[Radar] Track cross-tool token cost attribution as adjacent surface

[luoyuctl]

Background

Recent ecosystem scan surfaced a stronger adjacent surface around cross-tool token cost attribution for AI coding assistants. agenttrace already reports local session cost, token totals, tool failures, health, anomalies, and CI evidence. The overlap is local session data and cost visibility; the product difference is whether users primarily want diagnostic session health or detailed budget attribution across many coding tools, projects, models, and task types.

Evidence

• Tavily ecosystem scan on 2026-05-04 surfaced getagentseal/codeburn: https://github.com/getagentseal/codeburn
• The search snippet describes token usage, cost, and performance tracking across 16 AI coding tools.
• The snippet says CodeBurn reads session data directly from disk, prices calls with LiteLLM, and reports spending by task type, model, tool, project, and provider.
• The same Tavily scan surfaced Anthropic issue #49588 requesting programmatic Claude Code session token usage for cost attribution: [FEATURE] Expose session token usage to MCP servers and hooks for cost attribution anthropics/claude-code#49588
• That issue describes phase-level token measurement needs for agentic development workflows and notes that debug logs / OpenTelemetry contain usage and cost fields, but require extra plumbing.
• Duplicate check for cost attribution token cost coding tools local session found no exact open issue.

User value

Users evaluating agent sessions may need to explain not only whether a session was unhealthy or failure-prone, but also which tool, model, project, phase, or workflow drove spend across a local coding-agent stack.

Adoption rationale

Tracking this signal helps agenttrace keep its Cost Audit and Developer experience positioning precise. It may inform future report fields, parser metadata, or docs language without turning agenttrace into a full budget management product.

Suggested scope

• Keep as radar until there is user evidence that cost attribution should become a first-class agenttrace workflow.
• Compare current JSON/HTML overview fields against the attribution dimensions surfaced by adjacent tools: provider, tool, model, project, task type, and phase.
• If promoted, split into a focused product/parser issue such as clearer provider/model fields in JSON exports or a report section for top cost drivers.
• Prefer local session evidence and existing report outputs before adding external pricing or budgeting features.

Non-goals

• Do not add a billing dashboard from this radar item.
• Do not add remote telemetry, hosted storage, or account-level spend tracking.
• Do not change current cost calculations without a separate compatibility and pricing-source review.
• Do not claim exact cost attribution where upstream session logs only support estimates.

Acceptance criteria

• Maintainer decides whether this remains radar, informs docs/positioning, or becomes a focused product/parser task.
• Any promoted work names the minimum attribution dimension needed by users.
• Follow-up work preserves local-first privacy and keeps estimated-cost wording precise.
• Related adjacent tools and upstream usage-signal gaps are recorded before implementation begins.

Suggested lane

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

Risk

Medium. Cost attribution is valuable but can sprawl into pricing engines and budget management. The low-risk path is to track the signal and only promote narrow report/export improvements with evidence.

Source

source/radar: Tavily ecosystem scan and duplicate check on 2026-05-04.

[luoyuctl]

Additional ecosystem validation (2026-05-04):

MCP search refreshed the cross-tool cost-attribution signal and surfaced more adjacent examples:

• mm7894215/TokenTracker: https://github.com/mm7894215/TokenTracker
• getagentseal/codeburn: https://github.com/getagentseal/codeburn
• cc-switch discussion mentioning cross-tool token/cost/reasoning/subagent activity: TokenTelemetry — See exactly what your Claude Code, codex, gemini cli and 6 more coding agents sessions cost, think, and do (100% local, open source) farion1231/cc-switch#2431

Interpretation: the strongest repeated signal is local cross-tool usage and cost visibility, not a need for hosted budget management. This keeps #103 as radar. If promoted, the smallest agenttrace-shaped follow-up should compare current JSON/HTML overview fields against provider/tool/model/project attribution, while preserving estimated-cost wording and local-first behavior.

Event: Evidence
Actor: Product
Scope: #103
State change: none
Evidence: MCP search results for cross-tool token/cost/session observability on 2026-05-04.
Next owner: Human

[luoyuctl]

Additional ecosystem validation (2026-05-04):

MCP search again reinforced #103 rather than producing a distinct new opportunity:

• bhvbhushan/codeledger: token-level cost intelligence for AI coding sessions.
• Dicklesworthstone/coding_agent_usage_tracker: multi-assistant usage tracking across Codex, Claude, Gemini, Cursor, and Copilot.
• androidZzT/cc-statistics: native dashboard for Claude Code, Gemini CLI, Codex CLI, and Cursor token/cost tracking.

Product routing: keep as radar. The repeated pattern is cross-tool cost/usage attribution; still no need to open a new implementation issue until the maintainer chooses a narrow agenttrace workflow such as provider/tool/model/project cost-driver fields in existing reports.

Event: Evidence
Actor: Product
Scope: #103
State change: none
Evidence: dedicated MCP search, Tavily unavailable in this session.
Next owner: Human

[luoyuctl]

Additional ecosystem validation (2026-05-04):

Dedicated MCP search surfaced another adjacent research signal:

• AI Observability for Developer Productivity Tools: Bridging Cost ...: https://arxiv.org/html/2604.17092v1
• Search summary describes AI observability for developer productivity tools with real-time token tracking and configurable model/cost visibility.

Duplicate checks for the paper title and developer productivity tools token tracking model cost observability found no exact open issue. This still reinforces #103 rather than requiring a new implementation issue.

Product routing: keep as radar. The agenttrace-shaped opportunity remains local-first cross-tool token/cost attribution in existing reports, not hosted real-time monitoring.

