-
Notifications
You must be signed in to change notification settings - Fork 3
plat 036
| Coordination | Value |
|---|---|
| Assigned agent | Codex |
| Ticket state | runtime_reverify |
| Last synchronized | 2026-08-07 |
- Priority: P2
- Owner: shared cost telemetry writer and its UI/API projection
-
Source workflow:
Workflow/tectonicusadaytrading -
Source finding:
HARNESS-COST-CONTEXT-PCT-SATURATION
context_usage_percent reports 100% for every recorded call. The recorder
divides cumulative context_window_usage by a single-call model context window,
so the numerator and denominator describe different scopes. Tectonicus recorded
643,364 cumulative tokens against a 200,000-token single-call window and still
rendered 100% everywhere.
Pulse and the UI cannot distinguish normal context pressure from saturation. This produces false LLM/Ops evidence and can drive unnecessary truncation, model, or workflow decisions.
Persist and display either a same-call numerator/denominator pair, or mark the metric unavailable when only cumulative usage exists. Never clamp an incomparable ratio to 100%. Keep cumulative-session utilization as a separately named metric if it is useful.
Coding-CLI adapters (Codex CLI and Claude Code, interactive and structured)
now mark transcript/CLI usage as not being a current-context snapshot.
mcpagent keeps that aggregate for pricing, but clears the context numerator
and omits the percentage/denominator from emitted telemetry. Native API
providers retain the existing current-prompt behaviour.
This deliberately reports context utilization as unavailable rather than inventing a percentage from aggregate accounting. The focused mcpagent test uses the observed 643,364-token/200,000-token shape and proves it no longer emits a saturated percentage; Claude Code and Codex CLI transcript tests prove the marker is carried across the adapter boundary.
The 2026-08-07 Upwork review tried to turn an absent coding-CLI
context_usage_percent into a telemetry regression. Reviewer guidance now
states that this absence is the intended representation when only cumulative
transcript usage exists. It must not be filed as a platform defect or used as
proof of context pressure.
Two calls with different actual context utilization produce different percentages; a cumulative-only record is labelled unavailable rather than saturated; and a Pulse cost summary does not use an unavailable value as optimization evidence. A fresh coding-CLI workflow run is still required to complete runtime verification.
Auto-synced from docs/ on main. Edit there, not here.