Skip to content

Latest commit

 

History

History
383 lines (313 loc) · 32.8 KB

CHANGELOG.md

File metadata and controls

383 lines (313 loc) · 32.8 KB

Changelog

This file records every user-visible change to Doberman, one line per change, newest release first. Planned work lives on the roadmap board; each release also gets its own notes on GitHub.

Not yet released: changelog.d/, compiled into the next version.

v0.21.0 — 2026-11-15

Opt-in Windows Hello-required mode: 2FA approvals become Hello signatures, no secret on disk

Added

  • Windows: opt-in Hello-required mode (doberman 2fa hello enroll) makes 2FA prompts a Windows Hello gesture and deletes the TOTP secret, so none exists to steal (#802)
  • Hello-required mode: approvals are Hello-key signatures checked against the key pinned at enrollment; tool-call signatures are kept for later checking (#833)
  • Hello-required mode: loosening commands take a Hello gesture, and doberman doctor passes the Windows agent-sandbox check once no 2FA secret remains (#834)

Changed

  • Windows Hello approval no longer uses winsdk, which had no wheels past CPython 3.12 (#800)
  • Cursor hooks now wait 660 seconds, like Claude Code, so a slow Hello approval is not cut off; re-run doberman install-hooks --host cursor (#834)
  • Windows Hello now runs through the built-in Windows PowerShell 5.1, so Smart App Control does not block it; the winhello extra is deprecated and empty (#837)

Fixed

  • Windows Hello verify and key-creation prompts are now bound to the window in front, like signing, so they should not open behind other windows (#838)

v0.20.0 — 2026-11-13

Safer dashboard approvals, single-use 2FA codes, and a shadow judge that works

Security

  • Fixed Git push protected-branch bypasses from option variants (-qd, --del, --force-w) and indeterminate refs (#742, thanks @shravan606756)
  • WebhookAuditSink logs only the webhook host, never the full URL, when a POST fails (#775, thanks @harmehak0173)
  • A TOTP code is accepted once: a code that already approved an action, or an older one, is refused while still in its window (#776, thanks @dchaudhari7177)
  • doberman plugins enable trusts a plugin for one entry-point group, so a same-named plugin elsewhere stays off; old bare names still work (#779, thanks @dchaudhari7177)
  • doberman dash answers only loopback Host headers, so a DNS-rebinding page gets a 400 instead of the dashboard (#793)

Added

  • Deleting a gitignored directory that git never committed now needs authentication; build and cache directories are exempt (#653, thanks @harshitagrawal2O)
  • doberman dash answers AUTH prompts from Claude Code, Codex and Cursor hooks; an approval expires when its 90-second window closes (#764)
  • Payment-card detection now recognises Diners Club, JCB and UnionPay numbers, from the Discover Global Network IIN table; Luhn still required (#777, thanks @dchaudhari7177)
  • LocalEgressBroker takes a list of enforcement probe targets, tried in order, so the self-check works where github.com is unreachable (#780, thanks @dchaudhari7177)
  • python -m doberman runs the CLI when the doberman command is not on PATH (#795)
  • doberman doctor checks whether Claude Code's sandbox keeps your 2FA secret from the agent; Limitations explains how a script can answer its own prompts (#797)

Changed

  • The optional judge reads its key from DOBERMAN_JUDGE_API_KEY, so you never need to export ANTHROPIC_API_KEY (#796)

Fixed

  • Fixed OpenClaw workspace policy resolution (#745, thanks @shravan606756)
  • Bare-host fingerprinting skips 29 more filename extensions such as .mjs, .pyi, .db and .patch; no registered TLD added (#778, thanks @dchaudhari7177)
  • The optional shadow judge now parses Haiku's fenced JSON reply; before, every live call abstained (#794)

v0.19.0 — 2026-10-18

An ambient activity monitor, tighter command and hook checks, and a wave of contributor fixes

