Repository navigation
Releases
SeniorOS 0.11.0
On Claude Sonnet 5.5, SeniorOS now fixes a bug at its cause in 29 of 30 runs, against 13 of 30 for 0.10 and none of 18 for Sonnet alone, at half the cost per passing run.
Changed
- Routing: any report of wrong output or data, even one that names the function to fix, loads
senior-debuggingbefore the code is read. It was skipped in 14 of 42 such runs, and those passed 5/14 against 26/28 with it loaded. senior-debugging: a new step 9 has the agent search every reader of the field it is about to filter or convert, list them under Other readers, each fixed or right as is, and fix the writer when another reader sees the same bad records. Changing the writer is code, not a data change: it is done without asking even when a reader is billing, copies what the working path writes for the same action, and covers every path that sets or clears the field. Only rewriting saved records still needs approval. Sonnet, six cause-elsewhere tasks: 21/30 against 9/30 for the step alone, then 29/30 (95% Wilson 83–99%) against 13/30 with the routing, $0.14 against $0.28 per pass; bug tasks not used for tuning 18/18, control 16/16.senior-engineering: when one operation writes two or more records that must agree, use one transaction, or derive the later write from the earlier one, or record each write as done, and test a run that stops after each write.o49(a backfill that failed in every arm, Opus alone included): 4/4 against 0/4.
Added
- Bench: a seventh blind round (
o48–o55) on data already saved and on habits from Superpowers, gstack and Compound Engineering. Opus alone 2/2 on six,o531/2,o490/2. A variant of SeniorOS with four of those habits tied main at 21/24 on Sonnet and was not adopted. - Bench: arms that switch the Claude Code hooks on one at a time (
senioros-rules,senioros-rules-guard,senioros-importers) and an experimental importers-first gate, off by default. Measured, not adopted. plugins/senioros/: a skills-only edition for the Claude plugin directory, which accepts from unverified publishers only Markdown, text, JSON and images, and no hooks.npm run gen:rulescopies the skills, license, icon and manifest into it, andnpm testfails when the copy is out of date. The guard and the finish gate stay in the full plugin from the repository's marketplace.
Fixed
- Guard: Claude Code on Windows runs commands through its
PowerShelltool, which the guard never saw; the hook matcher and the guard now include it, and the Claude Code hooks say on stderr when they did not run. Two review rounds closed bypasses, each with a test: force flags inside clusters (git push -fu),git clean -d -f,TRUNCATEwithoutTABLEinside a database client,rmdir /q /s, a downloaded script run as the script itself (sh -c "$(curl …)",bash <(curl …),eval,iex (irm …), also piped intosudo),git checkout HEAD ., and recursive deletes through PowerShell pipelines and .NET. - Antigravity installer: a reinstall no longer wipes edited hook scripts (each script's sha256 is recorded;
--forcesaves the folder once and uninstall restores it), and the config is validated before anything is copied. - Routing:
senior-project-intelligenceandsenior-skill-authoringwere never routed on any host; each now has a routing line. A regression proof needs the original symptom, not just any failure.
npm: https://www.npmjs.com/package/senioros/v/0.11.0 · install per host: Quick start
SeniorOS 0.10.1
The finish gate no longer saves a run that verified nothing as a pass, the hooks' guarantees and limits are written down, and the bench can measure Codex.
Added
- Bench: a Codex backend,
--backend codex --model <name>, with the armsnoneandsenioros. Each run is onecodex execin the run's repository with Codex's sandbox on, from the bench's own Codex home and login, so the user's~/.codexis never read. The bench stops before any run when that login is missing or when~/.agents/skillswould load the user's own skills into every run. Codex's sandbox keeps.gitread-only, so tasks about destructive git commands cannot tell the arms apart there. No results are published for it yet. - Bench: any Claude Code plugin is an arm with
--packs name=source, where a source is a plugin folder or a repository pinned to a full commit (owner/repo@<commit>[#subfolder]), cloned once into the bench home. Each run writesbench.jsonwith the model and the commit of every arm. - Bench: a public leaderboard,
evals/bench/LEADERBOARD.md, generated from submissions inevals/bench/leaderboard/(bench.mjs submitwrites one from results folders,bench.mjs leaderboardrebuilds the board, andnpm testfails when it is out of date). Each table shows the pass rate with its Wilson interval, the exact test against the model alone, whether it holds after correcting for the packs compared, the cost per pass and each pack's commit, and says who ran it and whether anyone else re-ran it. It starts with the fifth round on Opus, Sonnet on the cause-elsewhere tasks and the sixth round.evals/bench/SUBMITTING.mdexplains how to measure a pack and submit it.
