[Radar] Track local coding-agent session search as adjacent product surface

[luoyuctl]

Background

A current adjacent tool, coding-agent-search (cass), focuses on indexing and searching local coding-agent session history across many agent tools. agenttrace currently emphasizes observability, health, cost, anomalies, reports, and CI evidence. The overlap is local session discovery and multi-agent history, while the product promise differs: searchable timeline versus diagnostic overview.

Evidence

• Tavily ecosystem scan on 2026-05-04 surfaced Dicklesworthstone/coding_agent_session_search: https://github.com/Dicklesworthstone/coding_agent_session_search
• The README describes a unified TUI and CLI for indexing and searching local coding-agent history.
• Supported sources listed include Codex, Claude Code, Gemini CLI, Cline, OpenCode, Cursor, Aider, Copilot CLI, OpenClaw, Kimi Code, Qwen Code, and others.
• The project exposes commands such as index --full, index --watch, search --robot, status, health, capabilities, introspect, and sessions.
• Duplicate checks for session search index local agent history, coding-agent-search cass, and local session history searchable timeline found no exact open issue.

User value

Users with many local agent sessions may want to answer a different first question than agenttrace currently optimizes for: not only which session is unhealthy, expensive, or failure-prone, but also where a prior conversation, file, tool call, or workflow appeared across history.

Adoption rationale

Tracking this adjacent surface helps keep agenttrace's roadmap honest about its scope. It may inform future report navigation, session lookup, or integration language without diluting the current Reliability value and CI evidence focus.

Suggested scope

• Keep as radar until there is user evidence that search belongs in agenttrace rather than remaining a separate adjacent tool category.
• Compare current agenttrace --overview, session list, and JSON output against the searchable-timeline use case.
• If promoted, split into a narrow product issue such as searchable report navigation, session lookup by name/path/model, or export fields that another search tool can consume.
• Prefer interoperability and clear positioning before adding a full search index.

Non-goals

• Do not add a search engine or watcher from this radar item.
• Do not expand parser scope solely to match another tool's source list.
• Do not change current health, cost, or CI gate behavior.
• Do not position agenttrace as a session archive until product scope is decided.

Acceptance criteria

• Maintainer decides whether this remains radar, informs docs/positioning, or becomes a focused product task.
• Any promoted task names a concrete user workflow and the minimum artifact needed.
• Follow-up work preserves agenttrace's current local-first observability focus.
• Duplicate or adjacent tool references are recorded before implementation work begins.

Suggested lane

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

Risk

Medium. Search can become a large product surface quickly. The low-risk path is to track the signal and only promote a narrow workflow with evidence.

Source

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

[luoyuctl]

Additional ecosystem validation (2026-05-04):

Dedicated MCP search refreshed the local session-search signal and surfaced another directly adjacent repository:

• Dicklesworthstone/coding_agent_session_search: https://github.com/Dicklesworthstone/coding_agent_session_search
• Search result describes a unified TUI/CLI for indexing and searching local coding-agent session history across 11+ providers including Codex, Claude, Gemini, Cursor, and Aider.

Product routing: keep as radar. This reinforces the local session discovery/search surface, but it should not become implementation work until a concrete agenttrace workflow is chosen, such as searching within already parsed sessions, jumping from search results to TUI detail, or exporting searchable metadata without prompt content.

Event: Evidence
Actor: Product
Scope: #101
State change: none
Evidence: dedicated MCP search; Tavily unavailable in this session.
Next owner: Human

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #101
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 Dicklesworthstone/coding_agent_session_search described as a unified TUI/CLI for indexing and searching local coding-agent session history across Codex, Claude, Gemini, Cursor, Aider, and other providers: https://github.com/Dicklesworthstone/coding_agent_session_search
Next owner: Product / Human

Product read: this reinforces that local session search is an adjacent surface, but it is still not enough to create an implementation task. agenttrace already has sorting/filtering and diagnostics, so any future scope should be framed around user journeys agenttrace uniquely owns: cost/time/health triage plus trace evidence, not generic full-text search.

Suggested next check: keep watching for users asking to find past sessions by prompt/tool/model/error evidence inside agenttrace; only promote from radar if there is direct user feedback or a narrow acceptance path.

