cc-session-tools 0.9.0

Claude Code session management CLI tools (ccd, ccr, ccs, claude-code-usage, ccst) and hook library (cccs_hooks)

claude-code-session-tools

Three concerns, one repo, for life on the Claude Code CLI:

1. Session management - start, resume, find, and relocate Claude Code sessions from the shell, with tagged dated session directories that don't pollute your repo root.
2. Usage analytics - parse ~/.claude/projects/**/*.jsonl into tokens-and-dollars breakdowns by project, session, model, MCP server, plugin, and tool.
3. Hook library - Python package (cccs_hooks) providing Claude Code PreToolUse / PostToolUse / UserPromptSubmit / Stop hook implementations, invokable via ccst hooks run <name>.

The repo ships five CLIs and three bundled skills:

	What it does
ccd <tag>	Start a new session with a pre-created cc-sessions/<date>-<tag>/ directory and a tagged display name.
ccr <fragment>	Resume an existing session by typing any substring of its name.
ccs <query>	Search across your sessions by name, file contents, or transcript messages, in the current project or across every configured root.
claude-code-usage	Multi-dimensional usage analytics CLI: query/group/filter by project, session, model, MCP server, plugin, tool, day/week/month/year. Reconciles dollar totals against ccusage.
ccst <noun> <verb>	Umbrella CLI for hook and skill management. ccst hooks install merges hook entries; ccst hooks run <name> runs a hook by name; ccst skills install symlinks bundled skills into ~/.claude/skills/.
Skill: find-claude-code-session	Wraps ccs. Lets a Claude Code session locate one of your prior sessions by name or content and offer a ccr command to resume it.
Skill: move-session	Move, rename, or move+rename a session while keeping its ~/.claude/projects/<encoded-cwd>/<uuid>.jsonl transcript resumable.
Skill: claude-usage	Wraps claude-code-usage. Lets a Claude Code session answer "how much have I spent on Opus this month?" without you typing the CLI yourself.

If you've ever tried to remember which 1f4a8b3c-... UUID is the session where you were debugging that flaky test last Tuesday, or wondered which project burned through last week's Opus budget, this is for you.

Installation

Prerequisites

• Python 3.11+ (3.12+ recommended)
• The claude CLI on your $PATH. Install it first via the official Claude Code instructions and verify with claude --version.
• ccusage (optional) - if on $PATH, claude-code-usage reconcile cross-checks dollar totals against it. Skipped gracefully if missing.
• ripgrep (optional) - ccs --contents prefers rg; falls back to threaded Python grep if missing.

Install the tools

The simplest path is pipx, which installs the commands into an isolated venv and puts them on your $PATH:

pipx install cc-session-tools

If you use uv:

uv tool install cc-session-tools

Either way, ccd, ccr, ccs, claude-code-usage, and ccst will be available on your $PATH. Verify:

ccd --version
claude-code-usage --version
ccst --help

Installing from source (pre-release or offline):

git clone https://github.com/raffishquartan/claude-code-session-tools.git
cd claude-code-session-tools
uv tool install .

Upgrade

# pipx
pipx upgrade cc-session-tools

# uv
uv tool upgrade cc-session-tools

Installing from a local clone: if you keep a local clone for development or to stay on the latest commit, refresh the global install with:

uv tool install ~/repos/claude-code-session-tools

Do NOT run uv tool install from inside a git worktree. That overwrites the global install's source pointer with the worktree path, which breaks all five CLIs when the worktree is deleted. Use uv run pytest to test inside a worktree, and run uv tool install ~/repos/claude-code-session-tools from outside the worktree after merging.

Install the skills

Run ccst skills install to symlink all bundled skills into ~/.claude/skills/:

# Dry run (shows what would be created)
ccst skills install

# Write the symlinks
ccst skills install --apply

Manual symlinks still work if you prefer:

ln -s "$PWD/skills/find-claude-code-session" ~/.claude/skills/find-claude-code-session
ln -s "$PWD/skills/move-session"             ~/.claude/skills/move-session
ln -s "$PWD/skills/claude-usage"             ~/.claude/skills/claude-usage

The skills shell out to the installed CLIs - they don't import the Python library directly, so the only requirement is that ccs / claude-code-usage are on $PATH.