Security

  • doberman doctor now fails when Claude Code's disableAllHooks turns every hook off, instead of reporting the hooks intact (#691, thanks @vortsghost2025)
  • A host-hook delete whose path was long enough to be redacted still gets its blast-radius preview and post-approval recheck (#692)
  • Path-qualified or .exe spellings of destructive verbs and git --attr-source <tree-ish> no longer slip past the command scanner (#693)
  • The approval prompt and phone push mask bearer tokens and password flags, not only name=value secrets (#694)
  • Activity path classes reject raw credential filenames like .ssh/id_rsa and accept wildcard and Windows drive classes like /etc/* and C:/Users/x/.aws/* (#700, #707, thanks @Souptik96)
  • git push --delete/-d and git push origin :<branch> against a protected branch now block like a force-push (#711, thanks @AmirF194)

Added

  • doberman monitor run and status score non-inline activity through the same engine; alert-only, never enforcing (#699, thanks @Maqbool61)
  • doberman decision-log-prune --dry-run lists the rows a prune would delete without deleting them (#735, thanks @harmehak0173)
  • doberman doctor warns about an enabled plugin no installed package provides, and plugins enable says so too; neither loads plugin code (#738, thanks @dchaudhari7177)
  • doberman log --jsonl and the dashboard feed carry each decision's session_id; the feed also names agent_role, auth_path and human_confirmed (#756)

Fixed

  • A root-level conftest.py and a shell mv/git mv of a test file now raise AUTH instead of passing silently (#689, thanks @Som0111)
  • doberman approvals ttl changes, including denied raises, now appear in doberman policy-history (#746, thanks @Santoshkumarpuppala)
  • doberman tune no longer counts ambient-monitor rows toward friction rates or standing-elevation proposals (#708, thanks @Maqbool61)
  • doberman dash refuses a busy or out-of-range --port before printing the URL, and a negative --last is a usage error (#736, #737, thanks @dchaudhari7177)
  • doberman doctor and approvals status refuse a --path that does not exist instead of creating it (#740, thanks @dchaudhari7177)
  • The MCP proxy reuses a SQLite connection only within its own task and event loop, so AUTH workers and child tasks get their own (#656, thanks @be-student)
  • Setup step totals skip refused tuning, a bad message tone lists the valid ones, and policy-versions --show --json prints compact JSON (#701, #733, #734, thanks @a27851305-hue, @harmehak0173)

Docs

  • docs/CLI.md documents plugins, policy-file and memory seed, and a test fails when a visible command is missing from it (#739, thanks @dchaudhari7177)

v0.18.7 — 2026-09-09

Approve or deny from your phone, and three ways past the command scanner are closed.

Security

  • Active-role and artifact-pin config are re-parsed only when file content changes, closing a same-tick edit that could serve a stale role or pin (#552)
  • Destructive commands wrapped in builtin or eval are now detected instead of silently passing (#555)
  • A post-approval BLOCK now revokes the role elevation just granted, instead of leaving it active until its TTL expires (#557)
  • The decision log classes extensionless paths under a directory (.ssh/id_rsa) like any other filename instead of storing them verbatim (#605)
  • A destructive or egress command wrapped in strace, flock, unshare, or taskset now classifies the same as its unwrapped form instead of silently passing (#654)
  • The MCP admission scan now flags whole-script look-alike server names, not only mixed-script ones (#655, thanks @Som0111)
  • Claude Code hook entries now pin a timeout that outlasts Doberman's challenge ceiling, so a timed-out hook cannot fail open before Doberman denies (#660)
  • A delete approved through a host hook now shows its blast radius and is re-checked before it runs, denying if the files changed since approval (#665)

