ccstory 0.7.0

Claude Code usage recap with narrative. ccusage tells you the bill, ccstory tells the story.

ccstory

Your Claude Code week, in plain English. Reads ~/.claude/projects/**/*.jsonl locally and writes a categorized recap with active hours, costs, and a per-bucket narrative.

Sibling to ccusage: ccusage tells you how much you spent · ccstory tells you what on.

Who this is for

• People who want to write a weekly status without scrolling scrollback.
• People who saw a ccusage number and want to know what kind of work those tokens went to.
• People who do a Sunday-night reflection on what they actually shipped.

Quick start

pipx install ccstory
ccstory init
ccstory week

That's it. init is a one-time auto-categorize step that scans your recent sessions; ccstory week produces the recap. Full report saves to ~/.ccstory/reports/recap-*.md.

The default What shipped section may query GitHub and PyPI metadata. For a first run with no network access at all, use:

ccstory init --skip
ccstory week --minimal --classify folder --no-artifacts

--no-artifacts alone disables ccstory's GitHub/PyPI lookups while keeping the normal narrative flow, which may invoke your installed Claude Code CLI.

Demo

╭──────────────── Claude Code Recap · May 5 – 12, 2026 ────────────────╮
│                                                                      │
│  ★ Top focus  coding  10.9h  (53% of active time)                    │
│    ↳ Built /show-routine slash command using bash+python to fetch…   │
│                                                                      │
│  Active  20.6h  Sessions  74   Output  2.92M                         │
│  Turns   3,692  Cache     96%  Cost    $1,608                        │
│                                                                      │
│  Time by category                                                    │
│  coding          ███████████████░░░░░░░░░░░░░   10.9h    53%         │
│  writing         █████████░░░░░░░░░░░░░░░░░░░    6.2h    30%         │
│  research        █████░░░░░░░░░░░░░░░░░░░░░░░    3.5h    17%         │
│                                                                      │
│  Full report → ~/.ccstory/reports/recap-2026-05-10_2026-05-17.md     │
│                                                                      │
╰────────────────────────────── ccstory ───────────────────────────────╯

The markdown report adds a header + 2-4 bullet points per bucket plus per-session one-liners. Run with --llm-narrative to upgrade per-session lines from the instant first/last-message fallback to claude-polished prose:

Re-running upgrades retroactively. If you viewed a window in the default (instant) mode first, re-running it with --llm-narrative upgrades those cached fallbacks to polished summaries — so ccstory month --llm-narrative polishes weeks you already skimmed. Already-polished sessions are reused (no re-burn) unless their prompt version is stale; add --refresh to force every in-window summary to regenerate (e.g. after a claude model upgrade you want reflected).

### coding

Shipped the /show-routine slash command end-to-end this week — bash+python
wrapper to surface scheduled-task output, plus a routine-detail bookmark
flow after the live debug session on Wednesday.

- 2026-05-10 03:24 · 123m · 212 msg — Built /show-routine slash command using
  bash+python to fetch scheduled-task output and surface it inline.
- 2026-05-08 12:30 · 67m · 294 msg — Debugged hook race condition in
  background-task notification dispatch; landed fix in main.

Usage

Basic

Command	What it does
ccstory init	One-time auto-categorize from recent sessions
ccstory	Current month so far (default window)
ccstory week	Past 7 days
ccstory month	Current month
ccstory 2026-04	A specific month
ccstory trend	Last 8 weeks of sparklines
ccstory category list	Show your custom bucket rules
ccstory category set <bucket> <keyword>…	Pin a project to a bucket
ccstory category unset <bucket> <keyword>…	Remove a keyword from a bucket

Advanced

Window

Command	What it does
ccstory all	Entire history
ccstory trend --weeks 12	Custom trend range
ccstory trend --months 6	By calendar months

Narrative depth

Flag	What it does
--minimal	Numbers only, no per-session lines
--llm-narrative	claude -p per-session prose (slow, opt-in)
--no-aggregate	Skip the per-bucket synthesis