Event: Evidence
Actor: Product
Scope: #103
State change: none
Evidence: dedicated MCP search; Tavily unavailable in this session.
Next owner: Human

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #103
State change: none; keep lane/radar,status/needs-human
Evidence: Tavily search was attempted first but failed with a plan usage-limit error. Product then used search fallback and found adjacent cost/token tools and surfaces:

• getagentseal/codeburn: positioned around token usage, cost, and performance across AI coding tools: https://github.com/getagentseal/codeburn
• junhoyeo/tokscale: token usage and cost tracking across multiple AI coding agents: https://github.com/junhoyeo/tokscale
  Next owner: Product / Human

Product read: cross-tool cost attribution remains a visible adjacent category. Current agenttrace scope already covers local cost/tokens/time plus health/latency/tool-failure context, so this should stay radar unless a concrete gap appears, such as per-project cost attribution, task-type attribution, or a mismatch users can verify against upstream billing/export data.

Suggested next check: compare user requests against existing by_agent, by_model, session list sorting, and JSON output before creating parser or report tasks.

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #103
State change: none; keep lane/radar,status/needs-human
Evidence: Post-v0.4.6 master f354b39 now covers part of this radar through existing local report fields, verified with a clean build from /tmp/agenttrace-product-origin:

• --demo --overview -f json exposes by_agent and by_model cost groupings.
• The same report exposes incident_timelines plus summary.tool_authority and surfaces.authority_categories.
• --demo --overview -f json --baseline <local overview> emits deterministic baseline_comparison regression fields including cost_delta_pct, token_delta_pct, new_tool_authority_categories, and new_high_authority_tool_use.
  Next owner: Human

Product read: keep this as radar rather than promoting an implementation issue. agenttrace now has local session cost grouping and regression evidence, but it still does not intentionally cover project/task/provider/phase-level attribution or billing-dashboard workflows. A future promoted issue should name one narrow attribution dimension and preserve estimated-cost wording plus local-first behavior.

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #103
State change: none; keep lane/radar,status/needs-human
Evidence: Ecosystem scan on 2026-05-20 surfaced stronger current evidence for the cross-tool token/cost attribution radar:

• Claude Code official monitoring docs describe OpenTelemetry usage, cost, session, and tool-activity telemetry, while explicitly keeping official billing separate from telemetry-derived values: https://docs.claude.com/en/docs/claude-code/monitoring-usage
• Stanford Digital Economy Lab working paper How Do AI Agents Spend Your Money? reports large variance in token consumption for agentic coding tasks and frames token spend as a first-class debugging/evaluation concern: https://digitaleconomy.stanford.edu/publication/how-do-ai-agents-spend-your-money-analyzing-and-predicting-token-consumption-in-agentic-coding-tasks/
• Adjacent observability products such as AgentDog and Optelly continue to position token usage, tool calls, and cost visibility as core agent observability fields: https://agentdog.io/ and https://optelly.com/
  Product read: this strengthens the adoption rationale for local cost/token attribution, but it still does not justify a broad billing dashboard or hosted telemetry direction. agenttrace should keep this as radar until a narrow local-first workflow is chosen, such as provider/model/project cost-driver fields in existing reports, clearer estimated-cost confidence language, or export fields that downstream local tools can consume.
  Next owner: Human / Product monitor

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #103
State change: none; keep lane/radar,status/needs-human
Evidence: Ecosystem scan on 2026-05-20 could not use Tavily because no Tavily MCP tool was exposed in this session, so Product used search fallback and prioritized official/current sources:

• Anthropic Verified session-report plugin generates explorable HTML reports from local ~/.claude/projects transcripts, including tokens, cache efficiency, subagents, skills, and expensive prompts: https://claude.com/plugins/session-report
• Claude Code official session docs expose /export for readable current-conversation export and document controls for session persistence: https://code.claude.com/docs/en/sessions
• AgentSight positions a local Claude Code JSONL parser around usage analytics, cache efficiency, and cost estimation: https://agentsight.org/
  Product read: this strengthens the local report and token/cache attribution signal, especially around cache efficiency, subagent/skill attribution, and expensive prompt identification. It still should remain radar rather than becoming a broad billing or dashboard task. If promoted later, the smallest agenttrace-shaped scope should name one local report dimension, such as cache-efficiency evidence, subagent/skill attribution when available in local logs, or clearer cost-driver fields in existing JSON/HTML reports.
  Next owner: Human / Product monitor

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #103
State change: none; keep lane/radar,status/needs-human
Evidence: Ecosystem scan on 2026-05-20 could not use Tavily because no Tavily MCP tool was exposed in this session, so Product used search fallback and prioritized current local-first sources:

• ccusage now documents a Codex session report that groups Codex CLI usage by local session, last activity, model switches, tokens, and cost columns: https://ccusage.com/guide/codex/session
• ccusage all-source docs describe local usage-file reporting across Claude Code, Codex, OpenCode, Amp, and pi-agent: https://ccusage.com/guide/all-reports
• Token Tracker positions a local-first dashboard across multiple AI coding agent CLIs, including Claude Code, Codex CLI, Cursor IDE, Gemini CLI, OpenCode, OpenClaw, Hermes Agent, Kimi Code, oh-my-pi, and others: https://www.tokentracker.cc/
  Product read: this strengthens the user value for local session-level cost/token attribution and multi-source comparison. It still does not justify a broad billing dashboard, hosted telemetry, or external pricing integration. If promoted later, the smallest agenttrace-shaped scope should name one existing-report improvement, such as clearer session-level cost-driver fields, model/provider grouping confidence, or export fields that downstream local dashboards can consume.
  Next owner: Human / Product monitor