0.2.0

0.2.0 - 2026-08-04

Added

• Tool & skill usage capture (src/capture/tool-usage.ts) — every run now records
  which MCP tools the agent called and which skills it loaded, not just what it was
  given. A green run that never reached for the configured MCP server is a different
  result from one that called it a dozen times, and that distinction is the metric the
  testbed exists to produce.
  • Per-tool aggregates (calls / errors / total duration), split into mcp / skill /
    builtin, plus an ordered timeline of the first 500 invocations.
  • Never-used lists: MCP servers that were configured but never called, and installed
    skills (.agents/skills/ + .claude/skills/) that were never invoked.
  • Read from opencode's SQLite store (<data dir>/opencode/opencode.db) read-only, so
    a live interactive session's store is never write-locked. Falls back to parsing the
    permission lines in <data dir>/opencode/log/*.log (marked source: "log", no timings
    or error status) when the db can't be read, and records null rather than a misleading
    all-zero result when neither is readable.
  • Headless/matrix entries collect once right after the agent exits; interactive sessions
    (which have no observable end) refresh on the stats collector's 30-second tick.
  • Surfaced in the History grid (MCP·Skill column — amber when some configured tooling
    went unused, red when none of it was touched), the History detail panel, report.html
    and summary.json (tools.mcp, tools.skills, tools.serversUnused,
    tools.skillsUnused — assertable from CI), both history exports, and a one-line run-log
    summary.
• Prompt images — attach reference mockups, sketches, Figma exports or screenshots to
  the agent's prompt, so "build this screen from the picture" becomes testable.
  • New read-write bind mount ./prompt-images/ → /prompt-images (PROMPT_IMAGES_DIR),
    created by run.sh / run.ps1. Read-write on purpose: browser uploads land in the same
    host folder a terminal-driven matrix config reads from — one folder, one namespace.
  • A Prompt images picker in both the Interactive and Matrix setup forms: thumbnails of
    what's in the folder, click to attach, upload straight from the browser, delete, rescan.
  • New pipeline stage attach-images (4c) copies the selection into the generated
    project's prompt-images/ (deliberately not dot-prefixed, so opencode's file browser and
    @-mentions can see it). Headless runs pass them to the agent as real attachments
    (opencode run "<prompt>" --file <img> …); interactive runs stage the copies and log the
    @prompt-images/<file> mentions.
  • REST surface: GET /api/prompt-images (list), GET /api/prompt-images/file?name=
    (serve one), POST /api/prompt-images?name= (raw bytes — no multipart dependency),
    DELETE /api/prompt-images?name=.
  • Matrix config gains an images field (alias promptImages) applied to every entry; an
    entry may name a single file or a whole folder, expanded to the images inside.
  • The attached set is recorded per run in History (with thumbnails in the detail panel) and
    in a matrix's report.html / summary.json.
  • Tunables: PROMPT_IMAGE_MAX_BYTES (per-file upload cap, default 10 MB) and
    PROMPT_IMAGE_MAX_COUNT (images per run, default 8; 0 disables the feature and the UI
    picker with it).
  • Vision is a paid-model feature: opencode's free/keyless hosted models silently drop
    attachments, so an image run on one degrades to a text-only prompt that still reports
    success. The pipeline warns when images are attached with no API key, and both setup
    forms and the README say so.
  • New reference doc prompt-images/README.md.
• Token breakdown — the input / output / reasoning / cache split now travels alongside
  the total, in the matrix report.html, summary.json (per-entry objects and the matrix
  totals, as tokensBreakdown), the History detail panel's per-model usage, and both
  history exports.

Changed

• Form controls across both setup views switched to the outlined Ignite UI variant
  (inputs, textareas, combos), the matrix prompt field moved from a hand-styled <textarea>
  to igc-textarea, and placeholder text got a legible color.
• run.sh / run.ps1 create and mount ./prompt-images, and their --help output now
  lists all bind-mounted host folders.
• The stats SSE stream replays the last tool-usage frame on connect, so a reconnecting client
  isn't blank until the next 30-second tick.
• History grid column widths reduced to eliminate the horizontal scrollbar.
• vendor/entry.js explicitly registers IgcFileInputComponent: as of
  igniteui-webcomponents 7.2.4 it is exported from the package index but missing from
  defineAllComponents(), so without this the upload control never upgrades.

Fixed

• Turning skills off actually turns skills off. ig new runs its own ai-config pass
  internally, and with --agents / --assistants absent it fell back to the CLI's
  interactive checkbox defaults (generic + claude) rather than to nothing — so a
  "no skills" run got the full .agents/skills/ + .claude/skills/ + AGENTS.md set
  installed at scaffold time, which the later ig ai-config --agents none could not undo.
  Every such baseline was silently contaminated, invalidating any matrix comparison against
  it. The scaffold now passes --agents={{agents}} --assistants=generic, and the prune stage
  additionally sweeps .agents/skills/, AGENTS.md and .claude/ as a guarantee
  (stripGeneratedAgentConfig).
• The prompt-image picker no longer shows stale entries after an upload or delete.
• Restored a missing import in src/pipeline/pipeline.ts.
• Removed a leftover AGENTS.md shipped by the locally installed skills.

Security

• Bumped body-parser 1.20.5 → 1.20.6 (Dependabot).

Full Changelog: 0.1.0...0.2.0