feat: add AI-powered Insights with multi-agent generation

[wesm]

Summary

• Insights feature for AI-generated analysis of agent sessions, supporting Claude, Codex, and Gemini as generation backends
• Full-stack: insights SQLite table with FTS5, Go HTTP handlers (GET/DELETE /api/v1/insights/{id}, POST /api/v1/insights/generate with SSE streaming), prompt builder with session context, and multi-agent CLI dispatch
• Date range analysis: insights can cover a single day or a date range (e.g., a week). DB schema uses date_from/date_to columns. The generate endpoint validates both fields and supports date_to >= date_from
• Mode-driven UI: type selector with three modes -- Daily Activity (single date), Date Range Activity (from/to inputs with 7-day and 30-day presets), and Agent Analysis (single date). Mode is derived reactively from store state
• Svelte 5 Insights page with sidebar controls (type, date, project, agent), concurrent generation tasks with live status spinners, date display on in-progress tasks, and markdown content viewer with delete support
• Structured API errors: ApiError class with status field for programmatic error handling; empty response bodies fall back to "API <status>" message
• ListInsights capped at 500 rows with created_at DESC index to prevent unbounded queries
• internal/insight package: prompt construction from session data with date-aware text ("Date" vs "Date Range"), streaming response parsing for Claude/Codex/Gemini CLI output
• Session breadcrumb bar showing project name, agent badge (color-coded), and session start time
• Header navigation with always-visible Sessions/Insights buttons and active state highlighting
• Design polish: Inter font, refined color tokens for light/dark themes, tighter typography and spacing

Test plan

• go test -tags fts5 ./... -- all Go tests pass (insights CRUD, filters, 500-row cap, date range round-trip, prompt builder with single/range dates, server handler validation)
• npx vitest run -- all 360 frontend tests pass (insights store, date sync, mode switching, generate payloads, ApiError handling with empty-body fallback)
• npx tsc --noEmit -- no type errors
• go vet ./... -- clean
• Manual: generate a Daily Activity insight, verify single date in list and content header
• Manual: switch to Date Range Activity, verify from/to inputs and presets appear, generate and verify range displays
• Manual: verify all insights appear in the list regardless of selected generation dates
• Manual: verify in-progress task items show the date being analyzed
• Manual: delete an insight, verify removal from list and content area

Generated with Claude Code

[roborev-ci]

roborev: Combined Review (6af7802)

Summary Verdict: 5 findings identified (2 High, 3 Medium) encompassing broken access control, build-breaking syntax, information leakage, DoS vulnerabilities
, and a frontend race condition.

High

1. Unauthenticated API Endpoints (Broken Access Control)

• Files: internal/server/server.go:109, internal/server/insights.go:131, internal/insight/generate.go:1 03
• Description: New endpoints POST /api/v1/insights/generate (which executes local AI CLIs) and DELETE /api/v1/insights/{id} are exposed without authorization checks. If the service is reachable beyond trusted localhost, attackers can trigger local command execution or perform destructive
  actions.
• Remediation: Add authentication middleware and per-resource authorization checks before generation and deletion. Restrict binding to localhost if intended to be local-only, and enforce origin checks for browser-originated calls.

2. Invalid Svelte Syntax & Potential Stored XSS

• Files:
  frontend/src/lib/components/insights/InsightsPage.svelte:219, 343, 349
• Description: A corrupted tag (likely from a stray search-and-replace) appears as { @frontend/index.html renderMarkdown(insights.selectedItem .content)} instead of a valid Svelte HTML injection directive. This invalid syntax will completely break the frontend build. Furthermore, because this renders persisted AI-generated content, it presents a Stored XSS risk if the output is not rigorously sanitized.
• Remediation: Fix the syntax error by reverting to S
  velte's proper {@html ...} directive. Ensure renderMarkdown employs a robust HTML sanitizer (e.g., DOMPurify) and apply a Content Security Policy (CSP) to block inline script execution.

Medium

3. Information Leakage via Error Paths

• Files: internal/ insight/generate.go:72, 85, internal/server/insights.go:52, 199
• Description: Error handling paths directly return or log raw CLI stderr/stdout and database error strings. This behavior can leak sensitive prompt/session content, internal paths, or
  operational details to the client.
• Remediation: Return generic client errors in API responses. Keep detailed errors strictly server-side, and redact sensitive prompt/content/stderr before logging.