Added

  • Ambient activity bus (FM.1): ActivityEvent, SQLite bus, and the doberman.collectors entry-point seam (#512, thanks @Maqbool61)
  • Expanded the dependency-admission popular-package seed (pypi, npm, cargo) with 89 more real, download-ranked names, cutting false AUTH prompts on legitimate installs (#554)
  • The decision log now records a delete preview's file/dir counts and cap/git/outside-repo flags, with the effect digest stored as a keyed HMAC (#556)
  • Phone push approvals through ntfy: doberman phone setup sends two-factor and elevation challenges to your phone with Approve and Deny buttons (#609)
  • protected_branches role.yaml key widens force-push blocking to extra branches, always adding to the defaults, never replacing them (#620)
  • Corpus report emits auth_gated: attacks stopped at AUTH rather than BLOCK, by shape; docs/BENCHMARKS.md gains the table (#659).
  • The decision log now records which code path resolved an authentication and whether a person approved it; both surface in doberman log --jsonl (#663)

Changed

  • Excluded-project checks are re-parsed only when file content changes, cutting redundant JSON parsing on every host-hook call (#618)

Fixed

  • An oversized shell payload no longer stalls the destructive-command scan; a segment over 64 KiB is truncated and asks for approval instead (#549)
  • Read-only git commands (log, grep, show, ...) no longer BLOCK/AUTH just because an argument mentions a force-push or history-rewrite flag (#550)
  • The delete blast-radius preview now flags a .GIT directory the same as .git on Windows, where they are the same folder (#558)
  • The AUTH blast-radius preview no longer understates a delete's size when directories dominate; it now shows both counts on a capped scan (#558)
  • The auth prompt shows the command the agent asked to run, with credential-shaped tokens masked and long paths kept, instead of a wholesale <redacted> target (#599)
  • Every additive schema ALTER now tolerates a racing process adding the column first, not just #614's fix (#619)
  • The update nudge now orders rc/beta/alpha/dev suffixes: an installed rc is nagged toward its final, a final never toward a pre-release (#621; #666)
  • Terminal explanations wrap overlong non-path slash tokens within the configured width while preserving paths and URLs intact. (#657, thanks @be-student)
  • Phone approvals honor the configured --wait on the 2FA tiers; the approval budget no longer caps it at 90 s (#671)
  • TUNING.md phone section rewritten: every AUTH tier reaches the phone, a tap replaces the TOTP code, silence auto-denies after 10 minutes (#671)
  • Telemetry stays silent during pytest runs (PYTEST_CURRENT_TEST is a kill switch), and the daily summary counts approved and denied AUTH prompts (#673)

Docs

  • Plain-English rewrite of the adapter, example, rule_lab, and rules-data READMEs: terms defined on first use, same commands and facts (#600).
  • Plain-English README, CONTRIBUTING, SECURITY, and docs index; new ROADMAP.md states the vision, and Known limitations move to docs/LIMITATIONS.md (#601).
  • Plain-English pass over the docs/ guides: terms defined on first use, shorter sentences, same commands, tables, numbers, and reason codes (#602).
  • PARITY.md: the secret-egress taint floor is now proven on Codex (#632, thanks @jasperdingg)
  • README drops its Known limitations section; the catalog stays in docs/LIMITATIONS.md via docs/BENCHMARKS.md, which also notes the lifecycle-hook vector and judge evidence (#661).
  • PARITY.md: the gitignored-delete gate (AN-1) is now proven on Codex (#664, thanks @harshitagrawal2O)
  • PARITY.md: the policy-weakening row is now proven on every host hook and the MCP proxy, not only the CLI gate (#672)

v0.18.6 — 2026-09-04

Closes detection bypasses and rebuilds setup, the dashboard, and the TUI.

