AI Tool Conversation History Summary

Analyze AI tool conversation histories for a given period and generate themed work summaries.

Scheduling

Goal

Collect AI tool conversation history for a date or window and synthesize it into a themed, project-oriented recap with saved Markdown output.

Intent signature

• User asks for daily recap, weekly/monthly summary, standup notes, work log, tool usage pattern, or AI conversation history analysis.
• User wants conversation histories grouped by work content rather than raw chronological logs.

When to use

• Summarizing a day or period of work activity
• Understanding the overall flow of work across multiple AI tools
• Analyzing tool-switching patterns between sessions
• Preparing daily standups, weekly retros, or work logs

When NOT to use

• Git commit-based code change retrospective -> use oma retro
• Real-time agent monitoring -> use oma dashboard terminal
• Productivity metrics -> use oma stats get

Expected inputs

• Date, relative date, time window, or tool filter
• Conversation history available through oma recap --json or fallback sources
• Desired daily or multi-day recap scope

Expected outputs

• Markdown recap saved to .agents/results/recap/{date}.md or range filename
• TL;DR, overview, themes/projects, miscellaneous or side projects, and tool usage patterns
• User-facing summary in configured response language

Dependencies

• oma recap --json
• Optional Claude fallback history at ~/.claude/history.jsonl
• .agents/oma-config.yaml for language behavior

Control-flow features

• Branches by date resolution, window length, available tool history, and daily vs multi-day output shape
• Reads local history data and writes Markdown recap files
• Groups by content, not by tool

Structural Flow

Entry

1. Resolve requested date or window.
2. Collect normalized conversation history.
3. Decide daily versus multi-day output structure.

Scenes

1. PREPARE: Resolve time range and tool filters.
2. ACQUIRE: Collect history through CLI or fallback; retain completion evidence where available.
3. REASON: Group by content and classify each item as requested, in progress, or completed from its evidence.
4. ACT: Write recap Markdown in the required format.
5. VERIFY: Check that every completion claim has direct evidence, then check grouping, language, and output path.
6. FINALIZE: Save and display summary.

Transitions

• If no date is specified, use today via --date (bare --window is a rolling window ending now, not calendar-aligned).
• If window is 3 days or longer, group by project instead of day chronology.
• If CLI is unavailable, use Claude fallback only and report scope limits.
• If tasks are under threshold, group them into Miscellaneous or Side Projects.

Failure and recovery

• If history is unavailable, report missing source and requested range.
• If timestamps are ambiguous, use configured timezone and state assumption.
• If extracted data is sparse, produce a concise recap and note limited coverage.

Exit

• Success: recap file exists and summary is displayed.
• Partial success: missing tools/history or fallback-only coverage is explicit.

Logical Operations

Actions

Action	SSL primitive	Evidence
Resolve date/window	INFER	Natural-language date rules
Collect history	CALL_TOOL	oma recap --json or jq fallback
Read extracted records	READ	Conversation history
Group and classify themes/projects	INFER	Time/content rules plus prompt, progress, completion, receipt, or artifact evidence
Validate output shape	VALIDATE	Daily or multi-day template
Write recap	WRITE	.agents/results/recap/
Report summary	NOTIFY	Displayed recap

Tools and instruments

• oma recap --json
• jq fallback for Claude history
• Markdown output templates

Canonical command path

oma recap --date YYYY-MM-DD --json
oma recap --window 7d --json
oma recap --json  # rolling last 24h, not "today"

Resource scope

Scope	Resource target
LOCAL_FS	Conversation history and recap output files
PROCESS	oma recap, jq, date commands
USER_DATA	Conversation prompts and project activity
MEMORY	Theme grouping and summary notes

Preconditions

• Requested time range can be resolved.
• At least one history source is available.

Effects and side effects

• Writes recap Markdown under .agents/results/recap/.
• Reads local conversation history data.

Guardrails

1. Evidence status: A prompt alone proves a request, not a result. Mark work completed only with an explicit completion/result message, a receipt, or an artifact that supports the stated outcome. Mark it in progress with progress evidence; otherwise call it requested. Do not infer completion from a tool invocation or elapsed time.
2. TL;DR required: Top 3 supported outcomes. Use "completed" only when the evidence-status rule permits it; otherwise summarize requested or in-progress work plainly. Project name + status/outcome. No tool names or unnecessary detail.
3. Overview: After TL;DR, describe the flow. Start with "I" as subject and preserve evidence status.
4. Daily: themes by time block (15+ min). Rest goes to "Miscellaneous".
5. Multi-day (3d+): sections by project, ordered by activity. Read like a sprint report, not a daily log.
6. 2-4 bullets per theme/project: Concise essentials only. Don't enumerate every step.
7. Themes by content: Group by actual work, not by tool.
8. Time range (daily only): (AM/PM/Evening HH:MM~HH:MM). AM: 12:00, PM: 12:0018:00, Evening: 18:00~.
9. Save results: Write markdown to .agents/results/recap/.
10. Response language: Follows language setting in .agents/oma-config.yaml if configured.
11. No em dashes: Use commas, periods, or parentheses instead of — (em dash).