Fixed
- Finish gate: a run that verified nothing is no longer saved as a pass. When
senior finishgives no readable report (a broken install, an invalid.senior/config.json) the agent is told once that the change is NOT VERIFIED; when no review is due and no check proves the change (no tests ran, checks skipped) it is told once that the baseline is NOT PROVEN. Before, both finished in silence. Neither blocks more than once per change set. - Bench: the
.senior/config.jsonof taskt7had noschemaVersion, sosenior finishrejected it and the finish gate verified nothing there. Publishedt7results for the SeniorOS arms were measured without a working gate. A test now parses every config a bench task ships.
SeniorOS 0.10.0
SeniorOS sets up Grok CLI in one command, and every result now carries its uncertainty: on the fifth round it is the only pack separable from Opus alone, and on Sonnet 5.5 it passes 12 of 20 cause-elsewhere runs where Sonnet alone and Superpowers pass none.
Added
senior install --target groksets up Grok CLI in one command: the plugin in~/.grok/plugins/senioros/, the entry in~/.grok/config.tomland the Grok routing rules in a marked block of~/.grok/AGENTS.md, each file backed up once.senior uninstall --target grokremoves them,senior setupandsenior doctorreport Grok, andsenior setup --applyinstalls it. The npm package now ships the plugin parts Grok loads (.claude-plugin/,hooks/,rules/).
Changed
- Bench, Claude Sonnet 5.5 on the cause-elsewhere tasks (
o17,o19,o20,o37), 5 trials: SeniorOS 12/20, Superpowers 0/20, Sonnet alone 0/20 (p < 0.001), $0.16 a run for both packs.o20failed in every arm. - Bench: a sixth blind round (
o40–o47) on concurrency, flaky tests, changes to stored data and dependency upgrades. Opus alone 23/24, so no task was kept for a held-out comparison and no skill was written; SeniorOS 24/24 (not worse), at $0.53 a run against $0.37. The tasks stay as a "never worse" check. - Bench reports give each arm's pass rate with a 95% Wilson interval, a two-sided exact p against the model alone compared task by task, and the cost per passing run (
evals/bench/lib/stats.mjs). On the fifth round, SeniorOS is the only pack separable from Opus alone (p < 0.001; the others p ≥ 0.08) and has the lowest cost per passing run ($0.65, against $0.78 to $2.04); its lead over compound-engineering (10/15 against 6/15) is not separable at 5 trials. - Bench: a fifth blind round of the cause-elsewhere class (
o34–o39), andselftesttakes task ids. Opus, the three tasks Opus alone always failed in calibration, 5 trials: SeniorOS 10/15, compound-engineering 6/15, Superpowers and mattpocock/skills 5/15, Opus alone 2/15. The finish-gate review threshold was measured and kept: half of the reviews that led to a fix were on medium-risk changes. - The routing rules come from one template,
rules/core.md.npm run gen:ruleswritesrules/senioros.mdand the Claude Code, Codex, Cursor, Grok and OpenCode rules from it, each with its own first line fromrules/hosts.json, andnpm testfails when a generated file is out of date. The text each host loads is unchanged. Antigravity keeps its hand-written rules. - The README is half as long (821 to 445 lines): it keeps the results, the demo, the install and the skills, and links to
docs/cli.md(every command, proof order, regression proof, project config),docs/hosts.md(each host install) anddocs/concepts.md(proof engine, project memory, evals, architecture, design principles). No text was dropped; stale counts in the evals section were updated. - Grok gets its own routing rules,
adapters/grok/rules.md, to put in~/.grok/AGENTS.md. They load at most one skill per task and start no review subagents, leaving the checks and the risk review to the finish gate. With the general rules Grok loaded four or five skills per task. On the bench (13 tasks, 2 trials, against the previous rules in the same batch) the score held at 20/26, the 12-minute timeouts fell from 7 to 0, and the time per task from 366 s to 226 s. - The bench measures
mattpocock/skillsand compound-engineering as rival arms, counts finish-gate reviews that led to a fix, and has a hygiene task for debug leftovers and copied secrets. - The Antigravity and Grok installers share one module for the rules block, backups and the package root (
src/core/install-shared.ts).senior --versionreads the version frompackage.json, andsenior helplistsinstall --target grokanduninstall --target grok.