Security

  • The destructive-command rule now scans a command-shaped argument whatever the tool is called, closing a gap where renaming the tool avoided detection (#519)
  • Every rule/detector/auth-provider plugin is opt-in by name (doberman plugins enable <name>); role elevation now always also asks the local provider (#520)
  • A password lockout now persists to disk across restarts, with the same 15-minute cooldown TOTP already had, instead of resetting on every new CLI process (#521)
  • A single-use elevation grant is now claimed atomically; a losing concurrent call is blocked instead of both being allowed through (#522)
  • The destructive-command rule and the egress walker now scan both sides of a single &, since both cmd.exe and POSIX shells run both commands (#524)
  • A host-namespaced MCP tool (mcp__<server>__write_file) now classifies by its real tool name, so path and role rules apply instead of treating it as unknown (#525)
  • Password guesses are counted before the hash runs, so parallel guesses can't dodge the lockout; an unrecordable attempt is refused (#526)
  • A tool call's command line is rebuilt from command and args with boundaries kept, so a list-form payload can't dodge the scanner (#527)
  • doberman setup --yes can no longer lower the security mode or preference weights past the possession-factor gate (#530)
  • The GUI auth dialog no longer lets a keypress during closing flip an expired denial into an approval (#536)
  • An unanswered approval now expires within 10 minutes instead of 20, so a two_factor challenge nobody answered can't stay approvable past its real deadline (#545)
  • Reverse shells (nc -e, socat EXEC:, an inline socket that spawns a shell) now hard-block instead of prompting; probes and plain sockets are unchanged (#570)
  • A Tk auth dialog no longer changes the hook's exit code on teardown, where a denied or expired action could execute if the hook crashed (#575)
  • Mail addresses obfuscated as [at]/(dot), spaced @, or bare at/dot words in untrusted input now trip the echo tripwire like literal ones (#578)
  • The destructive-command rule now sees through shell syntax, sudo/env/nice/timeout-style wrappers, and nested command substitutions instead of stopping at them (#580)
  • Process kills (kill, pkill, taskkill, Stop-Process, os.kill) and interpreter one-liners that spawn a subprocess now require authentication (#580)
  • Every command-scan bound now fails upward to authentication instead of skipping, and a package-manager verb hidden behind wrapper options no longer gets the implied-registry pass (#580)

Added

  • Every saved policy now gets a content-hash version in .doberman/policies.db; doberman policy-versions lists them, --show prints one, --verify checks the catalogue (#513)
  • doberman setup now asks which hosts to guard, with detected hosts preselected, --host repeatable, and a doctor pass at the end (#528)
  • The dashboard gets verdict/text filters, keyboard shortcuts (/, arrows, a/d, ? for help), a live countdown on pending cards, and a light/dark toggle (#533)
  • doberman tui is now a full decision browser: verdict/risk chips, a full-screen why view, filtering, and jump/copy keys; doberman log --why prints the same explanation (#534)
  • doberman setup --dry-run now previews the mode, preferences, and every file it would write, with nothing persisted (#535)
  • The GUI auth dialog is rebuilt: a bounded, expandable command panel, a severity chip for risk, and a live countdown extendable up to 10 times (#536)
  • The destructive-command rule now recognizes raw-socket egress shapes (/dev/tcp, netcat/socat exec, openssl s_client, an inline socket payload) and steps them up to AUTH (#538)
  • Added an experimental, offline-only BYO-model judge behind the [judge] extra; it is not wired into any live decision yet (#539)
  • A new verification-integrity rule pack requires authentication for git commit --no-verify-style bypasses and for deleting or renaming a test file (#541)
  • A new offline dependency admission gate blocks a known-malicious package name and authenticates a likely typosquat (one character off a popular name) (#543)
  • A host, URL, or email seen in a WebFetch/WebSearch result now raises a later egress to that exact value from allowed to authentication-required (#547)
  • A delete-class command (rm, del, Remove-Item, ...) reaching an AUTH challenge now shows a bounded file/directory count, and re-blocks if the filesystem changed since approval (#548)
  • install-hooks now fingerprints its own hook registration and warns if a hook is later stripped or changed; doberman doctor reports intact, diverged, or untracked (#561)
  • Benchmark harness gains RedCode-Exec, MSB, and LLMail-Inject suite adapters (DOBERMAN_BENCH_*_DIR, no data vendored); measured numbers and documented gaps live in docs/BENCHMARKS.md (#562)
  • Opt-in --replay-session harness mode replays each case in an isolated session with the real post-decide floors; every report is labeled session_replay: true/false (#562)
  • Command-shaped benchmark actions classify egress through the proxy's own destination extractor, so the harness measures shipped behavior instead of under-reporting it (#562)
  • The auth dialog window now carries the Doberman mark as its icon instead of Tk's default feather (#565, thanks @thesageak)
  • Experimental Cursor native-hooks adapter: doberman hook cursor gates Cursor's tool-use events on the same decision engine as Claude Code and Codex (#568)
  • doberman install-hooks --host cursor wires Cursor's hooks and a session heartbeat; doberman doctor gets a new "Cursor hooks" check (#571)
  • doberman serve --url now fronts a remote MCP server (Streamable HTTP or --transport sse) through the same policy chokepoint as a local command (#574)
  • A repo-committed doberman.policy.yaml is now resolved on every action, so teams can review policy changes in pull requests; dropping a glob needs doberman policy-file --accept (#579)
  • doberman memory seed --from traces.jsonl warms a fresh install's behavioral baseline from operator-supplied allowed-action traces (#579)
  • whole_script_confusable flags a token written entirely in Latin-lookalike Cyrillic or Greek letters (an all-Cyrillic "paypal") and steps it up to AUTH (#579)
  • A doberman.cost_observers plugin can now receive loop anomalies via an optional on_loop_anomaly hook (#579)
  • A worked example detector plugin ships at examples/plugin-detector/, alongside docs/EXTENDING.md documenting every entry-point group (#579)