Comparison block (vs-previous, auto-attached to week/month)

Flag	What it does
--no-compare	Skip the entire block
--no-compare-narrative	Keep numeric deltas, drop the prose

Coding agent

Flag	What it does
--agent all (default)	Every agent ccstory can read
--agent claude	Claude Code only (~/.claude/projects)
--agent codex	OpenAI Codex only (~/.codex/sessions)

Also accepted by ccstory trend, so a trend line and a week over the same range describe the same population. See Multiple coding agents for what the numbers mean once more than one agent is in the window.

Session classification mode

Flag	What it does
--classify folder	Folder-name rules only
--classify content	claude -p reads each session
--classify hybrid	User rule wins, else content (default)

Export

Flag	What it does
--for=obsidian	YAML frontmatter + [[wikilinks]]

Output format

Flag	What it does
--format=card	Force the Rich terminal card (default in a real tty)
--format=markdown	Force the full Markdown report to stdout
--format=auto (default)	Markdown when CLAUDECODE=1 or stdout is not a tty (piped / redirected), else card

The auto-detect means asking Claude Code "show me my week with ccstory" renders an actual Markdown report in the chat instead of ANSI escape codes. The Markdown body is the same content saved to ~/.ccstory/reports/ (recap-*.md for the default window, trend-*.md for ccstory trend), just printed to stdout so the chat can render it inline. In markdown mode all progress / status lines route to stderr, so stdout is a clean Markdown stream you can pipe.

Refresh (apply rule changes retroactively)

Flag	What it does
--refresh	Re-do this window's cached work: re-classify after a rule edit, and (with --llm-narrative) force-regenerate every per-session summary
--refresh-all	Wipe the entire content-classification cache, not just this window

Trend output

Hours by bucket
total          ▁▄▆▇▃█    16.5h   avg 9.0h   ▲ +183%
coding         ▁▂▃▄▁█    10.2h   avg 3.3h   ▲ +1148%
writing        ▁▇█▆▁▁     6.2h   avg 4.1h   ▲ +51%
research       ▁▃▅█▆█     3.5h   avg 2.0h   ▲ +75%

Overall
output         ▁▁▁▄▁█     3.0M   avg 0.8M   ▲ +260%
cost           ▁▁▂▃▁█   $1,643   avg $463   ▲ +255%
burn %         ▁▁▂▃▁█     201%   avg 57%    ▲ +255%

The burn % row is API-equivalent cost as a percentage of your prorated monthly quota. Set monthly_quota_usd in ~/.ccstory/config.toml (default $3,500 ≈ Max 20x plan); set to 0 to hide the row.

Categories

ccstory classifies each session into two layers:

• Area (layer 1) — the coarse bucket (coding, investment, …). Trend and compare aggregate at this layer, and its numbers are the stable contract downstream tools (dashboards, the MCP get_recap / get_trend shapes) read.
• Project (layer 2) — the normalized project-folder leaf (e.g. ccstory, stock). Projects emerge automatically from your session folders — no extra config — and the recap card, markdown report, and --json break each area down by project. This breakdown is computed at read time, so it adds no cache and never re-classifies history.

Four default areas, matched against the project folder name:

Area	Keywords (sample)
investment	investment, stock, portfolio, trading, ticker, etf, finance
writing	blog, newsletter, post, docs, content, article
coding	app, sdk, cli, plugin, mcp, server, frontend, backend, lib, …
other	playground, scratch, sandbox, experiment

Unmatched projects fall back to coding. Customize in ~/.ccstory/config.toml:

default_bucket = "coding"

[categories]
# An entry equal to a project's normalized leaf is an *exact member* of that
# area. Substrings still work as a fuzzy fallback, so existing configs keep
# resolving exactly as before.
"learning"   = ["info-collector", "ai-project-research"]
"investment" = ["stock", "kol-collector"]

Two matching tiers. The resolver checks exact membership first (the project's normalized leaf listed verbatim under an area), then falls back to the older token-needle fuzzy match. Because an exact member always wins over an earlier area's fuzzy hit, you can delete the section-ordering workarounds fuzzy matching used to force (listing one area before another just so a shared substring resolved the way you wanted). Listing the same project under two areas prints a warning at load and keeps the first.

