[Radar] Track upstream session-log fidelity gaps for cost and tool failures

[luoyuctl]

Background

External agent ecosystems keep surfacing a shared reliability gap: local session logs are useful, but some tools either omit cost details, undercount usage, or report tool failures without enough error context. agenttrace should track this as a radar item before deciding whether parser warnings, docs notes, or fixture work are needed.

Evidence

• Claude Code issue #23254 reports subagent token usage is shown by /cost but not persisted to local session JSONL, which limits downstream local reporting: Persist subagent token usage to session JSONL files anthropics/claude-code#23254
• Claude Code issue #27361 reports JSONL transcripts can miss final message stop events, making output token counts unreliable for local cost tools: [BUG] JSONL session transcripts missing final message_stop event, output token counts unreliable anthropics/claude-code#27361
• Claude Code issue #55430 shows large repeated tool outputs and full-file writes dominating token usage in a real session, with transcript-size evidence: Excessive token consumption: pre-push hook logs + Write vs Edit pattern anthropics/claude-code#55430
• Claude Code issue #51218 reports prompt-cache expiry during active sessions causing sudden token spikes after idle periods: [BUG] Prompt cache expires during active session, causing massive token spikes on next prompt anthropics/claude-code#51218
• Codex issue #2420 reports the CLI can show tool failed while local logs do not include the underlying error detail: Logs Don't Show Errors On Tool Failure openai/codex#2420
• Duplicate checks for Claude Code JSONL token usage subagent cost, tool failure logs Codex CLI, cache token spike Claude Code, and session log fidelity upstream cost accounting found no exact open issue in this repository.

User value

Users reading local reports need to know when a metric is derived from complete local evidence versus when the upstream session log may be missing cost, failure, or final-output details.

Adoption rationale

Clear confidence boundaries improve Developer experience and Reliability value. They help users trust agenttrace reports while avoiding false precision when an upstream tool did not persist enough evidence.

Suggested scope

• Keep this as radar until at least one minimal public fixture or reproducible local sample is available.
• Decide whether agenttrace should add parser-level confidence notes for known upstream log gaps.
• Decide whether docs should mention source-specific limitations for cost and tool-failure attribution.
• If fixture evidence becomes available, split concrete parser or product issues by source tool.

Non-goals

• Do not infer private billing data that is not present in local logs.
• Do not upload or request user transcripts.
• Do not change parser behavior without fixture-backed evidence.
• Do not treat unrelated model-routing or hosted observability products as direct requirements.

Acceptance criteria

• Maintainer decides whether this remains radar, becomes docs guidance, or splits into parser/product issues.
• Any follow-up issue names the source tool and the specific missing or unreliable field.
• Follow-up acceptance criteria require local fixture evidence or a reproducible public sample.
• agenttrace public copy avoids overclaiming exact cost attribution where the upstream log is known to be incomplete.

Suggested lane

lane/radar, priority/P2, status/needs-human

Risk

Medium. Overreacting could add noisy warnings; ignoring the signal could make reports look more precise than the underlying logs support.

Source

source/radar: Tavily scan of public GitHub and ecosystem signals on 2026-05-04.