Changed

  • doberman setup now shows the mode in force, scopes its doctor pass to your wired hosts, and offers doberman demo before you leave (#530)
  • Deny is now the dashboard's primary action and needs the same two-step confirm as Approve; lowering the security mode needs the same confirm gesture (#533)
  • doberman setup now ends honestly: -- Setup complete/pending/incomplete -- with exit codes 0/1/3, and its output wraps and colors like doberman status/doctor (#535)
  • The subjective-layer benchmark diagnostic reports a held-out-benign false-positive rate next to the AUC and gains a seeded --suite devsession corpus that engages the full ensemble (#542)
  • Tk auth dialogs show one decision per screen: subtitle, reassurance, and keyboard note fold into "What is this?"; the severity word prints once (#590)

Fixed

  • The auth dialog now shows a keyboard hint (Tab/Arrows: switch - Enter: confirm - Esc: deny) and grows to fit it instead of clipping (#507, thanks @thesageak)
  • The dashboard's mode-change form now actually closes, and approve/deny, live-feed, and refresh failures show in the UI instead of failing silently (#533)
  • An unanswered approval-memory lookup no longer hangs the hook, falling back to a prompt after 5 seconds; an aborted setup run reports what was written (#535)
  • The GUI auth dialog now scales correctly under DPI scaling and works with screen readers (#536)
  • Fixed a doberman tui crash on startup caused by a mount-timing race in the first row's highlight (#546)
  • A mailbox destination (local@domain, mailto:) no longer triggers a false "embeds credentials" prompt on every mail send in Light/Balanced (#564)
  • doberman hook cursor now reads stdin as raw UTF-8 bytes, so the BOM Windows cursor-agent emits no longer turns every hook into a deny (#571)
  • doberman hook pre now answers Cursor-shaped payloads instead of failing closed, which had denied every Cursor shell command on machines with global hooks installed (#576)
  • The test suite no longer touches your real Claude settings: every test gets a throwaway home and a guard fails the run if it changes (#577)
  • The store skips its schema migration when already current, about 11 fewer fsyncs per decided action (#591)

Docs

  • Added docs/AUTHORITY_TIERS.md, documenting which layer of a decision may BLOCK versus only step up to AUTH (#544)
  • Added docs/CONTROL_COVERAGE.md, mapping Doberman's controls to the OWASP Top 10 for LLM Applications and the NIST AI RMF (#579)
  • docs/BENCHMARKS.md: RedCode rows re-measured on main after the reverse-shell and command-walk hardening; in-scope ASR 0.104 → 0.010, only 7 print-only variants remain (#584)

v0.18.5 — 2026-08-30

Decision-log retention, update nudges, and dashboard polish.