Aliases (optional). Fold variant folder-leaf names onto one canonical project with a [projects] table — useful when the same work shows up under more than one folder name:

[projects]
"infocollector" = "info-collector"   # both roll up as one project

Area overrides. Folder rules can be overridden per-session by content (--classify content / hybrid), where one batched claude -p call re-buckets sessions by what they were actually about. An override changes a session's area only — its project is the physical fact of which folder the work happened in, never reassigned. Results cache in ~/.ccstory/cache.db so reruns are free.

Multiple coding agents

ccstory reads Claude Code (~/.claude/projects) and OpenAI Codex (~/.codex/sessions, plus archived_sessions) by default. Codex sessions are attributed to a project from the cwd the transcript records, folded through the same rules Claude Code project folders get — including git worktrees, so a detached checkout counts toward the repo it came from rather than becoming its own one-off project.

Time is reported once, not per agent. Agents run concurrently: a Codex review and a Claude Code session routinely occupy the same ten minutes. Summing their active time double-counts that overlap — on a real week here, raw per-agent time added up to 177h against a deduplicated wall clock of 64h. So:

• Total active time is the wall clock across every session, deduplicated — the same number ccstory has always reported, now spanning all agents.
• The Coding agents block reports shares, not hours. Each agent's share is its raw interaction time relative to the others'. Shares are not durations and do not add up to the total.
• Session share is shown next to time share, and they disagree on purpose. Many short Codex reviews against fewer long Claude Code sessions shows up as 75% / 25% of time but 51% / 49% of sessions — that gap is the finding.
• N× parallel is raw agent time ÷ wall clock: how much of the work overlapped.

Token counts and costs still cover Claude Code only — Codex usage appears in the time breakdown but not in the cost numbers, and the report says so inline.

Use --agent claude to get the pre-multi-agent numbers back.

What shipped

Time tells half the story; the other half is what the time produced. Each report includes a What shipped section — per-repo output metrics for the repos you actually worked in during the window:

| Repo     | Commits | PRs merged | Releases | Stars   |
|----------|--------:|-----------:|----------|--------:|
| ccstory  | 5       | 3          | v0.6.1   | 42 (+6) |
| myapp    | 21      | 1          | –        | 12      |

- PyPI **ccstory**: 107 downloads (last week)

• Repos are inferred from session working directories — no config needed. Worktrees collapse into their main repository.
• Commits come from local git (works offline, counts all branches). PRs merged / releases / stars need the gh CLI; the lookup sends the GitHub repo slug and requests recent merged-PR/release timestamps plus the current star count. ccstory applies the report window locally. Without gh, those columns degrade to –. PyPI downloads send the package name to pypistats.org for packages auto-detected in active repos' pyproject.toml.
• The artifacts collector never sends conversation text, prompts, summaries, commit contents, or local paths. It uses only repository/package metadata.
• Stars delta compares against the last snapshot taken before the window, so it becomes meaningful from your second run onward.

Skip all GitHub/PyPI metadata calls per run with --no-artifacts, or persistently via config:

[artifacts]
enabled = false            # no GitHub/PyPI metadata lookups
exclude = ["playground"]   # substring match on repo path
pypi = ["my-package"]      # extra packages beyond auto-detection

Narrative depth

## What you did is 2-4 goal threads (bold header + bullets) by default. For real retrospectives, --narrative goes deeper:

ccstory week --narrative per-category   # header + bullets per bucket instead
ccstory week --narrative both           # overall first, then per-bucket

Each bucket costs one claude -p call, cached until its exact input or prompt changes — rerunning the same window is normally free. A bucket whose synthesis fails (or that has no real summaries) is simply omitted; the report never blocks on it. In --json mode the same text lands in buckets[].narrative.