Why bother?

Claude Code stores each session as a ~/.claude/projects/<encoded-cwd>/<uuid>.jsonl transcript and exposes them through claude --resume. That works, but the picker shows untagged sessions in opaque order, the working files for a session sprawl into your repo root, and there's no built-in way to grep across past conversations or see where your tokens went.

These tools add:

1. A tagged, dated session directory under <project>/cc-sessions/<YYYYMMDD>-<tag>/ with working/ and out/ subdirs - the convention Claude Code's session memory hooks expect when you want scratch space and deliverables that don't pollute your repo.
2. Resume-by-fragment so you can type ccr flaky instead of scrolling through the picker.
3. Cross-session search so ccs --contents --global "GraphQL retry" finds every conversation that mentioned it.
4. Usage analytics so claude-code-usage query --since 2026-04-01 --group-by project,model answers where the spend went.
5. Skill wrappers so the Claude Code agent can do all of the above on your behalf when you ask in natural language.

Configuration: where do your sessions live?

ccd refuses to start a session if your current working directory isn't a direct child of one of your configured session roots. This sounds annoying but turns out to be a feature: it stops you from accidentally starting a session in /tmp or in ~, and it lets ccr/ccs find sessions across your projects without you telling them where to look.

Roots are configured via two environment variables - both optional, but you'll want at least one:

CLAUDE_SESSION_TOOLS_REPO_ROOT - the loose root

Point this at the directory whose direct children are your projects (the typical case is ~/repos):

export CLAUDE_SESSION_TOOLS_REPO_ROOT="$HOME/repos"

A "session root" means: if $REPO_ROOT/foo/ exists, then cd ~/repos/foo && ccd my-tag is allowed and creates ~/repos/foo/cc-sessions/20260509-my-tag/. Sessions started two levels deep (~/repos/foo/sub/) are rejected unless you pass --force.

Under the loose root the only naming rule is no spaces in the tag. Tag suffixes can be anything else: bugfix-7, redesign, try-out-thing.

CLAUDE_SESSION_TOOLS_PROJ_ROOT - the strict (namespaced) root

Pointing at this is opt-in. It's useful if you keep a separate directory for "Claude Code project" workspaces (think one folder per long-running theme, like ~/cc-claude-code/migration/, ~/cc-claude-code/oneshot/):

export CLAUDE_SESSION_TOOLS_PROJ_ROOT="$HOME/cc-claude-code"

Under the strict root, two extra rules apply:

1. Project directory names must match [a-z0-9]+ - lowercase, no dashes, no underscores.
2. Tag suffixes must start with <project-name>- followed by a descriptive label.

So in ~/cc-claude-code/oneshot/, ccd oneshot-config-cleanup is fine but ccd config-cleanup is rejected with a friendly error. The strict root also enables ccd's Levenshtein typo prompt: if you type oneshet-foo, it offers to correct it to oneshot-foo.

You can configure either, both, or neither. With neither set, you'll need ccd --force to start any session, and ccr/ccs won't find anything.

Why two roots, and which should I use?

	REPO_ROOT (loose)	PROJ_ROOT (strict)
Where you point it	A directory you already use for code, e.g. ~/repos	A purpose-built directory for Claude Code workspaces, e.g. ~/cc-claude-code
Naming conventions	None beyond no-spaces	Project name [a-z0-9]+, tag <project>-<label>
Typo protection	Off	On (Levenshtein-checked against project name)
Best for	Day-to-day work in existing repos	Long-running, themed Claude Code projects you want kept tidy

Most users only need REPO_ROOT. Configure PROJ_ROOT later if you find yourself wanting tighter conventions for a specific subset of work.

