plat 088

← Pulse platform issue index

PLAT-088 — every scheduled workflow and Pulse turn was billed to chat, so Pulse cost could not be measured

Coordination	Value
Assigned agent	unassigned
Ticket state	implemented — attribution fixed and tested; runtime reverify pending
Last synchronized	2026-08-11

• Priority: P1 — not a crash, but it makes the "what does Pulse cost us versus the goal work?" question unanswerable, and silently inflates the chat bucket with automation spend nobody typed.
• Owner: cost scope attribution (handleQuery observer construction, pkg/costobserver)
• Found on: 2026-08-11, while answering "how do we measure cost we spend on Pulse vs the goal process".

Evidence

Every cost row for the instagram scheduled run (schedule-manual--bae435e5_1786432476119732000, 2026-08-11) landed in scope='chat' — including the Pulse Gate, Review+Fix parent, and Finalize turns, which the logs show running at 08:56, 08:58, 09:48 and 09:53 UTC:

08:56:19  chat  schedule-manual--bae435e5_…  $1.56   ← Pulse Gate turn
08:58:36  chat  schedule-manual--bae435e5_…  $0.82   ← Pulse Review+Fix turn
09:48:05  chat  schedule-manual--bae435e5_…  $1.06   ← Pulse Finalize turn
09:53:59  chat  schedule-manual--bae435e5_…  $3.18   ← Pulse post-finalize turn

Only Pulse's delegated background children were tagged correctly, because their execution ids literally contain the string "pulse" (bg-pulse-review+fix:-engineering…), which matchPhaseScope matches:

09:19:38  pulse  bg-pulse-review+fix:-engineeri  $11.14
09:34:58  pulse  bg-pulse-review+fix:-engineeri  $12.87
09:40:28  pulse  bg-pulse-review+fix:-engineeri  $4.90
09:47:20  pulse  bg-pulse-review+fix:-engineeri  $4.56

So Pulse's measured cost for that run was $33.47 when the real figure was roughly $40 — and the missing ~$6.62 was charged to chat. Corroborating the scale of the problem: the ledger contains zero rows with scope='builder' in its entire history, because the code path that should produce them cannot.

Root cause

handleQuery captures isWorkflowPhase := req.AgentMode == "workflow_phase" (server.go:3215), then destructively rewrites the field a hundred lines later, purely to route the request down the standard agent path:

// Convert to multi-agent mode so it falls through to the standard agent path
req.AgentMode = "multi-agent"        // server.go:3321

The cost observer is constructed ~2200 lines further down (server.go:5561) and inferred its scope from that rewritten value:

inferCostScope(req.AgentMode, workflowPhaseID)   // ("multi-agent", "workflow-builder")

InferScope matches "multi-agent" against neither chief nor workflow, so it fell through to its default — ScopeChat. The routing rewrite and the cost attribution are ~2200 lines apart with no stated relationship, so nothing flagged that one silently destroyed the other's only input.

Separately, no signal on that path could have identified a Pulse turn even without the rewrite: a scheduled Pulse turn runs in the same session, with the same agent mode and the same workflow-builder phase id, as the workflow orchestration turns before it. Mode and phase genuinely cannot tell them apart.

Fix

Two parts, matching the two causes:

1. Honor the scheduler's explicit marker. New costobserver.ScopeForScheduledLLMRole(llmConfigSource) maps the stamp the scheduler already sets when it swaps in the Pulse/maintenance LLM (applyPulseLLMToReqMap → llm_config_source = "scheduled_pulse") onto a cost scope. scheduled_auto_improve maps to Pulse deliberately: Goal Advisor and Strategy Auditor are Pulse modules that merely run on the maintenance LLM. Returns "" for anything else so callers keep their own default rather than inheriting a wrong one.
2. Stop reading the rewritten mode. Scope resolution now uses isWorkflowPhase — captured before the rewrite — to pass the real workflow_phase mode, so scheduled orchestration turns resolve to builder rather than chat.

Priority order is marker first, then mode+phase, because only the marker can separate Pulse from workflow inside one scheduled session.

Consequence for existing data

Historical rows are not back-filled and must not be read as if they were. Every scheduled run before this fix has its Pulse-stage and workflow-orchestration spend sitting in chat. Any "Pulse vs goal" comparison over pre-2026-08-11 data understates Pulse and overstates chat. (Distinct from, and additional to, the 15,235 legacy unknown-scope rows, which stopped being produced on 2026-07-12 and are a separate pre-attribution artifact.)

Verification

• go build ./pkg/costobserver/ clean; go vet clean; gofmt clean.
• New tests (plat088_scheduled_role_scope_test.go): every scheduled-role marker maps to the right scope and unknown markers return ""; InferScope("workflow_phase", …) yields builder while the rewritten "multi-agent" yields chat — pinning the exact regression rather than implying it.
• The full ./cmd/server/... suite could not be run at commit time: another session's in-flight CreateSchedule/UpdateSchedule signature refactor leaves interactive_workshop_manager.go uncompilable. Confirmed by stashing that work that HEAD builds clean and that no build error is attributable to this change. Re-run the suite once that refactor lands.
• Not yet reverified live — needs a restart, then one scheduled run; expect its Pulse turns to appear under scope='pulse' and its orchestration turns under scope='builder', with neither in chat.

Acceptance

• A scheduled run's Pulse Gate / Review+Fix / Finalize turns are recorded as scope='pulse', not chat.
• A scheduled run's workflow-orchestration turns are recorded as scope='builder', not chat.
• chat contains only genuine interactive chat.