| name | quicker-agent-session-review |
| description | Analyze QuickerAgent chat exports (quicker-agent-chat-export JSON) to score authoring flow, detect tool/schema issues, and recommend prompt/tool/skill optimizations. Use when the user provides an exported thread JSON or asks to improve agent-gui behavior from real sessions. |
QuickerAgent Session Review
Offline analysis of exported chat threads to improve prompts, skills, tools, and agent flow.
When to use
- User attaches or paths to
quicker-agent-*.json under QuickerAgent exports
- User asks to optimize agent-gui after a failed or inefficient authoring session
- After manual benchmark runs — explain why the trace was slow or wrong
Input format
Exports use format: "quicker-agent-chat-export" (see agent-gui/lib/chat-thread-export.ts).
Required fields: thread, stats, messages[] with assistant tool-* parts and optional metadata.contextReport / agentTurnState.
Workflow
- Validate export format/version.
- Extract user prompt, tool timeline (with errors), token metrics, runtime metadata.
- Run rules
- Session rules:
agent-gui/lib/agent-session-analysis/session-rules.ts
- E-axis trace rubric:
agent-gui/lib/agent-eval/trace-rubric.ts
- Match user prompt to
authoring-tasks.json when possible.
- Report markdown + optimization hints mapped to repo paths.
- Optional F-axis: if action id known and task has
verify.mockProfile, run qkrpc action run --mock --assert.
CLI (deterministic)
From repo root:
pwsh -NoProfile -File ./scripts/Invoke-AgentSessionAnalysis.ps1 -Latest
pwsh -NoProfile -File ./scripts/Invoke-AgentSessionAnalysis.ps1 -ExportPath "C:\...\quicker-agent-*.json"
From agent-gui/:
pnpm agent-session -- --latest
pnpm agent-session -- "C:\Users\...\exports\quicker-agent-*.json"
pnpm agent-session -- export.json --json
Cursor slash command: /session-analysis
Exit code 1 when trace rubric fails or severity=error findings exist.
Report sections
| Section | Content |
|---|
| Summary | tokens, tool count, intent, matched benchmark task |
| Tool timeline | ordered calls, ERROR markers |
| Findings | session rules + E-axis violations |
| Optimization hints | ranked targets (instructions, skills, tools, Plugin) |
Scoring reference
- E-axis (auto):
docs/agent-authoring-benchmark.md — patch/get order, inline patch, step_runner_get
- A–D, F: human or
pnpm agent-eval -- <task> --verify-mock --judge
- Outcome oracle:
docs/quickerbench-design.md when task is IO-focused
Optimization map (common findings)
| ruleId | Likely fix |
|---|
schema-validation-error | Tool description + object examples in *-tool.server.ts |
token-baseline-high | model-tool-definitions.ts, tool-intent-filter.ts |
docs-call-heavy | Preload pattern skill in agent-skills/prompt.ts |
C-duplicate-search | Richer search results in StepRunnerCatalogMapper.cs |
create-then-read-data | docs/skills/quicker-authoring/prompt-tier0.md |
trace-rubric E violations | Flow fixes per benchmark doc |
Related
- Export UI: chat thread export →
%AppData%/QuickerAgent/exports/
- Full trace rules: reference/trace-rules.md
- Report template: reference/report-template.md
- Prompt assembly:
docs/agent-gui-prompt-structure.md
- Harness:
docs/superpowers/specs/2026-06-19-agent-gui-cursor-harness-design.md
Do not
- Treat UI-only tasks (MsgBox, text window) as QuickerBench pass/fail
- Recommend changes without citing export evidence (tool name + errorText)
- Guess step-runner
inputParams — analysis only; fixes go through search→get docs