feat(agent-gui): replace processing row with per-turn progress bar#1406
feat(agent-gui): replace processing row with per-turn progress bar#1406devRickyyy wants to merge 1 commit into
Conversation
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_59ab4761-a27f-4d82-b205-47e3f0889074) |
The "Planning next moves" row is now a live progress bar that follows the active turn: model phase (waiting for response vs responding) with a per-phase elapsed timer anchored to message/tool completion times, plus cumulative per-turn token counters for providers with the tokenUsage capability (Claude Code, Codex). Phase and timing are derived in the renderer from canonical turn and message timestamps. Token counters are accumulated daemon-side per turn (Claude: message_start input plus delta output; Codex: thread total baseline diff) and flushed through turn transitions into the durable WorkspaceAgentTurn.tokenUsage field (new workspace_agent_turns token_usage_json column). Signed-off-by: devRickyyy <74520253+devRickyyy@users.noreply.github.com>
a6bf856 to
dedbd04
Compare
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| event.Type != activityshared.EventRootProviderTurnCompleted { | ||
| patch.Turn = &agentsessionstore.WorkspaceAgentTurnPatch{ | ||
| TurnID: turnID, | ||
| CapabilityRefs: activityCapabilityReferencesFromEvent(event), |
There was a problem hiding this comment.
🔴 Per-turn token counts never get saved, so the new usage display stays empty
The per-turn token counts are dropped from each turn update (statePatchFromSessionEvent at packages/agent/daemon/runtime/reporter_state.go:76-80) before the turn is saved, so the input/output token numbers never reach the transcript.
Impact: The new ↑/↓ token counters on the active-turn progress bar are always blank, and no turn ever persists its token usage, making the entire per-turn token feature inert.
Why the structured field is never populated
The daemon computes token counters and emits them as turn.updated events whose Payload.Metadata["tokenUsage"] map carries inputTokens/outputTokens (see turnTokenUsageEvent at packages/agent/daemon/runtime/turn_token_usage.go:27-33). Every downstream stage reads a structured field, not the metadata: cloneWorkspaceAgentStatePatch copies patch.Turn.TokenUsage (packages/agent/daemon/activity/compat.go:410), turnTransitionFromStateInput copies turn.TokenUsage (services/tuttid/service/agent/activity_projection_turns.go:120), and the store/OpenAPI/WS payloads read that transition field.
The only producer of patch.Turn is statePatchFromSessionEvent. Its if patch.Turn != nil block (packages/agent/daemon/runtime/reporter_state.go:76-80) extracts metadata["fileChanges"] into patch.Turn.FileChanges but has no equivalent line copying metadata["tokenUsage"] into patch.Turn.TokenUsage. Neither applyLifecycleSnapshotToPatch nor applyExplicitTurnLifecycleToPatch sets it either. As a result patch.Turn.TokenUsage is always nil, and every downstream copy is nil.
This also means the new test TestStatePatchFromSessionEventMapsTokenUsage (packages/agent/daemon/runtime/turn_token_usage_test.go:333), which asserts patch.Turn.TokenUsage.InputTokens == 1200, cannot pass against the current file.
Fix: mirror the fileChanges precedent, e.g. after line 79 add extraction of metadata["tokenUsage"] (inputTokens/outputTokens via payloadInt64, tolerating float64) into a &agentsessionstore.WorkspaceAgentTurnTokenUsage{...}.
Prompt for agents
statePatchFromSessionEvent in packages/agent/daemon/runtime/reporter_state.go is the sole producer of patch.Turn (a *agentsessionstore.WorkspaceAgentTurnPatch). The per-turn token feature emits turn.updated events with Payload.Metadata["tokenUsage"] = { inputTokens, outputTokens } (see turnTokenUsageEvent in turn_token_usage.go). Every downstream stage (compat.go cloneWorkspaceAgentStatePatch, tuttid turnTransitionFromStateInput, the store, and the OpenAPI/WS payload) consumes the structured patch.Turn.TokenUsage field, NOT the metadata. But the `if patch.Turn != nil` block (around lines 76-80) only copies metadata["fileChanges"] into patch.Turn.FileChanges; it never copies metadata["tokenUsage"] into patch.Turn.TokenUsage. Consequently the token counters are computed and emitted but never persisted or transported, and the test TestStatePatchFromSessionEventMapsTokenUsage would fail. Add a metadata extraction alongside fileChanges that reads inputTokens/outputTokens (JSON metadata decodes numbers as float64, so use payloadInt64 which already handles float64/json.Number) and assigns patch.Turn.TokenUsage = &agentsessionstore.WorkspaceAgentTurnTokenUsage{InputTokens: ..., OutputTokens: ...} when the tokenUsage map is present.