Claude CLI calls, latency, and quota

There is no single fixed call total: it depends on init mode, uncached sessions, narrative depth, and which cache entries already exist. Every claude -p call runs through your installed Claude Code CLI and uses that CLI's signed-in plan/quota; ccstory does not use an API key or add a separate API charge.

Operation	Fresh claude -p calls	Cache behavior
ccstory init --quick	1 (usually ~10s)	One-time config proposal
ccstory init --deep	1 per 80 sampled sessions (up to 3 with the default cap of 200)	Writes per-session classification cache
ccstory init --skip	0	Uses local folder rules only
Hybrid/content classification	1 per 80 uncached or stale sessions	Reused until its prompt or category vocabulary changes
Overall narrative	0 or 1 on a cache miss	Reused while its rendered inputs and prompt are unchanged
Per-category narrative	Up to 1 per eligible bucket on a cache miss	Reused while that bucket's inputs and prompt are unchanged
Previous-window narrative	0 or 1 on a cache miss	Reused while its comparison inputs and prompt are unchanged
--llm-narrative	1 per uncached or stale session	Reused per session; --refresh deliberately regenerates

The default recap uses hybrid classification, an overall narrative, and a previous-window narrative; per-session LLM prose remains opt-in. Use --narrative per-category|both to trade the overall call for, or add, bucket calls. --no-aggregate, --no-compare-narrative, and --classify folder remove those call types; --minimal --classify folder makes the recap itself use zero Claude calls.

Deep/content classification is batched; per-session --llm-narrative work is linear and the CLI budgets roughly 40 seconds per cold session, showing an ETA before it starts. Aggregate call latency varies with Claude CLI startup and input size. A same-window rerun is usually cache-only, but new sessions, changed inputs/config, --refresh, or a newer prompt version can trigger fresh calls. Content classification carries accepted bucket names into later 80-session batches and enforces one run-wide vocabulary cap, preventing a large first run from fragmenting one theme into several near-duplicate labels.

If claude is absent from PATH, LLM classification and synthesis degrade gracefully: classification uses folder/fallback rules, per-session prose uses the local first/last-message fallback, and Claude quota usage is zero. This does not disable What-shipped metadata calls; add --no-artifacts for that.

JSON output

For dashboards, bots, and sync scripts — one machine-readable object instead of parsing markdown:

ccstory week --json          # shorthand for --format=json
ccstory month --format json
ccstory trend --weeks 8 --json

stdout is pure JSON (progress goes to stderr, same as markdown mode), so ccstory week --json | jq .totals.active_hours just works. The envelope carries schema_version (currently 1): renames/removals bump it, additive fields don't — consumers should tolerate unknown keys. Covers window, totals (hours/tokens/cost/cache), buckets, per-session lines, model breakdown, narrative, comparison, artifacts, and the pricing snapshot date. The markdown report file is still written either way; JSON is a view, not a replacement.

Obsidian export

ccstory --for=obsidian swaps the plain markdown for a PKM-vault-ready variant with YAML frontmatter and [[wikilinks]]:

---
date_start: 2026-05-10
date_end: 2026-05-17
active_hours: 20.6
top_focus: coding
buckets: [coding, writing, research]
cost_usd: 1608.42
output_tokens: 2920000
---

Queryable in Obsidian's Dataview / Bases (WHERE top_focus = "coding"). Bucket names with special characters are JSON-quoted so the frontmatter stays valid even for client: acme, inc.

Narrative language

ccstory delegates narrative writing to your local claude -p. By default it inherits whatever language Claude Code itself responds in; override it per run, per shell, or persist a per-tool choice.