Process

1. Resolve Date

Determine the target date or window from the user's natural language input. Default is today.

Resolution rules:

• Relative day references (today, yesterday, day before yesterday, etc.) → calculate --date YYYY-MM-DD
• Specific date mentions (month + day, or full date) → convert to --date YYYY-MM-DD
• Relative weekday references (last Monday, this Friday, etc.) → calculate the date
• Period references (this week, last 3 days, past 2 weeks, etc.) → convert to --window Nd
• No date specified → today, resolved to --date YYYY-MM-DD (bare --window 1d is a rolling 24-hour window ending now, not the calendar day)
• The CLI caps windows at 30 days (longer values are trimmed with a warning) — when a requested period gets capped, say so in the recap

2. Collect Data

Extract normalized conversation history via CLI.

# Today (calendar day, all tools)
oma recap --date $(date +%F) --json

# Last 24 hours (rolling window ending now)
oma recap --json

# Time window (rolling, ends now; capped at 30d)
oma recap --window 7d --json

# Specific date
oma recap --date 2026-04-10 --json

# Tool filter (supported: grok, claude, codex, gemini, qwen, cursor, antigravity)
oma recap --tool claude,codex --json

Fallback when CLI is not installed: process Claude history only via inline jq:

# Uses the system timezone; export TZ=<zone> first to override, and state the
# timezone assumption in the recap (per the Failure and recovery rules).
TARGET_DATE=$(date +%Y-%m-%d)
# macOS/BSD date: advance the calendar date, then parse both local midnights.
next_date=$(date -j -v+1d -f "%Y-%m-%d %H:%M:%S" "${TARGET_DATE} 12:00:00" +%Y-%m-%d)
start_ts=$(date -j -f "%Y-%m-%d %H:%M:%S" "${TARGET_DATE} 00:00:00" +%s)000
end_ts=$(date -j -f "%Y-%m-%d %H:%M:%S" "${next_date} 00:00:00" +%s)000
# Linux/GNU date alternatives (replace all three lines above):
# next_date=$(date -d "${TARGET_DATE} 12:00:00 tomorrow" +%Y-%m-%d)
# start_ts=$(date -d "${TARGET_DATE} 00:00:00" +%s)000
# end_ts=$(date -d "${next_date} 00:00:00" +%s)000
# Do not add a fixed 24 hours: DST calendar days can have 23 or 25 hours.

jq -r --argjson start "$start_ts" --argjson end "$end_ts" '
  select(.timestamp >= $start and .timestamp < $end and .display != null and .display != "") |
  {
    time: (.timestamp / 1000 | localtime | strftime("%H:%M")),
    project: (.project | split("/") | .[-1]),
    prompt: (.display | gsub("\n"; " ") | if length > 150 then .[0:150] + "..." else . end)
  }
' ~/.claude/history.jsonl

3. Theme Analysis and Grouping

Read all extracted data and analyze with the following criteria:

Grouping rules:

• Only classify as a separate theme if the work spans 15+ minutes (based on timestamp gaps and prompt count)
• Merge consecutive prompts on the same topic into one theme
• Collect sub-15-minute tasks into a "Miscellaneous" section
• Group by work content, not by tool

Cross-tool analysis:

• Track workflow when multiple tools are used in the same time window
• Example: "Designed in Antigravity -> Implemented in Claude -> Reviewed in Codex"
• Derive insights from tool-switching patterns

Extract from each theme:

• Core work performed
• Key decisions made
• Tool combinations used
• Artifacts produced (docs, code, config, etc.)

4. Output Format

Save results to .agents/results/recap/{date}.md and display simultaneously.

Use the markdown templates in resources/output-formats.md:

• Daily format (1d or specific date): TL;DR → Overview → time-blocked themes → Miscellaneous → Tool Usage Patterns.
• Multi-day format (3d+): project-driven sprint-report structure with Side Projects for small (<30 prompts) work; follow the multi-day grouping rules in the same file.

Response language follows language setting in .agents/oma-config.yaml.

5. Save Results

Save to .agents/results/recap/{date}.md. For window ranges, use {start-date}~{end-date}.md format.

# Example paths
.agents/results/recap/2026-04-12.md
.agents/results/recap/2026-04-06~2026-04-12.md

References

• Output format templates: resources/output-formats.md
• Recap CLI: oma recap --json
• Output directory: .agents/results/recap/
• Language config: .agents/oma-config.yaml
• Claude fallback history: ~/.claude/history.jsonl