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #101
State change: none; keep lane/radar,status/needs-human
Evidence: Current master f354b39 already covers a narrow lookup/triage subset without becoming a search product:

• Public README positions first-session triage around sorting and filtering by cost, duration, health, failures, anomalies, model, source, or text search.
• JSON overview exposes structured recent_sessions, by_agent, by_model, failure families, incident timelines, and tool surfaces that downstream tooling can index if needed.
• No current scope adds a watcher, full-text index, semantic archive, or session-history search daemon.
  Next owner: Human

Product read: keep this as radar. If promoted later, prefer a narrow agenttrace-owned workflow such as jumping from a report/session list to diagnostic evidence, or exporting searchable metadata without prompt content. Avoid a generic local history search engine unless direct user feedback asks for it.

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #101
State change: none; keep lane/radar,status/needs-human
Evidence: Ecosystem scan on 2026-05-20 found additional local session-history tools that reinforce this radar without creating a new implementation path:

• tobitege/codlogs: read-only Codex session CLI/desktop browser for finding sessions, exporting JSONL to Markdown/HTML, filtering by folder, and summarizing token usage for selected or filtered sessions: https://github.com/tobitege/codlogs
• hochonin93.codex-token-usage: VS Code extension that reads local Codex JSONL logs and summarizes local token usage by time range: https://marketplace.visualstudio.com/items?itemName=hochonin93.codex-token-usage
• AgentsView: local-first session browser/search/analytics surface for Claude Code, Codex, and other agent sessions, including FTS search and usage reporting: https://www.agentsview.io/
  Product read: these signals strengthen the adjacent category around local session browsing/search/export/usage analytics. agenttrace should still avoid becoming a generic session archive. If promoted later, the smallest agenttrace-shaped scope should stay tied to diagnostic evidence already owned by agenttrace: find sessions by health, cost, tool failure, incident timeline, model/source, or exported metadata without prompt-content indexing by default.
  Next owner: Human / Product monitor

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #101
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 2.1.141 adds claude agents --cwd <path> to scope the session list to a directory, and /feedback can include recent sessions for issues spanning more than the current session: https://code.claude.com/docs/en/changelog
• Claude Code Week 20 docs describe agent view as one screen for every Claude Code session, including running, blocked, and done sessions, with attach/return navigation: https://code.claude.com/docs/en/whats-new/2026-w20
  Product read: this strengthens the local session-list and session-discovery radar with official upstream evidence. It still does not justify building a generic search/archive product. If promoted later, the smallest agenttrace-shaped path should remain diagnostic lookup over completed local evidence, such as finding sessions by health, cost, tool failures, model/source, incident timeline, or export metadata without prompt-content indexing by default.
  Next owner: Human / Product monitor

[luoyuctl]

Event: Evidence
Actor: Product
Scope: #101
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/session-browser sources:

• jazzyalex/agent-sessions positions a native local session browser across Codex CLI, Claude Code, OpenCode, Gemini CLI, GitHub Copilot CLI, and OpenClaw, including search/filter, archive, resume, analytics, and limit tracking: https://github.com/jazzyalex/agent-sessions
• codex-session-history is packaged as an agent skill for searching Codex local session history from SQLite and session_index.jsonl: https://eliteai.tools/agent-skills/codex-session-history
• Qwen Code docs now show a user-facing /export json <session-id> flow for exporting conversation history, which reinforces the broader local-session retrieval/export pattern but still does not provide an agenttrace parser fixture: https://qwenlm.github.io/qwen-code-docs/zh/showcase/office-export-conversation/
  Product read: this strengthens the adjacent local session browsing/search/export category already tracked by [Radar] Track local coding-agent session search as adjacent product surface #101 and overlaps with [Radar] Track Qwen Code export and dual-output session surfaces #237 for Qwen export surfaces. It still should not become a generic search/archive implementation task. If promoted later, keep the scope diagnostic and local-first: find sessions by health, cost, tool failures, model/source, incident timeline, or metadata export without prompt-content indexing by default.
  Next owner: Human / Product monitor