Added

  • doberman decision-log-prune deletes resolved rows by age or row budget; pending approvals and the policy ledger are untouched (#461, #502)
  • doberman update checks PyPI once and prints the upgrade command; doberman status nudges when you are behind. Off under DO_NOT_TRACK, CI, or DOBERMAN_UPDATE_CHECK=off (#508)
  • Dashboard: a Copy details action on each pending card copies the redacted fields as JSON (#498, thanks @slegarraga)
  • doberman doctor reports the password factor and the optional dash/tui extras (#474, #475, thanks @slegarraga)
  • doberman egress-velocity [KNOB] [VALUE] shows or sets the burst, volume, and fan-out detection thresholds (#459, thanks @Maqbool61)

Changed

  • Changelog entries now compile from a per-PR changelog.d/ fragment instead of one shared file, so parallel PRs stop colliding (#476, thanks @slegarraga)

Fixed

  • install-hooks --dry-run now previews the exact command the installer writes (#463, thanks @slegarraga)
  • tune --json output is compact like every other JSON command (#470, thanks @slegarraga)
  • doberman log columns no longer shift for long action types like network_request (#472, thanks @slegarraga)
  • The dashboard stats strip now updates immediately instead of lagging behind the live feed (#491)
  • The dashboard header shows the real Doberman mark instead of a placeholder "D" (#490)

Docs

  • docs/README.md indexes every doc page (#492, #500, thanks @navaneethsankar07)

v0.18.4 — 2026-08-26

Friction reduction, part three: tap-to-approve 2FA, repeat-approval memory, safer uninstalls, and default telemetry.

Added

  • 2FA can be approved with a Windows Hello or Touch ID tap instead of a TOTP code (doberman 2fa methods enable)
  • doberman demo --quiet suppresses narration and prints only the summary line and exit code, for CI smoke tests
  • doberman doctor flags dangling hook entries when the installed doberman binary is no longer on PATH
  • doberman uninstall --global removes hooks, project state, auth factors, and the package device-wide, gated the same as any other removal
  • Repeated approvals get a five-minute memory: an identical action re-prompts with a one-click confirm instead of the full auth ladder

Changed

  • Breaking: anonymous usage telemetry is now on by default (five allowlisted events only, no paths/prompts/secrets); opt out with doberman telemetry off

Fixed

  • doberman uninstall now actually stops protection when hooks are installed globally, not just for the current project

v0.18.2 — 2026-08-26

Fewer spurious secret-detector prompts, opt-in telemetry, and a docs rewrite.

Security

  • Re-approving a changed tool pin now resets its learned familiarity and revokes the tool's scope tokens immediately

Added

  • Anonymous CLI telemetry is available as an opt-in (doberman telemetry on|off|status)
  • doberman tune reports friction telemetry and proposes possession-gated standing-elevations (#403)
  • MCP tool schemas are pinned on first use; a later mismatch requires doberman tools approve <tool_name> (#394)
  • Card numbers, IBANs, and SSNs in an outbound payload to an external destination now require authentication (#392)
  • A shell command that only dumps the environment (env, printenv, export, PowerShell Env: listing) now requires authentication (#455, thanks @QY-25123)
  • OpenTelemetry AuditSink forwards redacted decisions to any OTLP/HTTP collector (#368, thanks @Maqbool61)
  • doberman scan --mcp statically scans known MCP configs for risky patterns without running servers (#393)
  • The dashboard can change strictness mode directly, gated like doberman mode; live-feed rows now show risk level and source context
  • Every GitHub release now ships a CycloneDX SBOM (sbom.json) listing exact resolved dependencies