Session management CLIs

ccd - start a session

cd ~/repos/myproject
ccd bugfix-flaky-test
# Creates  ~/repos/myproject/cc-sessions/20260509-bugfix-flaky-test/
#                                            working/
#                                            out/
# And launches `claude` with -n 20260509-bugfix-flaky-test (tagged display name).

Useful flags:

Flag	What it does
--dry-run	Print what would happen (session dir, launch command) without creating anything or launching claude.
--force	Skip the root check and any naming-convention checks (escape hatch for one-off invocations outside your roots).
--debug	Enable verbose debug output (CCX_DEBUG=1).

Anything after the tag is forwarded to claude verbatim, so ccd my-tag --model opus works.

ccr - resume by fragment

ccr flaky        # resumes whichever session has "flaky" in its name
ccr 20260509     # resumes whichever session was started on that date

If multiple sessions match, ccr shows a numbered picker (1-9/0) if stdin is a TTY and there are 10 or fewer candidates; otherwise it prints them and exits. If exactly one matches, it execs claude --resume <basename> with the right working directory.

Useful flags:

Flag	What it does
--include-orphans	Also consider sessions whose cc-sessions/ directory is missing or has been cleaned up (resume by transcript UUID only).
--debug	Enable verbose debug output (CCX_DEBUG=1).

ccs - search

ccs flaky                              # name search in current project
ccs flaky --global                     # name search across all configured roots
ccs "GraphQL retry" --contents        # full-text search of working/out files
ccs "GraphQL retry" --messages        # full-text search of JSONL transcripts
ccs "GraphQL retry" --contents --messages --global  # combined, all projects
ccs flaky --since 2026-04-01          # only sessions from April 2026 onwards
ccs flaky --sort newest               # explicit sort (default)
ccs flaky --sort oldest

Results are ordered newest-first by session start date by default. --contents shows one line of context around each match; --messages searches the Claude transcript JSONL files and surfaces matching turns.

Useful flags:

Flag	What it does
--name	Search session basenames (the default; explicit opt-in).
--contents	Search text files inside each session's working/ and out/ directories.
--messages	Search Claude transcript JSONL files in ~/.claude/projects/.
--global	Search across all configured roots, not just the current project.
--since DATE	Only sessions started on or after DATE. Accepts YYYYMMDD, YYYY-MM-DD, 7d (days ago), 2w (weeks ago), 1m (months ago).
--before YYYYMMDD	Only sessions started before DATE.
--days N	Only sessions started within the last N days.
--sort {newest,oldest}	Sort order (default: newest).
--exclude-hooks	Hide hook-security-check sessions from results.
--json	Output results as a JSON array.
--null	Output null-delimited basenames (for xargs -0).
--debug	Enable verbose debug output (CCX_DEBUG=1).

Usage analytics CLI

claude-code-usage - tokens and dollars by every dimension you care about

The CLI parses your ~/.claude/projects/**/*.jsonl transcripts into a Pandas DataFrame (mtime-keyed Parquet cache means subsequent runs are fast), splits per-tool tokens evenly across tool_use blocks, and lets you slice the result.

Five subcommands:

	What it does
query	Multi-dimensional filter + group-by, output as markdown / CSV / JSON. The workhorse.
report	Render a full multi-section markdown report (project / model / time-bucket breakdowns at once).
children	List child sessions (hook-security-review, subagent dispatches) of a given parent session.
warm-cache	Populate or refresh the Parquet cache without producing output.
reconcile	Compare our totals against ccusage's authoritative figures, so we know the numbers are right.

A few examples:

# Total spend last month, grouped by project (top 10 by cost)
claude-code-usage query --since 2026-04-01 --until 2026-04-30 --group-by project --top 10

# Where Opus tokens went this week, by session
claude-code-usage query --since 2026-05-04 --model opus --group-by session

# How often each MCP server gets used, across the last quarter
claude-code-usage query --since 2026-02-01 --group-by mcp --sort token_total

# Daily spend trend for one project
claude-code-usage query --project myproject --group-by day --format csv

# Full report of last calendar month
claude-code-usage report --since 2026-04-01 --until 2026-04-30

# Cross-validate against ccusage
claude-code-usage reconcile --since 2026-04-01

Run claude-code-usage <subcommand> --help for the full grammar. A few flags worth knowing:

• --exclude-hooks strips out the bash-security-review.sh hook sessions, which would otherwise distort per-session cost breakdowns by ~$1.60 each.
• --include-children (when grouping by session) folds child-session tokens and cost into the parent row.
• --session-format {name,uuid,both} controls how the session column renders - by display name (default), by UUID, or both.

