ccdrift

A check for silent changes in Claude Code, read from your own session logs.

Claude Code keeps a transcript of every session on your machine. ccdrift reads them and tells you when something shifts that you'd otherwise only notice on your bill or your usage limits:

• Prompt caching gets worse. Less of each new prompt is read from cache, so Claude Code resends the conversation more often. ccdrift caught a real regression this way (Claude Code 2.1.233–2.1.258, August 2026) and follows such a regression until it's fixed.
• Tool-loop turns start missing the cache on the main thread: each miss writes the whole conversation to the cache again. In subagents, which often carry most of the spend, each day's read-back is judged instead, and a drop opens an incident.
• Haiku appears on the main thread, where your chosen model normally answers.
• A setting Claude Code picks changes: the main thread moves between the 1-hour and 5-minute prompt cache, or its effort level changes, including when a new model arrives with a setting of its own.
• New sessions start with a different amount of context, from the system prompt, tools, MCP servers or CLAUDE.md.
• Stop hooks start failing, as they can when an update changes what hooks receive.
• Requests start failing, or responses start stopping at the token limit, far more often than on the days before.
• A model starts thinking far more per response: thinking is billed as output, so it costs money and usage limits.
• Claude Code stops logging a field ccdrift relies on, so a silent change can't hide as a quiet week.

Each alert names the Claude Code version that was running, and what a regression has cost: tokens re-cached, extra Haiku responses or extra thinking tokens, and quotes the matching lines of Claude Code's own release notes in the log. Claude Code keeps them in the config folder that holds the transcripts: ~/.claude/cache/changelog.md, or $CLAUDE_CONFIG_DIR/cache/changelog.md when that variable is set.

It tells you when responses think far more, but not whether they think less: in one person's logs, effort swings more from day to day than a 70% cut in thinking moves it. See docs/findings.md.

Everything stays on your machine. ccdrift reads the transcripts and keeps a state file, a log, its own history of responses and, once a failing check has notified you, the time it did (check-state.json.last-failure-notice) in ~/.ccdrift. The history holds token counts, models, versions and settings, plus each transcript's path (which includes your project folder names) and session id; no prompt or response text. Only you can read the history and the log, as with Claude Code's transcripts; a log an older ccdrift made becomes private when you run ccdrift schedule install again. If your status line passes ccdrift its JSON (see Quota samples), ~/.ccdrift/quota.jsonl keeps how much of your usage limits was used and when, with the Claude Code version and model, and nothing else from it: no session id, path, project, repository or branch. ccdrift sends nothing anywhere, unless you give it a command to run with --exec.

Install

Needs Python 3.10 or newer.

uv tool install ccdrift
# or
pipx install ccdrift

For the code that hasn't been released yet, install from the repository instead: uv tool install git+https://github.com/rkolesnichenko/ccdrift.

Then schedule the check:

ccdrift schedule install                         # every hour, with notifications
ccdrift schedule install --at 18:30 --no-notify  # once a day at 18:30, log only
ccdrift schedule status                          # installed? how did the last run go?

On macOS this adds a launchd agent. On Linux it adds a systemd user timer, or a crontab line where systemd user sessions aren't available. Installing again replaces the job, and keeps the old one when the new one can't be set up. On Windows, run ccdrift check from Task Scheduler instead.

Hourly runs let ccdrift warn about cache misses within hours, once its history holds 200 or more new prompts in the two weeks before the latest week; the full verdict still takes days. Upgrading from 0.2.0, run ccdrift schedule install again: the job 0.2.0 installed keeps running once a day until you do.

Claude Code deletes transcripts after 30 days by default. From its first run on, ccdrift keeps its own history of every response it has read, so later deletions don't affect it; the history grows by about 200 MB a year for a heavy user, one with about 2,400 responses a day, subagents' included. Each check reads its last 90 days, or its last 60 days of use when those reach further back, and back to an older incident whose cost it works out; ccdrift report and ccdrift incident list read all of it. Its first run can only see what's still on disk, and it replays what it reads, its last 90 days, or its last 60 days of use when those reach further back, day by day, as if it had run all along, so a regression from before you installed it is recorded and reported once. To give that first run more to go on, and to keep transcripts for your own digging, set this in ~/.claude/settings.json:

{ "cleanupPeriodDays": 365 }

When an alert arrives

Each alert is sent once, except responses cut short: a run that deepens to 3 times the share last reported is sent again, naming the level it escalated from, and a regression that simply lasts is repeated about once a fortnight. A transcript that can't be read, or holds no response, alerts again only once a week has passed without one. A failing check is logged on every run and notifies at most once in 20 hours. Claude Code's documentation says the transcript format "is internal to Claude Code and changes between versions, so scripts that parse these files directly can break on any release", which is why the cache-metric alert exists.

