v1.4.4

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

The mod behaves as it did in 1.4.3. The repo has a privacy statement, and the plugin's listing has three links.

• PRIVACY.md says what the mod reads, what it stores and what leaves your machine, for the default settings and for each of the two experiments. The mod has no server, and the author receives nothing from it.
• The plugin's manifest names its documentation (the README), its support page (the repo's issues) and its privacy statement.

By default the mod makes no model calls, no network requests and no file writes, and sends the model nothing.

v1.4.3

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

The mod behaves as it did in 1.4.2. The README states one more thing.

• The file that tests credential redaction, hooks/shadow-candidates.test.ts, spells made-up values in the shapes the redactor has to catch: a GitHub token, a Slack webhook URL, an AWS key id. None is a real credential. The tests read no environment variable and no file, and send nothing.

By default the mod makes no model calls, no network requests and no file writes, and sends the model nothing.

v1.4.2

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

The mod behaves as it did in 1.4.1. The settings menu, the README and the plugin's listing changed.

• /config shows the four settings as free text, not a picker. Each setting's description names the values it takes. A value outside them counts as the default.
• The README has a new section, "What the mod reads, runs and sends". It lists every engine call the two experiments add, the full command of each program they run, and the two environment variables they read.
• The plugin has an icon, and the README header has a logo and a new tagline.

By default the mod makes no model calls, no network requests and no file writes, and sends the model nothing.

v1.4.1

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

The mod behaves as it did in 1.4.0. The README now states five things about the two experiments and the contributor commands that it left out or had wrong.

• The memory shadow log and the steering log are appended to and never trimmed. Each grows until you delete it.
• With XDG_STATE_HOME set to an absolute path, the logs go there and HOME is not checked. The README said a HOME that is not an absolute path was always refused.
• With steer set to on, a note can reach the model while its transcript notice fails to store. The band's ▲ and the sent: true log line still show the note.
• The memory shadow log quotes files, command output and web pages. Treat the log as untrusted text, and read a kept fact before you import it into a memory store.
• The contributor command claude -p x --plugin-dir . sends a one-word prompt on your account when you are logged in.

By default the mod makes no model calls, no network requests and no file writes, and sends the model nothing.

dreeft 1.4.0

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

Timeline

• The time the model spends streaming a tool call's arguments now counts as writing. Tool time starts when the model's response has ended and its tools start to run. Before, a long Write or Edit call read as slow tools while the model was still writing it.
• Tool time is the whole gap between two responses, so a wait on a permission prompt counts as tool time.

New setting: desktop (experimental, off by default)

• on also draws the band in the Claude Code desktop app. The band's drawing there has not been checked. VS Code and mobile get nothing with either value.

New setting: steer (experimental, off by default)

• An experiment that asks whether a short note, appended to a running turn, changes what the model does next.
• shadow: when a turn's context growth crosses 10 points of the window, and again at 20, the mod logs the moment. Nothing is sent to the model.
• on: for half of those moments, picked by a coin, the mod also appends one note the model reads, suggesting a subagent for large reads. A transcript notice and an amber ▲ on the band show each note that was sent.
• Both values write steer.jsonl beside the memory shadow log: one line per trigger and one per outcome.
• on changes what the model reads. No effect has been shown yet: that is what the log is for. The README gives the exact text sent and every field logged.

By default the mod makes no model calls, no network requests and no file writes, and sends the model nothing.

dreeft 1.3.6

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

Band

• Growth that draws as +10.0% is now amber, as the README states for 10 points or more.
• A band given no rows to draw in draws nothing.
• A usage report missing a token field no longer gives a broken growth figure.
• The focus row's ∴ ⟲ N fallback fits the narrowest band: a count too wide for the row drops whole.

Memory shadow log (experimental, off by default)

• The log file is set owner-only (600) on every append, so a file that was readable by others is tightened. Missing parent directories are created with your own umask; only the dreeft directory is 700.
• A HOME that is not an absolute path is refused: nothing is judged and nothing is logged.
• An absolute XDG_STATE_HOME moves the log to $XDG_STATE_HOME/dreeft/memory-shadow.jsonl.
• A turn whose append failed is judged again when its span comes up again, where before it was lost.
• The answer text kept for evidence is capped at 50,000 characters.