Changed

  • doberman message-tone human|technical switches auth-prompt wording between plain language and the technical format; human is now the default
  • The auth dialog and dashboard are restyled onto Doberman's brand system, with a live ON GUARD/ALERT status pill and a per-project dashboard tab title
  • Destination hosts in the decision log are stored as HMAC fingerprints instead of plain names

Fixed

  • The GUI auth dialog no longer silently fails to render off the main thread on macOS (#453, thanks @harshitagrawal2O)
  • The secret detector no longer flags ordinary identifiers, paths, UUIDs, or build tags as leaked secrets, and fails closed if it can't evaluate
  • Clicking a keyboard-highlighted auth-dialog button (Deny or Approve) now actually registers the click

Docs

  • Every page under docs/, the README, and CONTRIBUTING.md were rewritten for accuracy; new site at docs.trydoberman.dev

v0.18.1 — 2026-08-15

A documentation fix for broken images on PyPI.

Docs

  • The README's logo and demo GIF now use absolute URLs so they render correctly on the PyPI project page

v0.18.0 — 2026-08-15

Security-audit fixes, a new detection and role-governance layer, and CLI polish.

Security

  • The proxy output-secret gate now covers error results and structured or embedded response channels, closing a leak path (#378)
  • Control-plane state (the TOTP seed, password hash) and Windows-style path separators are both now correctly recognized as protected, closing two bypasses (#379, #380)
  • The lethal-trifecta floor (sensitive data, untrusted provenance, and an external destination) now also fires at the objective layer in Strict and Paranoid modes
  • The destructive-command rule now recognizes PowerShell and cmd deletes (Remove-Item, rmdir, del, Format-Volume), not just POSIX rm/dd/mkfs
  • Deleting an unrecoverable gitignored file (*.db, .env, *.key) now requires authentication instead of passing silently
  • The dashboard no longer writes its signed session URL token to server logs (#286)

Added

  • doberman uninstall removes this project's hooks and .doberman/, gated behind your password or 2FA (#375, thanks @QY-25123)
  • doberman role enable-default turns on a built-in least-privilege role for repos with no hand-written role file
  • doberman memory reset and doberman memory prune --older-than-days N manage the learned per-entity behavioral memory
  • A new detector flags suspiciously large base64-looking blobs in tool-call arguments as possible bulk exfiltration
  • doberman 2fa reset-lockout clears the TOTP lockout early, gated on your local password (#285)
  • doberman --install-completion enables shell tab completion (#280, thanks @slegarraga)
  • WebhookAuditSink posts every redacted decision to your own log pipeline via .doberman/audit_webhook.yaml (#317, thanks @Maqbool61)
  • Python 3.13 is now a supported and tested version (#328, thanks @jasperdingg)

Changed

  • An unanswered AUTH challenge is now logged and reported as a timeout rather than a denial (#281)
  • --help now groups commands into panels, and uninstall-hooks/setup state what's left behind and how to exit (#276, #279, thanks @slegarraga)
  • CLI diagnostics now use one consistent severity vocabulary: error:, warning:, note: (#346, thanks @floze-the-genius)

Fixed

  • The session correlator no longer flags reading a credential when the user's own earlier prompt named the destination
  • doberman doctor's Codex check no longer false-warns on Windows; denial messages no longer assume Claude Code under Codex

Docs

  • README and the setup guide warn that uninstall-hooks must run before pip uninstall doberman-core (#373, thanks @QY-25123)
  • docs/CLI.md documents every CLI command including log --jsonl; new docs/ADAPTER_GUIDE.md covers writing a host adapter (#295, #297, #316, thanks @AshSgDe29071999)
  • Docs confirm the MCP proxy blocks credential-carrying tool output too, not just the host hook (#374, thanks @jasperdingg)

v0.17.1 — 2026-08-07

Bug fixes, machine-readable CLI output, and a session-taint reset command.

