[lbx154]

Design rule 3 (no waiting in the dark), measured by the scenario eval.

What changed

Early words. While a Manager turn runs, its first model message (for example "我先看一下 core.py") is written to the shared Activity stream right away. It is a ui.argus row marked in_progress/live and carries the turn's message id (argus/webapi/manager_bridge.py). The final reply uses the same id and replaces it. Only the final reply goes into the transcript. The web feed hides a draft once its final reply, a later draft, or any later message exists (dropSupersededDrafts in frontend/web/src/lib/conversationEvents.ts). Map trace and map view skip drafts when they close a turn. The SSE route already streamed live deltas; this change also covers the blocking POST, page reloads, and other tabs.

Effort by need (argus/core/effort_policy.py, one place):

• Manager chat turn: medium. High only while the turn rules on a team stage question, because that decides where the project goes next.
• Bounded mission (direct mode, not a research/paper vertical, not a paper mission): Engineer round 1 and the Reviewer run at medium. A Reviewer-requested later round runs at high.
• Research, paper, and staged work: high.
• Nothing defaults to xhigh.
• An operator's explicit effort still wins: env, a /config switch, or hosting. ARGUS_SKILL_SELF_/MANAGER_REASONING_EFFORT sets the Manager. ARGUS_SKILL_ENGINEER_[INITIAL_]REASONING_EFFORT and ARGUS_SKILL_REVIEWER_REASONING_EFFORT set the team roles.

Scenario eval (20/20, checks 96/96 both runs)

Baseline runs/20261024-085712 vs this branch runs/20261024-092946. Medians in seconds:

Category	First response before → after	Completion before → after
Simple Q&A	20 → 2.6	30 → 20
Small code	32 → 3.0	65 → 62
Medium engineering	42 → 3.2	126 → 100
Research/writing	66 → 3.6	101 → 70

The whole run cost 105 quota units, down from 195 (estimated API price $1.81, down from $4.81).

First response now counts the in-progress first line, because that is when the person first sees words in the chat. Completion is when the answer or work is ready.

Tests

• New: tests/test_effort_policy.py.
• New vitest cases in frontend/web/src/test/conversationEvents.test.ts.
• pytest: tests/webapi, daemon, apps, manager, core, release_tools.
• Web: full vitest and tsc.
• ruff.
• Release artifacts rebuilt.

🤖 Generated with Claude Code