README and repo

• README: update, disable and uninstall steps, the band's width rule, each row's edge cases, what the shadow log's candidates are, and the exact commands CI runs.
• CI now tests the stated minimum, Claude Code 2.1.287, and the newest release.
• SECURITY.md says how to report a vulnerability.

By default the mod makes no model calls, no network requests and no file writes.

dreeft 1.3.5

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

Focus row

• A name the turn keeps coming back to now survives a long turn. Past 50 tracked names, the name with the lowest count goes first, the oldest on a tie. Before, the least recently seen name went, whatever its count.
• Justfile, Makefile, Dockerfile, Gemfile and Rakefile count as file names, as do .txt files and a .tmpl after a known extension (chezmoi.toml.tmpl).
• A Windows path in a tool's file argument counts by its last part.
• One malformed tool call in a step no longer costs the focus names of the step's other tool calls.

Memory shadow log (experimental, off by default)

• More credential shapes are replaced with [redacted] before the judge call and the log: vendor tokens (Google, GitLab, npm, Hugging Face, SendGrid, Stripe webhook, Slack webhook, age, AWS temporary key ids), the AWS secret key beside a key id, a password flag after a command known to take one (mysql -p, curl -u, docker login -p, --password), and more secret-named keys and headers (DB_PASS=, pwd:, Cookie:, Authorization: Token). A value holding , or ;, an escaped quote, or a / in a URL password is now redacted whole.
• A tool result is redacted before it is cut to 2,000 characters, so a token split by the cut no longer survives as a fragment.
• The repo's remote is quoted inside the judge's message, not placed on a line above it, and loses its credentials, query string and fragment in every scheme.
• The judge's reply is bounded: fact 400 characters, reason 160, name and topic 60.
• The code stamp on each line now also covers the redaction shapes, the name patterns and the shape of the judge's message.

Redaction is by pattern and can miss a secret in a shape it does not know. The README lists what is covered.

By default the mod makes no model calls, no network requests and no file writes.

dreeft 1.3.4

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

• Memory shadow log (experimental, off by default): each line gains code, eight hex characters naming the selection and judging code that wrote the line. Lines written before and after a change to the second-guess marker, the candidate limits or the judge's rubric can now be told apart. The stamp comes from the code a session has loaded, so a session that kept old code after an update stamps its lines as old.
• The log's schema goes from 1 to 2. A schema 1 line has no code.
• The band is unchanged.

By default the mod makes no model calls, no network requests and no file writes.

dreeft 1.3.3

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

• ⟲ N now counts a second-guess only where the thinking reverses itself: a sentence that opens with "Wait,", "Actually,", "Hmm,", "Oh,", "Oops," or "No,", or one that narrates a change of mind ("I realize", "turns out", "on closer inspection", "reconsider"). The same words used as verb or adverb ("wait for CI", "actually works") no longer count. Over the thinking of 60 recent sessions the old pattern matched 251 times and the new one 52.
• The memory shadow log builds its hedge candidates from the same marker, so its hedge spans change the same way.
• A session that was already running when the mod updated can keep the old marker until the session reloads plugins or restarts.

By default the mod makes no model calls, no network requests and no file writes.

dreeft 1.3.2

A Claude Code mod: a band above the prompt showing what the session's thinking is doing. Three rows: the names the turn keeps coming back to, a timeline of where the turn's time goes, and what the turn costs in context.

Install

Inside Claude Code:

/plugin marketplace add achiurizo/dreeft
/plugin install dreeft@dreeft

Needs Claude Code 2.1.287 or later, terminal surface. The mod API is early access, so a Claude Code update can break the mod until it catches up.

In this version

• A turn keeps at most 240 phase changes. Past that, the shortest phase folds into the one before it, so a turn of thousands of steps no longer grows the state the mod writes on every chunk. The turn's total time is unchanged; a very long turn can lose a sub-second burst from the strip and the totals.
• README: the settings example now uses the key a marketplace install reads, dreeft@dreeft. The bare dreeft key applies only to a clone loaded with --plugin-dir.
• README: the memory shadow log needs /bin/sh (macOS, Linux or WSL), and the list of what the mode sends now includes answer text and the repo's directory name when there is no origin.

By default the mod makes no model calls, no network requests and no file writes.