[Radar] Track local multi-agent observability dashboards as adjacent surface

[luoyuctl]

Background

Recent ecosystem scan shows a small but relevant adjacent product surface: local dashboards that observe multiple AI coding assistants as they run, including tool decisions, logs, token/cost signals, and terminal-style activity views. agenttrace currently focuses on local session traces, health, cost, anomalies, reports, and CI evidence. The overlap is local-first observability; the product difference is live multi-agent monitoring versus session/report diagnosis.

Evidence

• Tavily ecosystem scan on 2026-05-04 surfaced tobilg/ai-observer: https://github.com/tobilg/ai-observer
• The search snippet describes unified local observability for Claude Code, Gemini CLI, and OpenAI Codex CLI, including tool-decision and metrics/log fields.
• Tavily also surfaced tech4242/agenttop: https://github.com/tech4242/agenttop
• The search snippet frames it as an htop-style view for AI coding agents and notes differences in exposed tool names across Claude Code, Codex, and Gemini.
• Duplicate check for ai-observer agenttop local observability htop AI coding agents found no exact open issue.

User value

Users running several local coding agents may want an immediate answer to which agent is active, noisy, expensive, or repeatedly calling tools before they inspect a completed session report.

Adoption rationale

Tracking this signal helps keep agenttrace's positioning clear around local-first agent observability. It may inform future wording, parser priorities, or export compatibility without expanding scope into live process monitoring prematurely.

Suggested scope

• Keep as radar unless users ask for live monitoring, multi-agent comparisons, or better interoperability with local observability tools.
• Compare current agenttrace --overview, latest-session views, and JSON output against the live-dashboard use case.
• If promoted, split into a narrow task such as report fields for multi-agent comparison, parser metadata needed by dashboards, or README positioning.
• Prefer compatibility with existing reports before considering a live monitor.

Non-goals

• Do not build a live htop-style UI from this radar item.
• Do not add background watchers or process monitoring without separate product evidence.
• Do not expand parser support only to mirror another tool list.
• Do not weaken the current local session/report and CI evidence focus.

Acceptance criteria

• Maintainer decides whether this remains radar, informs docs/positioning, or becomes a focused product/parser task.
• Any promoted work names one concrete user workflow and the minimum existing artifact it builds on.
• Follow-up keeps local privacy and completed-session diagnosis as first-class constraints.
• Related external tools are recorded before implementation work begins.

Suggested lane

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

Risk

Medium. Live observability can become a separate product quickly. The low-risk path is to track the signal and only promote a narrow workflow with clear user evidence.

Source

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

[luoyuctl]

Additional radar validation (2026-05-04):

Tavily surfaced another adjacent local observability tool:

• matt1398/claude-devtools: https://github.com/matt1398/claude-devtools
• The search snippet describes a visual UI for inspecting Claude Code session logs, tool calls, token usage, subagents, and context window state.

Routing: still radar. This reinforces the local diagnostic dashboard signal, but it is narrower than agenttrace's multi-source/session-health scope because the surfaced evidence is Claude Code-specific. If promoted later, the smallest useful follow-up should compare whether agenttrace should expose richer completed-session drilldown fields rather than build a live or single-tool devtools UI.

[luoyuctl]

Additional ecosystem validation (2026-05-04):

Dedicated MCP search for 2026 agent observability surfaced more adjacent signals:

• 15 AI Agent Observability Platforms in 2026 result highlights limitations around simulation testing and semantic failure clustering.
• Agent Observability 2026: Evals, Traces, Cost Guide result frames common failures as tool failures, context failures, and runaway loops.
• Security-focused result reinforces local/privacy value when AI tooling may expose sensitive company data.

Product routing: keep #102 as radar. These signals support local completed-session diagnostics and explainability, but do not justify turning agenttrace into a hosted observability platform or broad eval product.

Event: Evidence
Actor: Product
Scope: #102
State change: none
Evidence: dedicated MCP search; Tavily unavailable in this session.
Next owner: Human

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #102
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 local multi-agent dashboard/session-manager examples:

• zanetworker/aimux: TUI dashboard for managing AI coding agent sessions, including status/model/tokens/cost: https://github.com/zanetworker/aimux
• asheshgoplani/agent-deck: terminal session manager for AI coding agents with token/cost tracking claims: https://github.com/asheshgoplani/agent-deck
  Next owner: Product / Human

Product read: this supports keeping local multi-agent dashboards on radar. It does not justify broadening agenttrace into session orchestration. agenttrace should preserve the observability/debugging boundary unless users ask for a narrow workflow that benefits the existing trace model, such as comparing live/current sessions with historical cost and health evidence.

Suggested next check: do not create an implementation issue until there is a concrete user journey, acceptance criteria, and no conflict with the local observability positioning.

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #102
State change: none; keep lane/radar,status/needs-human
Evidence: Current master f354b39 already covers the local completed-session side of this adjacent surface:

• README and launch kit position agenttrace around local history across multiple agent sources.
• --demo --overview -f json exposes by_agent, by_model, recent sessions, failure families, incident timelines, and tool surface evidence.
• CI/report work after v0.4.6 adds baseline-comparison and tool-authority evidence for completed-session regression review.
  Next owner: Human

Product read: keep #102 as radar. The useful boundary is historical observability and diagnosis across local agent logs, not live session orchestration, task routing, or a hosted monitoring dashboard. If users ask for more here, the narrow agenttrace-shaped follow-up would be richer completed-session drilldown or comparison, not a session manager.

[luoyuctl]

ok

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #102
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 sources:

• Claude Code changelog continues to expose subagent workflow changes, including forked subagents on external builds: https://code.claude.com/docs/en/changelog
• Claude Code subagent docs expose CLI-visible subagent configuration via claude agents: https://code.claude.com/docs/en/subagents
• Gemini CLI changelog includes built-in agent skills work in the current CLI surface: https://github.com/google-gemini/gemini-cli/blob/main/docs/changelogs/index.md
• OpenAI Codex releases continue to evolve session/thread and remote-control adjacent surfaces: https://github.com/openai/codex/releases
  Product read: current coding-agent CLIs are making agent, subagent, skill, session, and remote-control surfaces more explicit. This strengthens [Radar] Track local multi-agent observability dashboards as adjacent surface #102 as an adjacent observability-dashboard radar item, but it still does not justify turning agenttrace into live orchestration or a hosted dashboard. The agenttrace-shaped path, if promoted later, should remain completed-session comparison, report/export compatibility, or clearer local multi-source positioning.
  Next owner: Human / Product monitor

[luoyuctl]

Event: Handoff
Actor: Product
Scope: #102 -> #249
State change: none for #102; keep lane/radar,status/needs-human
Evidence: Product read-only check of origin/master at 6fb3917 found no evidence that agenttrace reverse-decodes Claude Code ~/.claude/projects/<encoded-cwd> into cwd, but also found no structured cwd metadata extracted from Claude JSONL body records. The external comment on #102 is therefore actionable only as a source-specific parser metadata correctness check, not as a live dashboard task.
Handoff to: Parser / Quality
Reason: verify the authoritative Claude Code cwd boundary before any future multi-agent comparison or report/export metadata uses project identity.
Expected output: focused fixture/test or parser note proving any exposed cwd/project metadata comes from JSONL body cwd, and that encoded project directory reversal is not used as authoritative cwd.
Acceptance check: #249 acceptance criteria; go test ./... and relevant CI checks pass.
Context: #249
Guardrail: Do not build a live multi-agent monitor, watcher, orchestration UI, prompt-content index, or new source support from this radar.