviour change:
- whole-tree `dart format` reformat (tall style)
- `dart fix` for unnecessary_underscores + use_null_aware_elements
No runtime behaviour change; `make test` green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A user-initiated denial (Deny & simplify) came back as an isError
tool_result and rendered as a prominent expanded-red "Bash · error"
block — pure noise, since the user chose it. It now folds to a muted,
collapsed "denied" card.
Built as a reusable filter rather than string-matching the note: DenyTool
carries a `quiet` flag, the session collects quiet denials' tool_use_ids,
and ConversationView renders any error whose id is in that set folded +
muted. Genuine tool failures (ids not in the set) keep the expanded-red
treatment. Adding future "expected error" cases is just adding ids.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
setPermissionMode only wrote the control_request; the mode was never
folded back into the session status, so the badge and Ctrl/Cmd+M looked
dead. Optimistically merge + emit the new mode immediately; the next
system/init still reconciles.
Co-Authored-By: Claude <noreply@anthropic.com>
Now that the editor split actually opens (T-197), it exposed latent
issues, plus a coincidental Claude-session crash in the same log:
- T-203: the _EditorDragHandle's slider Semantics had value +
onIncrease/onDecrease but no increased/decreasedValue, so Flutter
asserted on every semantics flush — add them. And opening the split
reparented the Claude pane (direct child → Column/Expanded), tearing
down its SelectableRegion mid selection-update ('selectable not in
this registrar' / 'inactive element'); a stable GlobalKey on the
workspace primary makes Flutter move the element instead.
- T-202: rate_limit_event.resetsAt arrives as a unix-epoch number but
was cast `as String?`, throwing in the stream-json line parser. Accept
a num (epoch) or an ISO string.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A --fork-session branch is spawned without --session-id, so claude mints
a new session id that only arrives in the init event; the ManagedSession
was left holding its placeholder. StreamJsonSession now captures
session_id from the first event that carries it and exposes it via
claudeSessionId / sessionIdResolved; the orchestrator folds that back into
ManagedSession.sessionId (idempotent for normal sessions). A fork can now
itself be resumed or forked.
T-185.
Co-Authored-By: Claude <noreply@anthropic.com>
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>
Each roster row shows a D/A/P mode badge reflecting the session's live
permission mode. Click cycles the safe trio default -> acceptEdits ->
plan and sends a set_permission_mode control_request to that session
(mirrors interrupt(); fire-and-forget). bypassPermissions is a footgun,
so it is reachable only on Shift-click and behind an inline confirm. A
clide.agent.set-permission-mode command gives the CLI parity.
T-181.
Co-Authored-By: Claude <noreply@anthropic.com>
The conversation pane now exploits the structured stream instead of
dumping tool input as JSON. ConversationController indexes tool_use by id
so a tool_result pairs back to its call and renders the Edit/Write diff or
is_error failure in place; per-tool bodies (Bash command+output, Read/Grep
file/query) reuse the shared renderers factored out of the permission
card. SessionStatus gains cost + contextWindow + rate-limit, read straight
off the init/result/rate_limit_event events, so the in-pane status line
reflects live state without the config probe.
Partial-message streaming is wired behind --include-partial-messages but
its event shape is unverified against the live binary and degrades to a
no-op if it differs — see T-184.
T-168.
Co-Authored-By: Claude <noreply@anthropic.com>
test / unit + widget + golden + a11y (push) Failing after 29s
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 34s
The sidebar had outgrown one scroll (stats + roster + config don't fit). A
sub-tab strip now switches between three surfaces: Activity (usage stats +
the primary session's live runtime), Team (the member roster, auto-fronted
when a team spawns and otherwise quiet), and Config (the Claude-environment
settings table over ClaudeConfig).
Activity and Config render their key→value rows through one shared table
(same label column + row pitch + header style) so toggling tabs doesn't move
anything. The expandable skills/agents/commands/permissions/MCP browser on
the Config tab is the follow-up (T-183). Exposes StreamJsonSession.status so
the runtime row can seed from the session's current state.
T-182, D-77.
Co-Authored-By: Claude Opus 4.7 (1M context) <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>
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 29s
A runaway turn had no escape: Escape was unbound once the slash typeahead
was closed, and there was no Stop affordance. Now the composer interrupts
the in-flight turn — Escape (when no typeahead is open) or a Stop button
shown while busy — over the stream-json control channel.
StreamJsonSession gains interrupt() (writes a {subtype: interrupt}
control_request; claude cancels the turn and ends it with a result) and a
busy/busyStream signal driven true on send and false on the next result.
The pane binds onInterrupt to the session and reflects busy reactively.
D-78.
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>
A permission-gated tool or AskUserQuestion already surfaces as a prompt
in the composer zone, so its raw tool-use card was redundant noise. The
session now tracks which tool_use_ids surfaced as a prompt; the
conversation view hides those tool-use cards. AskUserQuestion also hides
its result (the chosen answer is logged separately); permission-tool
results are kept — that's the useful outcome. The pane rebuilds the
view on each prompt change so the payload vanishes the moment its prompt
appears.
T-176, T-177, D-78.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Builds on the in-composer prompt surface (D-78):
- Permission prompts (T-175): Allow / Allow-and-don't-ask-again / Deny.
"Don't ask again" appears only when the request carries a
permission_suggestion and echoes it back as updatedPermissions. An
optional note rides Deny as the message, or Allow as a follow-up user
message (the protocol has no allow-with-message).
- AskUserQuestion picker (T-176): a single question renders bare; 2-4
questions step one at a time (nav shows "N · Header", ✓ when answered)
then a review/confirm screen. Each question offers an "Other" free-text
choice and a per-choice note; multi-select joins labels. A "chat
instead" escape denies the prompt so the user can type freely. On
submit the answer is echoed into the log, since the card is ephemeral.
- Collapsed tool cards (T-177): multi-line tool_use / tool_result start
collapsed behind a one-line summary; one-line output renders inline.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the Claude pane's tmux-TUI + transcript-tail backend with
Claude Code's stream-json control protocol (D-77/D-78). A
StreamJsonSession owns the `claude` process: its event stream feeds the
existing ConversationController, and permission / AskUserQuestion
prompts arrive as can_use_tool control_requests. Those surface as a
ToolPrompt in the composer zone — the pane swaps the text input for an
Allow/Deny card or an option picker while a prompt is open, so
interaction stays out of the conversation stream and the prompt buttons
don't fight the message-card hover chrome. The decision is written back
as a control_response (allow echoes updatedInput; AskUserQuestion
answers go in updatedInput.answers). Unsupported control subtypes are
answered with an error so a turn never hangs.
Session continuity is --resume (existing transcript) vs --session-id
(new); /clear and /resume respawn the process. The transcript reader
still backs the sidebar/status/team surfaces. T-165, T-166.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
test / unit + widget + golden + a11y (push) Failing after 25s
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 23s
The testable core of the stream-json pivot (D-77): StreamJsonSession drives
a claude process in stream-json mode — parsing its line-delimited events into
the existing ConversationItem / SessionStatus types (reusing
parseTranscriptChunk, since stream-json assistant/user events share the
transcript's message.content shapes), pulling permission-mode off the init
event, and sending user input as stream-json over stdin (with a local echo
so the user's own message renders immediately). The process is abstracted
behind StreamJsonProcess so it unit-tests without spawning; the real
ClaudeStreamJsonProcess wraps Process.start.
Not yet wired into the pane — the claude_pane integration (replace the
tmux/PTY spawn + TranscriptReader feed, route input through send) and live
verification are the next step.
T-165.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