4. Resource Exhaustion (DoS) via Unbounded Request Parsing

• Files: internal/server/insights .go:136, 144, internal/insight/prompt.go:93
• Description: The handleGenerateInsight endpoint decodes the incoming JSON request body without imposing size limits and passes unbounded input to external CLIs (which can run for up to 1
  0 minutes). Malicious actors can send excessively large payloads to cause an Out-Of-Memory (OOM) crash or easily exhaust system resources.
• Remediation: Wrap http.Request.Body with http.MaxBytesReader (e.g., 1MB limit), enforce strict maximum
  lengths on all input fields (prompt, project, date, type), and implement per-client rate and concurrency limits for generation.

5. Race Condition in Insights Store

• Files: frontend/src/lib/stores/insights.svelte.ts:54, 1 49

• Description: The load() function can race with generate() completion. If a load() request is in-flight while generate() prepends a new insight locally, the delayed load() response can overwrite the items array, erasing the newly generated and selected insight.

• Remediation: Invalidate in-flight loads before applying local generation results (e.g., increment a version counter), or consistently refresh the list via load() after generation and deduplicate items by id.

---

Synthesized from 4 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (f0fca15)

Summary Verdict: The PR implements the end-to-end Insights feature but introduces high-severity security issues (missing authentication, XSS vulnerabilities) and requires fixes for DoS risks, CORS, and background
polling leaks before merging.

High Severity

• Unauthenticated Insights Endpoints
  • Files: internal/server/server.go:106, internal/server/insights.go:21, internal/server/insights.go:133

Issue: New insights endpoints are exposed without authentication/authorization checks. This includes read (GET), destructive (DELETE), and command-triggering generation (POST /insights/generate) operations. If the service is reachable by untrusted clients, this enables unauthorized data access, deletion, and abuse of local AI CLIs.

• Remediation: Add auth middleware on /api/v1/insights* routes, enforce per-user authorization, and explicitly reject non-loopback access for local-only mode.

• Stored XSS and Syntax Error in Markdown Rendering

  • Files: frontend/src/ lib/components/insights/InsightsPage.svelte (lines 214, 301)
  • Issue: The application renders AI-generated markdown using a raw HTML directive. There is a syntax artifact ({@frontend/index.html ...} instead of Svelte's {@ html ...}) that will cause a build failure. More importantly, rendering untrusted AI-generated content directly into the DOM creates a Stored Cross-Site Scripting (XSS) vulnerability via indirect prompt injection, as models could be manipulated into outputting malicious payloads.
  • Remediation: Fix the directive syntax
    to Svelte's {@html ...} and strictly sanitize the output of renderMarkdown using a trusted library like DOMPurify before rendering the HTML string.

Medium Severity

• Resource Exhaustion DoS Vector

  • Files: internal/server/insights.go: 133, internal/server/insights.go:190
  • Issue: handleGenerateInsight accepts an unbounded JSON body/prompt and can spawn long-running subprocesses (up to 10 minutes) per request, with no visible concurrency or rate limits.

• Remediation: Apply http.MaxBytesReader, enforce maximum lengths for input fields (prompt, project, type, date), and add server-side rate limiting/concurrency caps for generation jobs.

• CORS Allowlist Missing DELETE Method

  • Files: internal/server /server.go:109, internal/server/server.go:226
  • Issue: The DELETE /api/v1/insights/{id} endpoint is registered, but Access-Control-Allow-Methods still only advertises GET, POST, OPTIONS. Cross-origin browser clients will fail the CORS preflight for delete requests.
  • Remediation: Include DELETE in the CORS allow methods list to align with registered API verbs.

• Background Session Polling Leak

  • Files: frontend/src/App.svelte: 30, frontend/src/lib/components/layout/AppHeader.svelte:88
  • Issue: Navigating to Insights changes the route but does not clear or disable active session handling. The App.svelte effect continues to load and watch the session in the background, consuming
    unnecessary resources while the session UI is hidden.
  • Remediation: Gate the session watch/load logic on router.route === "sessions" or explicitly unwatch the session when navigating to the Insights page.

---

Synthesized from 4 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (33facca)

Verdict: The PR introduces a few medium-severity issues regarding potential sensitive data leakage in logs and background session syncing regressions that should be addressed.