Incidents

While an incident is open, its days stay out of the baseline, so a regression that lasts for weeks is still judged against the days before it began. Against a rolling baseline, the August regression looked normal again within 8 days, while 5–10% of prompt turns kept missing the cache.

The first check replays the history it reads, its last 90 days, or its last 60 days of use when those reach further back, day by day and records the incidents it would have followed, with the days it would have opened and closed them on, then sends one alert about them, but only when it found any. ccdrift replay runs the same replay on any install, over all of the history, not just the last 90 days, without recording anything or sending an alert, and says whether each incident it finds is recorded.

ccdrift incident draft prints a Claude Code issue about an incident as Markdown, ready to paste into GitHub: what changed before, during and after it, by version, the release notes that may be related and your environment, as aggregates only. A cache incident's draft also shows what a missed turn looks like.

It drafts four alerts that point at Claude Code itself the same way: session-start, hooks, tool-loop and cut-short, the latest of that kind or the one starting on the day given. A session-start draft says what changed in how the sessions started, a hooks draft which events and tools it reached; two hooks alerts starting the same day are both drafted, one after the other. Claude Code's own tools are named; your projects, skills, agents and MCP servers are counted, never named.

ccdrift incident list                               every incident, its cost and versions
ccdrift incident add cache 2026-08-16..2026-09-04   record one from before ccdrift ran
ccdrift incident close cache                        end the open one as of yesterday (UTC)
ccdrift incident dismiss cache 2026-09-14           a false alarm: its days rejoin the baseline
ccdrift incident draft cache 2026-08-18             an issue draft with the evidence, aggregates only
ccdrift incident draft hooks 2026-09-05             the same about an alert: session-start, hooks, tool-loop, cut-short
ccdrift replay                                      the incidents the check would have followed

Status line

ccdrift status --short prints one line when something needs attention, and nothing otherwise: no check yet, a failing check, no check for 3 days, a check that can't see what it judges (no cache values, transcripts it couldn't read or found no responses in, or a field Claude Code stopped logging, for 3 days from the last check that found it), an open incident, failing hooks, cache misses rising, or tool-loop cache misses rising on the main thread. A state file it can't read shows as ccdrift: can't read state.

$ ccdrift status --short
ccdrift: cache ratio down since 08-18

It reads only the state file and always exits 0, so it's cheap and safe to call from the command your Claude Code status line runs. ccdrift status shows the last run, open and recent incidents, setting changes, and other changes: session start, hooks, days without cache values, transcripts it couldn't read or found no responses in, fields that stopped or started being logged, early warnings, tool-loop warnings, failed requests and responses cut short.

Quota samples

Claude Code's transcripts don't record how much of your plan's usage limits you have used, but the JSON it hands your status line command does, on Pro and Max plans: the share of the 5-hour and 7-day limits used and when each resets. Pass that JSON on to ccdrift and it keeps a sample whenever those numbers change, at most once a minute:

input="$(cat)"
# ... the rest of your status line ...
printf '%s' "$input" | ccdrift status --short --stdin

--stdin prints the same line --short does. Each sample, one line of ~/.ccdrift/quota.jsonl (beside the state file), holds its time, the Claude Code version, the model and the limits' numbers, and nothing else from the JSON. Nothing alerts on the samples yet: a rule on how fast a limit fills for the tokens you use needs weeks of them, which is why collecting starts first. ccdrift status counts them. With no samples after a few responses, your plan's limits aren't in the JSON, which Claude Code sends only on Pro and Max plans.

Alerts elsewhere

--exec runs a command through the shell for each alert, with CCDRIFT_ALERT (flag, recovered, persistent, history, early, loop, subagent_loop, setting, context, hooks, hook_coverage, failed_requests, cut_short, thinking, fields, blank_cache, unreadable, no_responses, digest or failed), CCDRIFT_TITLE and CCDRIFT_MESSAGE set. The new_fields, new_attachments and context_dropped alerts go to the log only and never run --exec. For example, to send alerts to ntfy:

ccdrift schedule install --exec 'curl -s -d "$CCDRIFT_MESSAGE" https://ntfy.sh/your-topic'

Anyone who knows an ntfy topic's name can read what's sent to it, so pick one that's hard to guess.

A command that fails or runs longer than 30 seconds is noted in the log and doesn't stop the check.