Bundled skills

The repo ships three Claude Code skills, designed to be symlinked into ~/.claude/skills/. They're thin wrappers around the CLIs so a Claude Code session can invoke them on your behalf in response to natural-language prompts.

Install all three at once with ccst skills install --apply (see Install the skills above).

find-claude-code-session

Wraps ccs. Triggers on prompts like "find my session about X", "did I work on foo before", "what session was I in when Y". Constructs the right ccs invocation, escalates from local to global search if the local hit list is empty, and presents results as ccr <fragment> commands you can paste.

move-session

Moves, renames, or move+renames a session directory while keeping the JSONL transcript resumable. Triggers on "move session to", "rename my session", "this session belongs in a different folder". Dry-run by default - you must pass --execute for any filesystem change. Validates source and destination against the same rules as ccd, copies the session directory tree, rewrites JSONL cwd fields to the destination path, and appends a tombstone record to the source JSONL so claude --resume on the old session explains where it went.

claude-usage

Wraps claude-code-usage. Triggers on usage questions: "how much have I spent on Claude Code", "tokens used this week", "which project costs the most", "Opus vs Sonnet", "any spike in usage recently". Picks the right subcommand and flags, runs it, and summarises the result in plain English.

See docs/design.md for the full design and CLI contract.

Hook library (cccs_hooks)

The cccs_hooks Python package provides Claude Code hook implementations. Install via uv tool install cc-session-tools or pipx install cc-session-tools to make the hook library available. Hooks are invoked through ccst hooks run <name>.

Modules

Module	Hook script	What it does
cccs_hooks.telemetry	—	Writes structured JSONL to ~/.claude/hooks/fires.jsonl; used by other modules.
cccs_hooks.transcript	—	Walks parent session transcript JSONL; shared by confirm_8digit.
cccs_hooks.confirm_8digit	enforce-8digit-confirmation.sh	8-digit confirmation guard for gated tools.
cccs_hooks.cache	—	SHA-256 command cache (CSV); used by bash_security_review.
cccs_hooks.bash_security_review	bash-security-review.sh	Tiered Bash security review with cache.
cccs_hooks.edit_write_audit	enforce-edit-write-audit.sh	PostToolUse sensitive-path + WORKLOG audit.
cccs_hooks.prompt_guard	prompt-guard.sh	UserPromptSubmit credential/injection pattern guard.
cccs_hooks.session_end	session-end-reminder.sh	Stop-event WORKLOG/uncommitted-changes nudge.

Running hooks via ccst hooks run <name>

Hook scripts invoke the dispatcher via ccst rather than calling python3 -m cccs_hooks.* directly. This means CCST only needs to be installed via uv tool install or pipx install - the hook modules do not need to be importable by the system Python. The shim contract is:

exec ccst hooks run <name> <<< "$INPUT"

Where <name> is one of:

Verb	Module
bash-security-review	cccs_hooks.bash_security_review
confirm-8digit	cccs_hooks.confirm_8digit
prompt-guard	cccs_hooks.prompt_guard
edit-write-audit	cccs_hooks.edit_write_audit
session-end	cccs_hooks.session_end

The dispatcher reads the event payload from stdin, calls the matching module's main(), and propagates its exit code.