Security

  • An AuthProvider plugin raising a non-Exception (like SystemExit) no longer leaves an action neither approved nor denied; it now counts as a denial

Added

  • doberman scan/doctor/policy-history --json and doberman log --jsonl emit machine-readable output through an explicit redaction allowlist (thanks NanoRisk6)
  • doberman scan --quiet suppresses the risk map for scripted use (thanks NanoRisk6)
  • doberman taint clear resets sticky session taint after verifying your strongest enrolled auth factor

Changed

  • doberman dashboard is renamed doberman session-summary; the old name still works (#267)
  • Verdict output is now colored (NO_COLOR-safe) with fixed-width labels, and AUTH prompts state their auto-deny deadline (#252, #265)
  • The dashboard's Approve control now needs two clicks to confirm; Deny stays one click, and light mode meets AA contrast (#261, #268)
  • The setup wizard now re-prompts on a mistyped mode instead of losing your answers (#266)

Docs

  • README states the defense-in-depth disclaimer once (#264, thanks @slegarraga); hook help now points to the real install-hooks command (#263, thanks @AshSgDe29071999)
  • New docs/CLI.md command reference and docs/REASON_CODES.md reason-code list (thanks NanoRisk6)

v0.17.0 — 2026-07-30

The egress-broker extension point, plus AUTH-timeout and path-matching fixes.

Security

  • An unanswered AUTH challenge now denies on a wall-clock deadline instead of hanging indefinitely
  • Trailing dots and spaces in a path are now stripped before matching a protected-path rule, closing a padded-spelling bypass on Windows

Added

  • A runtime egress-broker extension point lets a forward proxy verify and enforce network destinations; dormant until a broker is registered
  • doberman 2fa remove unenrolls TOTP
  • CI/CD config protection now covers GitLab CI, Jenkins, CircleCI, and Azure Pipelines, not just GitHub Actions (#164, thanks @harshitagrawal2O)

Changed

  • The devops role no longer allow-lists GitHub Actions workflows; CI config edits now escalate to authentication like every other role
  • CLI output is now color-coded by verdict and wraps long explanations to the terminal width

Fixed

  • The dashboard now keeps you signed in across a reload, and the 2-second poll no longer clears an in-progress TOTP entry

v0.16.0 — 2026-07-23

Raw shell and package commands are now scanned for hidden network egress.

Added

  • Doberman now parses the raw shell, package, and git command and detects a network destination hidden inside it, even through env or a wrapper
  • A command that both touches a session secret and reaches an external destination now hard-blocks; an ambiguous or unresolvable destination requires authentication

Fixed

  • The reported version number is now sourced from installed package metadata, so it can't drift from what's published

v0.15.0 — 2026-07-19

First PyPI release: the turn gate, a live dashboard, and a demo reel.

Security

  • A hardening pass across 17 pull requests closed proxy and host-hook secret-redaction gaps and persisted the TOTP lockout across restarts

Added

  • The turn gate now extends per-call lethal-trifecta protection across a whole multi-step turn, with provenance inheritance and risk clamping between related actions
  • doberman dash (pip install "doberman-core[dash]") is a local dashboard with a live decision feed, verdict stats, and interactive approve/deny for AUTH prompts
  • doberman demo runs a scripted attack reel — secret exfiltration, rm -rf ~, a force-push, Unicode-smuggled egress — through the real decision engine

Docs

  • New CONTRIBUTING.md guide for external contributors

v0.11.0 — 2026-06-17

First public release: tool mediation, guardrails, roles, and tiered authentication.

Added

  • The decision engine mediates every tool call into an allow, authenticate, or block verdict
  • Objective guardrail rules cover path confinement, destructive commands, external destinations, secret patterns, and smuggled-token channels
  • A subjective guardrail layer builds adaptive per-entity behavioral baselines and flags out-of-distribution and homoglyph token signals
  • Roles and repo boundaries, capability discovery, and tiered auth from a local confirm up to TOTP and scoped elevation
  • A local redacted audit log and raise-only policy-drift defense
  • doberman serve runs a stdio MCP proxy in front of any MCP server