agent-insights

Cross-tool version of Claude Code's builtin /insights. Everything runs locally; no session data leaves the machine.

SKILL_DIR below means this skill's directory; SCRIPT means python3 <SKILL_DIR>/scripts/agent_insights.py. All intermediate files live in ~/.agent-insights/ (override with --data-dir on every subcommand, e.g. for tests).

Arguments

One optional argument: number of days to analyze. Default 30, clamp to 1..365. Non-numeric input: tell the user, then proceed with 30.

Pipeline

Step 0 — Detect runtime

SCRIPT detect

Detects whether this skill is running on the host machine (where the user's agent logs live) or inside a sandboxed runtime such as Claude Cowork, whose Bubblewrap container does not mount the host filesystem. Reads the JSON result:

• If sandboxed is false, continue to Step 1.

• If sandboxed is true, the host machine's agent log directories are not reachable from this runtime. Do not run a full report from here: it would be empty or cover only the sandbox's own session. Instead, tell the user:

  • the detected runtime and confidence, and the signals that fired;
  • the recommendation verbatim (run /agent-insights from a non-sandboxed agent with direct host access, e.g. Claude Code in the terminal);
  • reachable_sources (what little, if anything, is visible from inside the sandbox).

  Then stop, unless the user explicitly asks to analyze only what is reachable inside the sandbox — only then proceed to Step 1.

Step 1 — Scan

SCRIPT scan --days <N>

Discovers sessions across all supported tools, filters to the window, dedupes, drops non-substantive sessions (no user messages or < 1 active minute) and prior insight-analysis sessions, caches per-session metadata, and prepares truncated transcripts for sessions that don't have cached facets yet (max 50 per run; re-runs only pay for new sessions).

Output (stdout JSON): runtime (same shape as Step 0, repeated here as a safety net), sources_detected, sessions_kept, per_tool, facet_extractions_needed, facet_extractions_deferred (uncached sessions beyond the 50-per-run cap), and batches (each with transcript_files and an output_file).

• Exit code 2 with an error message means zero sessions were found across every source. Report that to the user and stop; nothing to analyze. If the error's runtime.sandboxed is true, the cause is the sandboxed runtime (Step 0) — relay its recommendation instead of a generic "no sessions" message.
• Tell the user what was detected (tools + session counts) before continuing.

Step 2 — Facet extraction (skip when facet_extractions_needed is 0)

Read <SKILL_DIR>/references/prompts.md (section 1) for the facet extraction prompt.

If the running agent supports subagents (e.g. Claude Code's Task tool, Cursor): spawn one general-purpose subagent per batch, all in parallel in a single message. Each subagent prompt must contain:

1. The facet extraction prompt from prompts.md, verbatim.
2. The batch's transcript_files list. Each file starts with a === SESSION <session_key> === header.
3. Output instructions: read every transcript file; write ONE JSON array to the batch's output_file (use the Write tool), where each element is {"session_key": "<from the file header>", "facets": {<facet object per the schema>}}; return only a count of sessions processed.

If the running agent does not support subagents (e.g. OpenCode): do NOT attempt a parallel subagent run. Process the batches yourself in the main context, one batch at a time, sequentially: for each batch, read its transcript files, apply the facet extraction prompt, and write the same JSON array (shape as in item 3 above) to the batch's output_file before moving on to the next batch.

In either mode, if a batch fails or produces invalid JSON, retry it once; on second failure continue without that batch (those sessions still count in the deterministic stats).

Step 3 — Aggregate

SCRIPT aggregate --days <N> --analysis-model "<your model id>"

Pass --analysis-model with the model you (the main agent running this skill) are currently executing as. Use the base model id without context-window or runtime suffixes (e.g. claude-opus-4-8, not claude-opus-4-8[1m]). This records which model produced the analysis (facets + narrative) so reports can be benchmarked against an agreed standard model. It is the analyzing model, distinct from the per-session models found in the scanned logs. If you genuinely cannot determine your own model id, omit the flag (it records unknown).

Validates and ingests the facet batch files into the cache, then merges all cached sessions in the window into ~/.agent-insights/aggregate.json (also printed to stdout): version (the skill's metadata.version), analysis_model, totals (incl. total skill_invocations), per-tool breakdown (sessions, messages, hours, tool calls, skill invocations, models used), goal/outcome/satisfaction/friction/expertise_level/model/skills_used distributions, and narrative_context (session briefs + friction details) for the next step. Skill invocations are counted deterministically during the scan (no facet extraction): the Skill tool in Claude Code / Cowork logs, leading /name slash commands in Cursor and OpenCode user messages, and skill.invoked events in Copilot CLI / JetBrains agent logs. Names merge across tools, so /spec-lint and the spec-lint skill count as one.

If the scan reported facet_extractions_deferred > 0, only part of the window has facets. Loop Steps 1-3 until deferred is 0 — each round's scan prepares batches for the next 50 uncached sessions (already-cached ones cost nothing) — so the narrative is grounded in the full window. Then continue.

Step 4 — Narrative

Read <SKILL_DIR>/references/prompts.md (section 2) and follow it: it defines every narrative section's key and shape, plus the grounding rules. Using the aggregate output from Step 3, write ALL of those sections to narrative.json in the data dir (~/.agent-insights/narrative.json unless --data-dir was overridden). Also set a top-level version key, copied from the aggregate's version value, so the narrative file records the skill release that produced it.

Step 5 — Render

SCRIPT render

Produces a self-contained HTML report at ~/.agent-insights/report-YYYY-MM-DD_<days>-days.html, where <days> is the analyzed window taken from the aggregate (file mode 0600; it contains transcript snippets). The report header shows the skill version (read from the aggregate). Prints the path.

Step 6 — Tell the user

Output, in this order:

1. A stats line: N sessions · N user messages · Nh · across <tools detected> plus the date range.
2. The At a Glance summary: the persona one-liner from at_a_glance first ("You are a <level> <user_type>."), then the four parts as short markdown sections: What's working / What's hindering you / Quick wins / Ambitious workflows.
3. The report path, suggesting open <path> to view it.
4. One line inviting the user to dig into any section.

Notes

• The scan prints adapter warnings to stderr; surface them to the user only if a tool they expected is missing from sources_detected.
• Storage locations, format details, and parsing caveats per tool: <SKILL_DIR>/references/data-sources.md (read when debugging a missing/empty source).
• Requires Python 3 (stdlib only). If python3 is unavailable, stop and tell the user.