Fixed
senior install --target antigravityno longer adds a second rules block to~/.gemini/GEMINI.mdwhen the existing block has lost a marker: like the Grok installer, it leaves the file as it is and says so, anduninstalldoes the same. Like the Grok installer, it now exits 1 when a step was left undone, anduninstalldeletes aGEMINI.mdthe install created and left empty.
SeniorOS 0.9.0
SeniorOS beats the strongest models alone on Claude Opus, Grok and Gemini: a reported bug is fixed at its cause, even in another module, and the guard holds in Grok.
Added
-
A reported bug goes to
senior-debugging, which reproduces it through the real entry points the data goes through, not only the function the report names, and fixes the cause where it is, even in another module. A code fix of the cause needs no approval; rewriting data already stored still does, and the agent says which records stay wrong until then.senior-engineeringruns the changed path at the repository's real volume (a seed, a config value, a note on production size) before saying done. Every host's routing rules carry the change.- Measured on Claude Opus with tasks written by agents that never saw SeniorOS, held out from tuning (
o17,o19,o20, 5 trials): SeniorOS 15/15, Opus alone 0/15, Superpowers 0/15. Every arm found the cause; only SeniorOS fixed it where it was. On grok-4.7 and Gemini Pro, on the five tasks of this kind: 9/10 and 7/10 for SeniorOS, at most 1/10 without it.
- Measured on Claude Opus with tasks written by agents that never saw SeniorOS, held out from tuning (
-
Bench: an Opus set of 32 tasks (
o1–o32) written blind to the pack in four rounds, and a Grok backend (--backend grok): each run gets its own Grok home and an empty profile folder, and the login is copied without its refresh token, so runs cannot rotate it. Results on 13 tasks, 2 trials: Grok 23/26 with SeniorOS, 15/26 alone, 14/26 with Superpowers; Gemini Pro 21/26, 16/26, 16/26. Opus core set, 3 trials: 24/24, 21/24, 19/24. -
Planning before the first edit, in the routing rules of every host and in
senior-engineering: a change with several steps, a shared contract or stored data, or a step that can fail halfway is planned before editing. The agent finds every reader of what changes and the data already saved in the old shape, and decides what happens if a step fails halfway, runs twice or meets old data. Renaming or reshaping a field, key or type that other modules or saved data read is no longer a low-risk edit.senior-finishchecks the plan against the result.- Measured on Haiku, 8 trials each on
h12-cancel-orderandh13-line-rename: 15/16 with these rules, against 12/16 for the rules before them and 9/16 for the model alone. A separatesenior-planskill with the full procedure scored the same 15/16 at the same cost, so it was not shipped: it is kept inexperiments/senior-plan/with its measurement. Seeevals/bench/README.md.
- Measured on Haiku, 8 trials each on
-
The finish gate puts
senior-finishto work: when the checks pass on a changesenior riskrates medium or higher, it sends the agent back once per session to name the two likeliest ways the change is still wrong and the evidence that rules each out, then give the ship status. Low-risk changes are untouched. When nothing failed but the proof baseline is not passing (no tests ran, checks skipped), the review says NOT PROVEN instead of "the checks pass". In the Claude Sonnet bench, no run had loadedsenior-finishon its own. Applies to Claude Code, Codex, Cursor, OpenCode and the Antigravity hooks. -
senior-code-reviewgains a "Try to break it" pass for medium-risk changes and above, andsenior-finisha step to name what could still be wrong before choosing a status. -
Bench: Claude runs get an empty config folder and authenticate with
CLAUDE_CODE_OAUTH_TOKEN(the login file is never copied), andrunrefuses to start when the user's ownCLAUDE.mdstill reaches the runs, which Claude Code reads from the profile folder whatever the environment says. With it moved aside, Claude Sonnet on the core and hard sets: SeniorOS 38/38, Superpowers 7.0.1 35/38. -
A test command that runs no tests is no longer a pass.