Commands

ccdrift check [--notify] [--exec CMD] [--no-digest] [--source DIR] [--state FILE]   what the schedule runs
ccdrift report [--days N] [--by day|version] [--json | --html FILE] [--source DIR] [--state FILE]
ccdrift cost [--days N] [--by thread|agent|skill|plugin|mcp|model|project|branch] [--json] [--source DIR] [--state FILE]
ccdrift status [--short [--stdin]] [--state FILE]
ccdrift incident list [--source DIR] [--state FILE]
ccdrift incident add {cache|haiku|subagent-cache} START..END [--state FILE]
ccdrift incident close {cache|haiku|subagent-cache} [--state FILE]
ccdrift incident dismiss {cache|haiku|subagent-cache} START [--state FILE]
ccdrift incident draft {cache|haiku|subagent-cache|session-start|hooks|tool-loop|cut-short} [DAY] [--source DIR] [--state FILE]
ccdrift replay [--source DIR] [--state FILE]                           incidents the check would have followed
ccdrift peek [--source DIR]                                            the fields ccdrift reads
ccdrift schedule install [--at HH:MM] [--no-notify] [--exec CMD] [--no-digest] [--source DIR]
ccdrift schedule remove
ccdrift schedule status
ccdrift --version

report covers the last 21 days by default, and every version with --by version; cost covers the last 30 days with responses.

report --by version also shows each version's median session start size once it has 3 or more sessions, where automatic compaction started, up to 2 release note lines about caching, Haiku and default models, effort, the system prompt and tool definitions, hooks, or subagent models, and a sub-line under each version for why the cache missed, as Claude Code recorded it. ccdrift report also shows stop-hook runs, the models subagents ran on, and tool-loop cache misses on the main thread and in subagents, which report --by version shows as a share per version.

report --json holds aggregates only: no paths, session ids or project names. It adds a miss_reasons key of counts per cache-miss reason, on the day view's window as a whole and on each version's own record in --by version.

report --html FILE writes the day view as one self-contained page: the cache ratio, Haiku share and subagent read-back drawn per day, with the days of a recorded incident on that metric shaded and the flagged ones marked, a strip of each day's z under the chart with the cutoff across it, a line saying what those marks mean, then the table and the sections the terminal prints. It has no scripts and fetches nothing when opened, so it works offline, and it names your project folders as the terminal report does. Its last line says so, since a page is easier to send on than a terminal.

ccdrift cost says where the tokens went: by default it partitions the window by thread, agent type, skill, plugin, MCP server and model; --by project or --by branch reach the two dimensions the default view leaves out. Every section accounts for all of the window's tokens, not only the attributable slice, so a response can appear in more than one section: one that carries both a skill and a plugin counts once in each. It reports and never alerts: no rule, no threshold, no cutoff turns on any number here. A branch name is scoped to its repository, so a row like main can hold several at once; a row drawn from more than one project folder says how many, which is as far as the transcripts allow, since they record a working directory and not a repository root.

Dollars appear only where ccdrift could fit a price for the model from Claude Code's own cost records. One cutoff decides every dollar figure: a total, or a bucket's, is shown once the models with no price are under 1% of its tokens between them, and withheld above that, because a partial figure would be quietly wrong while an absent one is merely unhelpful. A bucket that withholds names the model that made it, in every section and not only the by-model one, and a line under the header names the models with no price whenever anything in the window has one.

Each model is priced from its own cost records: an input, output and cache-read rate, and a web-search rate when it searched, with cache writes fixed at 1.25 times input, or 2 times for those Claude Code kept for an hour. A response logs which writes those were; a cost record doesn't, so it takes the share from its session's responses. A fit whose error summed over its records is over 1% of their total cost is refused, and so is a model with fewer than four records, five if it searched, or whose cache reads rest on fewer than two.

ccdrift cost --json withholds project and branch, exactly as report --json already withholds project folders, and lists them under a withheld key rather than dropping them without saying so. It carries the evidence behind the money as well: the residual, record count and cache-read ratio of each fitted price under priced_models, and what it could not price, with that model's share of the window, under unpriced_models.

Transcripts are read from $CLAUDE_CONFIG_DIR/projects when that variable is set, otherwise from ~/.claude/projects. With --source pointed at one project's own folder, that folder is read as a single project. The state file, history and log live in $CCDRIFT_HOME, otherwise in ~/.ccdrift, readable by you alone. Schedulers don't see your shell's variables, so a schedule installed while either one is set keeps its value. Install the schedule again after moving or reinstalling ccdrift.

How it decides

For each complete UTC day, ccdrift looks at main-thread responses from the Claude Code CLI (Agent SDK sessions are your own scripts and are left out) and computes:

• the cache read ratio on turns that open with a new prompt, within an hour of the previous response and not right after a compaction;
• the share of responses from a Haiku model.