Running modules directly (debugging only)

Each module is also runnable as a Python CLI if you want to bypass the dispatcher and have cccs_hooks importable on sys.path (e.g. inside an activated venv, or when installed via uv tool install cc-session-tools):

python3 -m cccs_hooks.telemetry log --help
python3 -m cccs_hooks.bash_security_review  # reads JSON from stdin
python3 -m cccs_hooks.prompt_guard          # reads JSON from stdin

Hook management CLI (ccst)

The ccst umbrella CLI provides hook and skill management.

ccst hooks install

Merges hook entries from a source settings.json into ~/.claude/settings.json. Matching is by event type + matcher + command string; already-present hooks are never duplicated.

# Dry run (default) - shows what would be added
ccst hooks install \
  --source /path/to/source-settings.json \
  --target ~/.claude/settings.json

# Write the changes
ccst hooks install \
  --source /path/to/source-settings.json \
  --target ~/.claude/settings.json \
  --apply

The target file is written atomically (.tmp swap).

ccst hooks run <name>

Run a Claude Code hook by name. See the table above for the supported names.

ccst skills install

Symlink all bundled skills into ~/.claude/skills/.

# Dry run (default) - shows what would be created or skipped
ccst skills install

# Write the symlinks
ccst skills install --apply

# Replace wrong-target or conflicting symlinks
ccst skills install --apply --force

How it interacts with Claude Code's task lists

Claude Code lets multiple sessions share a single task list if they all set the same CLAUDE_CODE_TASK_LIST_ID environment variable. ccd and ccr derive this from the project layout:

• If your cwd is a direct child of a configured root (e.g. ~/repos/myproject), the task list ID is set to the project directory name (myproject). All sessions started in ~/repos/myproject share one task list.
• If your cwd is anywhere else (or both env vars are unset and you used --force), no task list ID is set and the session gets a private task list.

This means you can pick up a task created in yesterday's session from today's session in the same project, without any extra setup.

Typo protection (strict root only)

When you start a session under the strict (PROJ_ROOT) root, ccd checks whether your tag's first dash-separated term looks like a typo of the project directory name (Levenshtein distance ≤ 2):

cd ~/cc-claude-code/oneshot
ccd oneshet-fix-bug
ccd: 'oneshet' looks like a typo of project folder 'oneshot' (Levenshtein 1).
ccd: Start session with tag 'oneshot-fix-bug' instead? [y/N]

A second prompt fires if your first term is far from the current project name and far from every sibling project under PROJ_ROOT - in that case ccd offers to prepend the current project name. This behaviour is intentionally off under the loose root.

Sessions on disk

Each session directory looks like:

cc-sessions/20260509-bugfix-flaky-test/
  working/      # scratch files, notes, WORKLOG.md - whatever you want
  out/          # deliverables you might keep or hand off

Add cc-sessions/ to your project's .gitignore if you don't want session artefacts tracked.

Development

git clone https://github.com/raffishquartan/claude-code-session-tools.git
cd claude-code-session-tools
uv sync --extra dev
uv run pytest

Tests run on Python 3.11, 3.12, and 3.13 (see .github/workflows/ci.yml). CI also includes an install-check job that runs uv tool install . and verifies all five CLIs start up correctly - the direct guard against the editable-install/worktree failure mode.

When working in a git worktree: test your changes with uv run pytest or uv run python -m cc_session_tools.cli.ccd - do not run uv tool install from inside a worktree. After merging, run uv tool install ~/repos/claude-code-session-tools (or uv tool install cc-session-tools if installed from PyPI) to update the global install.

Limitations and caveats

• Linux and macOS only. Windows is not tested; the tools assume POSIX paths and os.execvpe-style process replacement.
• The session-management CLIs shell out to claude via os.execvpe. If claude isn't on $PATH, ccd and ccr will fail with the standard "command not found" error.
• claude-code-usage reads from ~/.claude/projects/ and writes a Parquet cache under ~/.cache/claude-code-usage/parquet/ (overrideable via --projects-dir and --cache-dir). Pricing data is loaded from data/pricing.json shipped with the package, refreshed lazily from LiteLLM upstream with a 7-day TTL.
• The strict-root convention is opinionated. If you don't want it, just leave CLAUDE_SESSION_TOOLS_PROJ_ROOT unset.

Licence

MIT - see LICENSE.