Precedence (high → low):

Source	Notes
--lang "Traditional Chinese"	One-off, this invocation only
CCSTORY_LANG=日本語 env var	Shell-scoped
language = "Spanish" in ~/.ccstory/config.toml	Persistent, ccstory-only
~/.claude/CLAUDE.md	Pasted verbatim, so it can carry richer directives
~/.claude/settings.json language	Set by Claude Code's /config UI
System locale ($LANG)	Auto-detected — zh_TW → Traditional Chinese, etc.
English	Final fallback

ccstory week --lang "Traditional Chinese"   # one-off
export CCSTORY_LANG="日本語"                  # shell-scoped
# or in ~/.ccstory/config.toml:
# language = "Spanish"

The value is dropped straight into the prompt as Respond in <value>., so any name Claude can parse ("Traditional Chinese", "日本語", "pt-BR") works.

Custom pricing

Default API list prices snapshot to 2026-07. Every human-readable report shows the snapshot date and warns once it is over 90 days old relative to the report window end. ccstory makes no pricing network requests at runtime; model rates ship with each release and come from the LiteLLM registry.

Override per-model rates in ~/.ccstory/config.toml:

[prices]
snapshot_date = "2026-08"

[prices.opus]
input       = 6.0
output      = 30.0
cache_write = 7.5
cache_read  = 0.6

Partial overrides are fine — unspecified keys keep their default. Defining a brand-new model ([prices.custom]) with only some keys defaults the rest to $0 with a warning so misconfig is loud.

How ccstory differs from ccusage

	ccusage	ccstory
Role	The bill	The story
Active hours (5-min gap heuristic)	—	✅
Activity categories	—	✅ folder rules + content-aware
Per-session narrative	—	✅ via local claude -p
Per-bucket synthesis	—	✅
Cross-period narrative	—	✅
Conversation logs stay local / no telemetry	✅	✅

Pair them — ccusage monthly for the spend, ccstory month for the breakdown:

ccusage monthly
ccstory month

Privacy and network behavior

ccstory never sends your conversation data to its own service or to the What-shipped metadata providers. There is no ccstory telemetry or account.

• Data source: ~/.claude/projects/**/*.jsonl — Claude Code's own logs.
• Narratives and classification: subprocess-call your locally installed claude -p. The Claude CLI contacts Anthropic using your signed-in session and plan quota; ccstory does not use your API key or operate a proxy.
• Pricing: ccstory makes no pricing network requests. Model prices ship with each release and come from the LiteLLM registry.
• What shipped: local git supplies commit counts. By default, gh may send a repo slug and request recent PR/release timestamps plus the current star count from GitHub; ccstory filters timestamps to the report window locally. The pypistats request sends a package name to pypistats.org. No conversation text, prompt, summary, local path, or commit contents are included.
• Cache: ~/.ccstory/cache.db (sqlite, per-session summaries).
• Reports: ~/.ccstory/reports/recap-*.md.

Disable GitHub/PyPI metadata calls with --no-artifacts or persistent [artifacts] enabled = false. For a fully no-network report, also avoid Claude CLI calls with --minimal --classify folder (and initialize with ccstory init --skip). Relevant implementations are ccstory/artifacts.py and ccstory/session_summarizer.py.

Requirements

• Python 3.11+ and pipx (brew install pipx on macOS, other platforms).
• Claude Code CLI on PATH — required for --llm-narrative, content classification, and the cross-period synthesis. Without it, narratives fall back to first/last user-message excerpts and --classify falls back to folder rules.

Implementation notes

• Time math: 5-minute gap heuristic — consecutive messages within 5 minutes count as active, longer gaps are "stepped away". Wall-clock dedup prevents parallel sessions from double-counting. The 5-min cap is a practical floor for "still at the keyboard"; comparable across periods even though not precise.
• Timezone: session timestamps are parsed UTC-aware. Window boundaries (week, month) are local-midnight aligned, so "this week" matches the calendar week you actually lived in. --weeks N for trend mode does the same.
• Cost comparison: cross-period diffs use output tokens, not total_tokens. In typical use ~96% of total_tokens is cache_read, which inflates with turn count and system prompt size and isn't a stable signal of work done. Output tokens stay comparable month over month.
• Pricing: prices are list prices snapshotted by date (default 2026-07); every human-readable report shows the snapshot date and warns when it is over 90 days old relative to that report's window end.