Over subagents' tool-loop turns (defined below), on days with at least 300 of them, it computes what each turn reads back of what the response before it in its transcript had cached, at most all of it. Each turn's share of its whole input read from the cache would track the work instead: its uncached part is the new tool result.

It counts failed requests the same way, on the main thread and in subagents alike (a request a subagent made is one Claude Code made), and the responses that stop at the token limit or refuse. Those are far too rare for a usual rate (9 failed requests that count, 4 sleep banners and 2 responses cut short in the six weeks this was built on), so each day is judged against the days before it instead.

Each day is compared with the 14 days before it, leaving out the days of open and recovered incidents, using their median and spread, with the spread floored at sampling noise. A day is deviant past z = −3.0 for the cache ratio, z = +3.5 for Haiku share or z = −3.5 for subagent read-back, and a metric is flagged once 3 of any 4 days in a row are deviant. These defaults were tuned on one person's logs; the research harness in lab/ measures how small a change they catch on yours.

Every run also follows new-prompt turns one by one with a likelihood-ratio CUSUM that tests the usual miss rate of the 14 days before the last week against 5%, the August regression's rate. It warns when the sum passes h = 4 (measured in lab/early_warning.py) within the last day, at most once a week, and not while a cache incident is open. It needs 200 or more new-prompt turns in those 14 days, so it stays quiet for the first three weeks or so of history.

A tool-loop turn is a response that doesn't open with a prompt, doesn't follow a compaction and comes within 5 minutes of the previous response in its transcript. It misses the cache when it reads less than half of what that response had cached.

Every run follows these turns one by one on the main thread with the same kind of CUSUM against the usual miss rate of the 14 days before the last week (at least 1,000 turns), against 2% with h = 3 (measured in lab/loop_cache.py). It warns when the sum passes h within the last day, at most once a week, and also while a cache incident is open. Subagents' turns get no warning: no setting measured there caught a planted rise quickly enough without false alarms, so they are judged by the day instead, as subagent read-back.

A session's start is the prompt size (input plus cache tokens) of its first response. The latest 3 sessions are compared with the 10 before them: a change is at least 25%, with each of the 3 more than 12.5% off on the same side.

Stop hooks count on days with 10 or more runs. They are failing when at least half the runs report an error on 2 such days in a row, after at least 5 such days in the 2 weeks before without that.

Thinking is judged per model, on days with 50 or more main-thread responses of which 90% logged a thinking token count: a day's mean count is compared with the median of the model's 14 such days before it (at least 5), when that median is 100 tokens or more. A day at twice the median or more is a rise, reported once and only when that day is within the last 2 weeks, or, for a check that first runs later, when the rise is still going on the model's latest day. A rise that starts within 2 weeks of the model's last raised day is the same one going on unless it is twice that day's level, so a heavy day that recurs, say every week, is reported once. The setting was measured in lab/thinking.py (G16) on one person's logs, whose 2026-09-10 rise it catches on its first day.

For each Claude Code version first seen in the last 2 weeks with 50 or more responses, a field logged on at least 90% of the responses in the 2 weeks before it and on under 10% of the new version's is reported.

Attachment records are counted per session, main thread only, CLI and Agent SDK sessions apart: for each Claude Code version first seen in the last 2 weeks with 5 or more sessions, a kind of attachment ccdrift doesn't read that is on at least 90% of them, and on under 10% of the 10 or more sessions in the 2 weeks before it, is reported once. Kinds that follow what you do, like plan mode or a hook that blocked, stay under that bar.

Claude Code records a reason on a response whose prompt did not match what it had cached. ccdrift counts those reasons in ccdrift report and in ccdrift incident draft, and no rule turns on them.

Linux notes

• systemd user timers run only while you're logged in, unless lingering is on: loginctl enable-linger $USER.
• The timer's units go in $XDG_CONFIG_HOME/systemd/user, otherwise in ~/.config/systemd/user.
• The timer appends to the log with StandardOutput=append:, which needs systemd 240 or newer.
• A timer catches up on a run missed while the machine was off; cron doesn't.
• Which scheduler ccdrift schedule install uses depends on the session: a systemd timer when a user session answers, cron otherwise (over SSH, say). remove and status look in both, and install removes a job the other one holds, so exactly one runs.
• Notifications use notify-send. They usually appear from a systemd timer but not from cron, so with cron, watch the log or use --exec.
• CI installs, checks and removes a real systemd timer on Ubuntu on every push; cron is covered by tests only. Reports from other distributions are welcome.

Uninstall

ccdrift schedule remove
uv tool uninstall ccdrift        # or: pipx uninstall ccdrift
rm -rf ~/.ccdrift                # state, history and log

Development

uv run --group dev --group lab pytest

License

MIT