Reflect

Reflect

Overview

Reflect is the official recall-reflect Recall extension. It is a discussion-first workflow for reviewing AI coding sessions through Recall's stable CLI JSON and JSONL protocol. The extension reconstructs a conversation-first timeline across sessions and surfaces observed patterns as questions for the user, not as automatic conclusions or edits.

Use this skill when the user wants project-level reflection on AI coding process, workflow friction, handoffs, repeated corrections, scope drift, or possible future calibration.

Reflect output is evidence for discussion, not permission to change skills, prompts, configs, instruction files, or project files.

Core Rule

Do not treat an observed pattern as a calibration target, workflow change, or proposal unless the user explicitly confirms the interpretation. Ask first, then act only on confirmed direction.

Scope Defaults

Scope to the current repo unless the user asks otherwise. Resolve the project directory with:

git rev-parse --show-toplevel

Run Reflect with structured output. Prefer user-facing dispatch when the extension is installed and managed by Recall:

recall reflect --project /absolute/project/path --format json

Direct extension invocation is also acceptable when testing or calling the extension binary explicitly:

recall-reflect --project /absolute/project/path --format json

Add --time 30d or --source <source> when the user narrows the review. Use --sync when recent sessions may not be indexed yet.

Workflow

1. Resolve the repo scope.

   • Use the repository root for current-repo reflection.
   • Use the user-provided path when they specify one.
   • Explain the scope briefly if there may be multiple worktrees or project directories.

2. Run Reflect.

   • Prefer recall reflect ... when Recall can dispatch to the installed extension.
   • Use recall-reflect ... for direct testing or explicit extension invocation.
   • Prefer JSON for structured inspection.
   • Use text output only when the user wants a human-readable report directly.

3. Read the timeline as conversation history.

   • Focus on user intent, agent response, corrections, decisions, and outcomes.
   • Do not turn the timeline into a tool-call log.
   • Do not quote large transcript blocks.

4. Present observed patterns as discussion prompts.

   • Use the prompt text from observed_patterns[].discussion_prompt when present.
   • If there are no observed patterns, summarize the timeline and say there is not enough repeated evidence yet to discuss a pattern.
   • Do not add confidence scores, personality labels, or identity judgments.

5. Discuss before proposing changes.

   • Ask which pattern, if any, the user wants to inspect further.
   • Ask whether the pattern is real, intentional, or a false positive.
   • Only after user confirmation may you suggest a calibration target, workflow rule, or follow-up implementation plan.