Agent Quest - v0.3.0

Agent Quest 0.3.0 — Context budget, skills & plugins, subagent families

Agent Quest now tells you how much of its context window each agent has used, which skills and plugins it is running under and what they cost, and makes every subagent read as part of the family of the hero that spawned it. Everything is read from the same local logs, for both Claude Code and Codex, and can be exported with one click so another AI can review a session.

Skills & Plugins — fleet view and AI-review export

Main agent — context, skills, delegation

Subagent — role badge, parent, context kept out

Highlights

• Context-window gauge — every hero carries a bar above its name (green → amber → orange → red, pulsing from 90%). The top bar shows a Ctx stat for the hottest main conversation and opens a dropdown comparing every live agent. The Detail Panel gets a Context section with tokens in use / window, peak, compactions and how the window was determined. An optional "context nearly full" alert fires at 85% and re-arms below 70%.
• Skills & plugins telemetry — a ✦ skill tag above the hero while a skill's instructions apply; the Detail Panel lists every skill loaded (how many times, ~tokens injected, attributed messages) and every plugin / MCP server called (calls, ~tokens of results), plus a line showing what all installed skills would cost if they were always loaded. The activity feed gains Skill and Compact rows and a Skills filter.
• Fleet view — the new ✦ button in the top bar opens Skills & Plugins: totals, installed inventory per provider, fleet-wide skill and plugin tables, and every agent's context at a glance. The panel is live, inventory included: installing a plugin or adding a skill shows up without a page refresh.
• Export for AI review — Copy MD, .md and .json buttons on each hero's panel and on the fleet view. The Markdown report covers session facts, context and token figures, activity breakdown, skills, plugins, delegation, the last message and a list of review questions; the JSON keeps counts instead of raw tool inputs.
• Subagent families — knights are always main sessions; subagents are pawns (archers for ultra-mode workflow squads), drawn smaller in the parent's color with a dimmer italic name and a SUBAGENT · role badge, while the parent wears a ⚑ N badge. Subagents nest under their parent in the Party Bar, and the parent's Delegation section shows the context its subagents processed, what they returned, and how much stayed out of the main conversation.
• Codex: new rollout format and subagents — Codex ≥ 0.160 writes item_completed events instead of the legacy per-kind events, which left Codex heroes standing still; both formats are now parsed. Codex sessions also get their context window (reported by Codex itself), a model badge, skill detection from SKILL.md reads, MCP plugin calls, and subagents spawned with spawn_agent are linked to their parent.

Notes

• Token figures marked ~ are estimates (characters ÷ 4). Context usage itself comes from the usage Claude Code records on every assistant message and from Codex's token_count events.
• Claude Code never writes the context window size to its logs. Agent Quest reads the model configured in settings.json (a [1m] suffix means the 1M window), infers 1M once a session exceeds 200k tokens, and otherwise assumes 200k. Override it with AGENT_QUEST_CLAUDE_CONTEXT_WINDOW=<tokens>.
• Both CLIs write two markers per compaction; Agent Quest counts each compaction once.
• Hero classes are now assigned by role instead of in rotation, so sessions that were archers or pawns before the upgrade become knights.
• Context bars and the context alert can be switched off in Settings. Subagents still never fire notifications.