Library usage (integration API)

ccstory is primarily a CLI, but a small set of functions is maintained as a semi-stable integration API for programmatic consumers — dashboards, scripts, and the MCP server below all call these instead of shelling out to the CLI:

from ccstory.recap import build_recap
from ccstory.time_tracking import collect_sessions, rollup_by_category
from ccstory.categorizer import classify, load_rules

result   = build_recap("week")                   # one call = full recap
sessions = collect_sessions(since, until)        # any window, tz-aware
rollups  = rollup_by_category(sessions)          # per-bucket hours/share
bucket   = classify(project_dir)                 # folder-rule bucketing
rules    = load_rules()                          # parsed ~/.ccstory/config.toml

build_recap() runs the same pipeline as the CLI (the CLI is a thin shell over it) and returns a RecapResult: rich objects (.sessions, .rollups, .usage, narratives, comparison) plus .markdown, .report_path, and .to_json() — the schema_version: 1 envelope, same shape as --json stdout. Keyword args mirror the CLI flags one-to-one (llm_narrative=, narrative=, classify=, …); pass a Rich Console via console= for progress output, or nothing for silence. An empty window raises RecapUnavailable instead of exiting the process.

Semi-stable means: signatures may still change with minor versions, but renames and behavior changes are called out in the changelog instead of happening silently. Everything else in the package is internal. The JSON envelope (--json, schema_version: 1) is the other supported contract.

MCP server

pip install 'ccstory[mcp]'
ccstory mcp   # stdio MCP server — read-only, no fresh `claude -p` by default

Point any MCP-aware client (Claude Desktop, Claude Code, or another local agent) at the ccstory mcp command and it can ask for your recap live in conversation instead of you running the CLI and pasting output back in. Example client config (Claude Desktop's claude_desktop_config.json, or Claude Code's MCP settings — same shape):

{
  "mcpServers": {
    "ccstory": {
      "command": "ccstory",
      "args": ["mcp"]
    }
  }
}

Four read-only tools:

Tool	Returns
get_recap(window, classify, allow_llm)	Totals, per-category active hours + narrative + a children per-project breakdown (name + hours), the overall narrative, top 5 sessions, cost.
compare_to_previous(window, classify)	Active-hours and cost deltas vs. the immediately preceding same-length window.
get_trend(period, count, classify)	Per-period series over the last count weeks/months (oldest first): active hours, cost, per-category hours. count clamped to 1..24.
list_categories()	The bucket rules ccstory classifies sessions into (user + built-in defaults).

window accepts week / month / all / YYYY-MM, same as the CLI; period is week or month. Default classify="folder" and allow_llm=False never trigger a fresh claude -p call — an MCP client may call these tools opportunistically mid-conversation, so nothing here should cost you latency or tokens unless you explicitly ask for it (classify="content" / "hybrid", or allow_llm=True on get_recap; compare_to_previous and get_trend stay cache-only under every parameter combination).

This is a third, distinct JSON contract, not the same shape as either of the two above: not --json / RecapResult.to_json() (which lists every session in the window) and not the Python function signatures. MCP responses are deliberately compact — top 5 sessions, not the full list — so they're cheap for an agent to read into its own context, and never include raw transcript text, only summaries.

Roadmap

• --json structured output — one general primitive over per-destination export flavors
• Optional PNG card export
• ccstory year — annual recap (Spotify-Wrapped style)
• Git commit / PR correlation — period-level What shipped section (per-session attribution still open, #11)

See the issue tracker for the full backlog.

Changelog

See CHANGELOG.md for release history.

License

MIT — see LICENSE.