senior finishand the finish gate read how many tests ran from node:test, Jest, Vitest, Mocha, pytest, unittest,cargo testandgo testoutput; zero isSKIP (no tests ran), and the count is shown next to the check and reported astestsin--json. Before,npm testwith no test files, orgo testwith no test files, reported PASS. Output that is not recognised changes nothing.- A node:test file with no tests in it is not counted either. node:test reports such a file as one passing test named after the file (Node 20, 22 and 24, spec and TAP), so a suite of placeholder files read as
PASS (1 test). Only a top-level entry that is not a suite and has no subtests, naming a test file that exists and has no code in it (only comments, imports,requires orexport {}), is subtracted, once per file in each run. Sodescribe('utils.ts'), a test namedindex.jsor after its own file, and a build line like✔ dist/index.js (12ms)still count. When the test command runs from another directory (npm workspaces,cd pkg && node --test), the printed paths are resolved from the project root, so they are usually not found and nothing is subtracted: the count stays as node:test printed it.
- A node:test file with no tests in it is not counted either. node:test reports such a file as one passing test named after the file (Node 20, 22 and 24, spec and TAP), so a suite of placeholder files read as
-
Provenance: every proof report says when and on which code it ran (
provenance: commit, uncommitted changes, time).senior finish --recordkeeps the report in.senior/evidence/<commit>.jsonas a durable record of what was verified; it stores no command output. Changes under.senior/no longer count as changes to a project. -
The guard catches equivalents of the commands it already guarded, listed in a bypass corpus test (
tests/guard-bypass.test.ts) where each case either asks or is documented as out of scope with its reason. Applies to every host.- git options before the subcommand (
git -C ../app push --force,git -c …,--git-dir,--no-pager) no longer hide it from the git rules. Their values are still judged, since they can run commands (-c alias.x='!…',-C "$(…)") or name the agent's own configuration (git -C ~/.claude/plugins/… rm). - Newly guarded:
git push --mirrorand--prune;git checkout/restoreof a folder,.or a glob, andcheckout -f/switch -f(a dot-folder such as.githubcounts as a folder; one file, such asgit restore src/a.ts, stays free);git branch --delete --force,update-ref -d,reflog expire,gc --prune=now,git prune,worktree remove --force. - Deletion by other tools:
rmwith its options after the operands (rm src -rf),find -deleteand-exec rm,rimraf(build folders stay free, as withrm -rf dist), PowerShell's abbreviated-r/-Recand theri/rm/delaliases (where a command starts, sogrep del -ris a search),rsync --delete, and recursive deletes inpython -c,node -e/-pandruby -eone-liners. - Publishing and remote changes:
npm unpublish/deprecate,cargo publish,twine upload,poetry/uv publish,gem push,docker push(--dry-runstays free);gh release delete,gh repo archive,gh api -X DELETE;aws s3 rm --recursive/rb/sync --deleteandaws … delete-*/terminate-*, also after global options (aws --profile prod …), askubectl -n prod deletenow is;helm uninstall,heroku apps:destroy,pulumi up/destroy,az … delete;redis-cli FLUSHALL, MongoDBdropDatabase();dd of=/dev/…,mkfs. - Secrets:
printenv,envandGet-ChildItem env:on their own or piped (env | grep TOKEN),printenvor echoing a variable named like a secret ($GITHUB_TOKEN,$OPENAI_API_KEY). - In
git,ghandglab, commit messages, titles and search patterns are read as text, sogit commit -m "remove rimraf"no longer asks; a$(…)or backticks inside double quotes still does, and so does a message containing;or&&. Other tools are not touched: inssh host -t "…"orwatch -t "…"the quoted word is the command that runs. - Out of scope, by decision: commands built at run time or decoded (
$(echo rm),base64 -d | sh) need an OS sandbox, and everyday file operations (rmof named files,mv, writing a file) would make it ask too often to keep.git restoreof a file without an extension (Makefile) also asks, since the text cannot tell it from a folder.