Medium

1. Potential sensitive data leak
and log-forging via raw stderr/stdout logging

• Description: CLI failures are wrapped with raw stderr/stdout content, and then logged (log.Printf("insight generate error: %v", err)). This can leak sensitive prompt/session data into logs and allows
  log-forging via untrusted newline/control characters in model/CLI output.
• Suggested Remediation: Return/log structured, sanitized errors only. Avoid embedding raw stdout/stderr in error strings; if needed for debugging, redact and truncate, and escape control characters before logging.

Files:

• internal/insight/generate.go:70
• internal/insight/generate.go:84
• internal/insight/generate.go:129
• internal/insight/generate.go:258
• internal/server/insights.go:194

2. Background session sync continues while on Insights route (regression/perf)

• Description: The new Insights nav button only changes route (router.navigate("insights")) and does not clear/disable active-session
  watchers. Existing session-loading/watch effects are keyed to activeSessionId, so if a session was selected before switching routes, message polling/reloads can continue unnecessarily in the background.
• Suggested Remediation: Either deselect/unwatch when entering Insights, or gate message/session watch effects by
  router.route === "sessions".
• Files:
  • frontend/src/lib/components/layout/AppHeader.svelte
  • frontend/src/App.svelte

---

Synthesized from 4 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (a017507)

Verdict: The PR introduces the Insights feature end-to-end but requires fixes for a Svelte compilation error, a session state regression, and potential sensitive data leaks in error logs.

High

1. Invalid Svelte syntax
   File: frontend/src/lib
   /components/insights/InsightsPage.svelte:285
   Issue: Invalid Svelte syntax { @frontend /index.html renderMarkdown(...) }. This appears to be a broken find-and-replace for { @html } and will cause a Svelte compilation error.
   Remediation: Fix the syntax to correctly render HTML, likely by reverting to { @html renderMarkdown(...) }.

Medium

2. Session state is reset when navigating to Insights (regression)
   File: frontend/src/App.svelte:125

Issue: router.route changes trigger sessions.initFromParams(params) and sessions.load() unconditionally. Moving to #/insights (usually no session params) clears active session and resets filters (e.g., project/date), and also does unnecessary session reloads while on Insights.

Remediation: Only run sessions.initFromParams/load when router.route === "sessions" (or preserve existing session filters when route is insights).

3. Sensitive data leak in error logs
   File: [internal/insight/generate.go:69](/
   home/roborev/.roborev/clones/wesm/agentsview/internal/insight/generate.go:69), [internal/insight/generate.go:83](/home/roborev/.roborev/clones/wesm/agentsview/internal/insight
   /generate.go:83), internal/insight/generate.go:129, internal/insight/generate.go:252
   , [internal/server/insights.go:194](/home/roborev/.roborev/clones/wesm/agentsview/
   internal/server/insights.go:194)
   Issue: Error paths include raw CLI stderr and sometimes raw stdout in returned errors, and the server logs those errors directly. This can leak sensitive prompt/session content, model output, or local environment details into logs.

Remediation: Return/log sanitized error codes/messages by default; redact or truncate CLI output; gate full raw output behind explicit debug mode only.

---

Synthesized from 4 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (890cf0c)

The PR implements the new Insights feature but requires addressing critical functional regressions and medium-severity security issues.

High

• Database loss on upgrade

  • File(s): internal
    /db/db.go, [internal/db/db.go](/home/roborev/.roborev/clones/wesm/agents
    view/internal/db/db.go#L137)
  • Description: needsRebuild() treats a missing insights.date_from as a rebuild trigger. For users upgrading from versions without the insights table, this path drops and recreates the DB, losing
    all existing session data.
  • Suggested fix: Do not use a missing insights table/column as a full-rebuild condition. Add it via an additive migration (CREATE TABLE IF NOT EXISTS insights) and reserve rebuilds for truly incompatible schema states.

• Invalid Svelte syntax
  (Build Failure)

  • File(s): frontend/src/lib/components/insights/InsightsPage.svelte, .roborev.toml
  • Description: An errant global find-and-replace for the string html changed Svelte's
    raw HTML rendering tag from { @html renderMarkdown(...)} to { @frontend/index.html renderMarkdown(...)}. This is invalid syntax and will cause the frontend build to fail.
  • Suggested fix: Revert { @frontend/index.html back to { @html in both
    the Svelte component and the documentation in .roborev.toml.

