plat 226

← Pulse platform index

PLAT-226 — orchestrator/Pulse/Builder overhead cost is provably merged, and the disambiguating signal doesn't exist yet

Coordination	Value
Assigned agent	Claude Code
Ticket state	root cause confirmed in code — fix deferred, needs live instrumentation design
Last synchronized	2026-08-29

• Priority: P1 in the queue, low severity per the finding itself.
• Owner: agent_go/pkg/orchestrator/context_aware_bridge.go — the shared event bridge every BaseOrchestrator-based session (workflow execution, Pulse review, Pulse fixer, the interactive Builder) constructs through the same single call site (base_orchestrator.go:167, NewContextAwareEventBridge).
• Findings: Upwork PUL-A8AB0913 (cost-ledger overhead attribution) and PUL-E717A5E1 (aggregated call/context telemetry) — same underlying mechanism, grouped here per the audit's own framing.

What the findings established

PUL-A8AB0913's own evidence (re-checked this pass, not re-derived): the per-execution cost ledger (costs/execution/__ungrouped__/<date>.json) is correctly ID-keyed per execution now — that part of an earlier diagnosis is resolved. But every execution record still carries exactly one attribution row, workflow_orchestrator:workflow_orchestrator, even when the actual spend is a mix of the real run-level orchestrator, a same-day Pulse pass, and Builder/interactive work. The finding's own reproduction on 2026-08-12.json shows this directly: a 17:25:31Z record that post-dates a 16:57:51Z run completion — i.e. this is a Pulse pass's own cost — booked under the identical workflow_orchestrator label as genuine orchestrator spend. PUL-E717A5E1 is the same symptom from a different angle: aggregated telemetry prevents item-level diagnosis of a workload's actual cache-hit composition.

Root cause, confirmed in code

context_aware_bridge.go's step-based token-attribution path (persistTokenUsage, around line 961) falls back to a single hardcoded bucket whenever neither currentPhase nor currentStepID is set:

attributedPhase := currentPhase
attributedStepID := currentStepID
if strings.TrimSpace(attributedPhase) == "" {
    if strings.TrimSpace(attributedStepID) != "" {
        attributedPhase = "execution_only"
    } else {
        attributedPhase = "workflow_orchestrator"
        attributedStepID = "workflow_orchestrator"
    }
}

This branch is reached by any token event that never had a plan-step phase/ID pushed onto the bridge — which includes the real run-level orchestrator, but also Pulse review turns, Pulse fixer turns, and the Builder/interactive workshop session, none of which are workflow-plan steps and so never call PushContext with a phase/step at all. All of them collapse onto the identical "workflow_orchestrator" label.

Checked for an existing disambiguating signal before concluding a new one is needed. currentAgentName (c.currentAgentName, already read into scope at the point this fallback runs, already surfaced elsewhere as orchestrator_agent_name event metadata) looked like a promising existing field to key off instead of hardcoding a shared bucket. It is not: searching every PushContext/PushContextRich call site in this codebase found exactly one caller that sets an agent name at all (background-task, in interactive_workshop_manager.go). Pulse review, Pulse fixer, and the main Builder session never call PushContext — so currentAgentName is empty for them too, for the identical reason currentPhase/currentStepID are empty. There is no existing signal anywhere in this bridge's current inputs that can distinguish these three sources from each other or from the real orchestrator.

Why this ticket stops at diagnosis rather than shipping a fix

A real fix is not a one-line change to the fallback branch — there is nothing to branch on yet. It requires deliberately calling PushContext/PushContextRich (or an equivalent new "session kind" field threaded into ContextAwareEventBridge) at the start of Pulse review turns, Pulse fixer turns, and the Builder/interactive workshop session — each a separate call site, not a shared one, since they're set up independently (the Builder session config is in interactive_workshop_manager.go; Pulse review/fixer dispatch is in pulse_worklist.go and related files). Missing one of the three, or mislabeling which existing bucket becomes "the real orchestrator" once the others are split out, produces a cost ledger that is confidently wrong rather than honestly merged — worse than the current state, and not something this investigation could verify without a live Pulse pass, a live Builder session, and a live scheduled run to check the resulting buckets against. That live verification loop is exactly the kind of check this session cannot perform from a static repo checkout, which is why PLAT-224 reached the same kind of stopping point for a different reason (unknown external-tool schema) and why this ticket does the same for this one.

Concrete next steps, for whoever picks this up next

1. Add a sessionKind (or reuse/extend currentAgentName, now that its real gap is confirmed) that Pulse review, Pulse fixer, and the Builder session each set once at session start via a small, explicit call — not inferred from the absence of phase/step, which is exactly what silently merges them today.
2. Change the fallback branch to attribute by that signal when phase/step are both empty, keeping "workflow_orchestrator" only for the case where none of the three explicit signals are set (the genuine run-level orchestrator, which is the one caller that legitimately never has a plan-step phase).
3. Verify against a real same-day run that includes a scheduled full run, a Pulse pass, and Builder activity — the exact reproduction shape both findings already used — and confirm the resulting costs/execution/__ungrouped__/<date>.json carries three separable attribution rows instead of one merged one.

Verification

N/A — no files changed. This is a diagnosis record with a concrete implementation plan for the next pass, not a shipped fix.