Was this helpful? React with 👍 or 👎 to provide feedback.
1 participant
Summary
Replaces the "Planning next moves" processing row in the AgentGUI transcript with a live per-turn progress bar that follows the active turn:
tokenUsagecapability (Claude Code, Codex/tutti-agent). Providers without an input/output split (ACP family) show phase + timer only.Phase and timing are derived in the renderer from canonical turn state and in-flight message timestamps — no new wire phases. Token counters are accumulated daemon-side per turn and flushed through turn transitions at ~1 Hz into a new durable
WorkspaceAgentTurn.tokenUsagefield:message_startusage (per-API-call input); daemon folds per-call cumulative output deltas into turn totals.thread/tokenUsage/updatedthread totals (keyed by threadId; reasoning output counted as output; stale/negative diffs clamped).fileChangesprecedent:turn.updatedmetadata →TurnTransition→ newworkspace_agent_turns.token_usage_jsoncolumn (migration included) → OpenAPI/WS turn payload. Final flush always lands before the settle transition (settled turns reject later transitions).Contract changes:
WorkspaceAgentTurnTokenUsageschema intuttid.v1.yaml, matching optionaltokenUsageon theturn_updateWS payload, regenerated clients, and the mirrored types inagent-activity-core.Documentation:
docs/architecture/agent-gui-node.md§3.2/§4.6 now record the tokenUsage ownership and the renderer-side progress-row derivation rule.Verification
cd services/tuttid && go test ./... && go build ./...; same forpackages/agent/daemonandpackages/agent/store-sqlite— all green, incl. new counter, repository round-trip, migration idempotency, and projection-guard testsgolangci-lint runon the touched daemon packages — 0 issuespnpm --filter @tutti-os/claude-sdk-sidecar test— 87/87pnpm --filter @tutti-os/agent-gui test— 227 files / 1999 tests (new projection, component, and transcript cases)pnpm --filter @tutti-os/agent-activity-core test— 208 pass; desktop bridge/adapter/activity-service suites green (39/39 on the reconcile-failure fixture)pnpm check:api-generated,pnpm check:event-protocol-generated,pnpm check:i18n,pnpm lint:ts,pnpm typecheck(27/27), electron/renderer/ui boundary checks, agent-gui degradation checkpnpm check:changed -- --push-ready— 32/32 lanes greenFallback Three Questions
fileChanges.Checklist
git commit -s.Note
Medium Risk
Touches turn lifecycle ordering (token flush before settle), multi-provider token accounting, and live transcript projection; mistakes could show wrong timers/tokens or drop final usage on settled turns.
Overview
Replaces the generic "Planning next moves" transcript row with a per-turn progress bar on the active turn: Waiting for response vs Responding, a per-phase elapsed timer from canonical turn phase and message/tool timestamps (not provider wire phases), and ↑/↓ token counts when the session has the
tokenUsagecapability. The row is suppressed for non-live turn phases and when another turn-specific progress row (e.g. compact) is already shown; it still hides during in-flight tool work via existing presentation rules.Daemon / persistence: Claude sidecar emits
usage_updatedon rootmessage_start; Claude and Codex adapters accumulate per-turn input/output (Codex diffs thread totals with baseline/replay handling), throttle mid-turn flushes to ~1 Hz, and always flush before settle. Counters persist onWorkspaceAgentTurn.tokenUsage(token_usage_json) through the sameturn.updatedpath asfileChanges, with types wired through activity-core, desktop adapter, and reconcile bridge coercion.Supporting fixes: timeline merge prefers latest
completedAtUnixMsfor messages/tools; completion times flow through detail projection for accurate awaiting-phase anchors; Claude lifecycle logging moved toclaude_sdk_diagnostics.go; interaction upsert/list helpers consolidated in store-sqlite.Reviewed by Cursor Bugbot for commit a6bf856. Bugbot is set up for automated code reviews on this repo. Configure here.