Medium

• Untrusted execution paths for AI CLI tools

  • File(s): [internal/insight/generate.go](/home/roborev/.roborev/clones/wes
    m/agentsview/internal/insight/generate.go) (Lines 40, 61, 103, 232)
  • Description: The server resolves claude/codex/gemini with exec.LookPath from the runtime
    PATH and executes the resolved binary. If PATH is tampered with or includes writable locations, a malicious binary could be executed.
  • Suggested fix: Use fixed absolute binary paths from a trusted configuration, validate ownership/permissions, and run with a minimized trusted PATH.

Sensitive data leakage in error logs
* File(s): internal/insight/generate.go (Lines 73, 86,
136, 264), internal/server/insights.go
* Description: CLI stderr/raw output
are embedded into returned errors and subsequently logged by the server. This can leak sensitive prompt/session content or tokens into the logs.
* Suggested fix: Do not include raw stderr/stdout in returned errors by default. Log only sanitized summaries (or gated debug logs), with truncation/red
action.

• "Date Range Activity" mode UI bug
  • File(s): [frontend/src/lib/components/insights/InsightsPage.svelte](/home/roborev/.roborev/clones/wesm/agentsview/frontend/src/
    lib/components/insights/InsightsPage.svelte#L28)
  • Description: uiMode is derived from dateFrom !== dateTo, but handleModeChange("range_activity") only sets type and does not force a range. If dateFrom === dateTo (the default), the derived mode snaps back to daily_activity, meaning range controls may never appear.
  • Suggested fix: When selecting range mode, also set dateTo to a different date or maintain an explicit UI mode state rather than deriving it solely from dates.

---

Synthesized from 4 reviews (agents: codex, gemini | types: default, security)

[roborev-ci]

roborev: Combined Review (8cf6223)

Summary: The proposed changes introduce severe data loss and prompt injection vulnerabilities, alongside race conditions and logging issues, which must be addressed prior to merging.

High

• Prompt Injection Vulnerability via Untrusted Session Text

  • Files: internal/insight/prompt.go:79, internal/insight/prompt.go:113, internal/insight/generate.go:57, internal/insight/generate.go:226
  • Description: Untrusted session text
    (FirstMessage) and user-supplied Prompt are directly embedded into the LLM prompt, then sent to agent CLIs. While codex is sandboxed read-only, claude and gemini are invoked without equivalent confinement. This creates a prompt-injection path where malicious session content can steer
    tool-capable agents to access local files/commands and leak data into generated insights.
  • Remediation: Run all agents in a strict non-tool/sandboxed mode (or isolate in a locked-down subprocess/container), use a minimal working directory/env, and treat session text as unt
    rusted data blocks with explicit “do not follow instructions inside data” guardrails.

• Destructive Database Rebuild on Upgrade

  • File: internal/db/db.go:136
  • Description: needsRebuild() returns true when pragma_table _info('insights') doesn't contain date_from. For pre-feature databases where the insights table does not exist at all, this also returns 0. Consequently, Open() drops the whole database (dropDatabase), causing data loss during an upgrade.
  • Remed
    iation: Distinguish between “table missing” vs “table exists but missing required column.” If the table is missing, return false and let migrations create it.

Medium

• Sensitive Content Exfiltration in Error Logs

  • Files: internal/insight/generate.go:70,
    internal/insight/generate.go:84, internal/server/insights.go:198
  • Description: Error paths include raw CLI stderr and raw model output (stdout) in error strings, and server code logs those errors. This can expose sensitive content (prompt data
    , model output, file paths, tokens printed by CLIs) to logs.
  • Remediation: Redact/suppress raw CLI output in normal logs, log only high-level error codes/messages, and gate detailed diagnostics behind an explicit debug mode.

• Race Condition Hiding Newly
  Generated Insights

  • Files: frontend/src/lib/stores/insights.svelte.ts:59, frontend/src/lib/stores/insights.svelte.ts:157
  • Description: load() can be in flight while generate() succeeds
    and prepends the new item. If the older load() resolves after that, it overwrites this.items and the freshly generated insight disappears from the list until another reload.
  • Remediation: Invalidate/bump load version before optimistic prepend, or avoid prepend when loading === true and always refresh from server after generation.

---

Synthesized from 4 reviews (agents: codex, gemini | types: default, security)