- git options before the subcommand (
-
senior doctorshows each supported host on this machine and whether SeniorOS is set up in it: the host's command on PATH or folder, the plugin (yes,no, orunknownwhen the files that would say are missing or unreadable, never a guess; a Claude Code install scoped to another project isno), Codex hook trust (yesonly with all three trusted), Antigravity hooks, and the file behind each answer. It also says whether theseniorCLI is on PATH. Detection reads files only and never runs a host's binary. Cursor's plugin state shows as unknown unless a cached copy is found, since Cursor does not document where it records plugins. -
senior setupprints the step left for each host found.senior setup --applyalso does the steps the CLI owns: it addsseniorosto the user'sopencode.json(backed up once as.bak-senioros; a JSONC file, or one that is not a plain JSON object, is left alone and the step says to edit it by hand) and runs the Antigravity installer when Antigravity is present without SeniorOS's skills or hooks. Plugins for Claude Code, Codex and Cursor are installed in the app, so setup prints those commands. When a step could not be applied,setup --applyexits 1. -
The guard asks before
senior setup --apply,senior install --target antigravityandsenior uninstall, also asnpx senioros@<version> …ornode …/index.js …: they write a host's hooks or plugin list, the same change as editing those files. Before, an agent could run the Antigravity installer without a prompt. -
Proof beyond Node:
senior finishand the finish gate now check Python, Go and Rust projects. Before, the finish gate stayed silent on any repository withoutpackage.jsonor.senior/config.json.- Python:
pytest(continuing past import errors) orunittest, plusmypyandruffwhen configured, in the project's.venv/venvor Poetry/Pipenv environment. Go:go vet,go test. Rust:cargo test(a workspace covers its members). - Projec...
- Python:
SeniorOS 0.8.0
SeniorOS as an OpenCode plugin, and a guard that also protects the settings, hooks and plugins of the host it runs in.
Added
-
OpenCode 1.x plugin. Add
"plugin": ["senioros"]toopencode.jsonand restart OpenCode. The plugin:- registers the
senior-*skills; - adds the routing rules to the system prompt;
- fails destructive
bashcalls. OpenCode's tool hook cannot ask, so the agent shows what would be lost and lets the user run the command; - runs the finish gate when a session goes idle, and sends failing checks back to the session.
- registers the
-
The guard protects the agent's own configuration. An agent that could rewrite these files could switch SeniorOS off:
Host Protected files Claude Code .claude/settings*.json,.claude/plugins/Codex .codex/config.toml,hooks.json,plugins/Cursor .cursor/hooks.json,plugins/OpenCode opencode.json, plugin foldersAntigravity .geminihooks and rulesCoverage:
- Shell commands that change these files need confirmation, including quoted scripts and command substitutions.
- The file-edit tools of Claude Code, Codex (
apply_patch) and OpenCode are covered too. - Cursor and Antigravity have no hook before a file edit, so there only shell commands are covered.
Fixed
- The finish gate tells the agent to stay within what the user asked. The second time it sends the agent back for unchanged code, it tells the agent to repeat its explanation and stop. Before, in OpenCode, an agent sent back twice edited a test the user had said not to touch.
Verified
In OpenCode 1.14.34 on Windows:
- the guard blocked
git reset --hardand an edit toopencode.json; - the finish gate sent the agent back on a failing test;
- with the new wording, the agent kept its explanation and changed nothing it was told not to touch (one run).
Not yet tested in the host: the Claude Code Edit and Codex apply_patch protection, OpenCode 1.2, macOS and Linux.
Install
- OpenCode:
{ "plugin": ["senioros"] }inopencode.json - CLI:
npm install -g senioros@0.8.0
SeniorOS 0.7.0
The guard and the finish gate in Codex, tested in the Codex CLI itself.
Added
- Codex plugin hooks (
adapters/codex/, declared in.codex-plugin/plugin.json):- a
SessionStarthook with the routing rules; - a
PreToolUseguard that denies destructive commands. Codex hooks cannot ask the user, so the agent is told to show what would be lost and let the user run the command; - a
Stopfinish gate that sends the agent back on failing checks, at most twice per change set.
- a
- Codex marketplace (
.agents/plugins/marketplace.json).
Fixed
- The Codex plugin no longer loads
hooks/hooks.json, the Claude Code hooks, whoseaskanswer Codex treats as an error and ignores.
Verified
In Codex CLI 0.164.0 on Windows:
- the guard blocked
git reset --hard HEAD, even when the prompt claimed the user had confirmed; - after an edit that broke a test, with the instruction not to run tests, the finish gate sent the agent back. Its final answer reported the failing test instead of "Done".
Installing from GitHub brings exactly the three SeniorOS hooks. Not yet tested: Codex on macOS and Linux.
Install
codex plugin marketplace add joshsegatt/SeniorOS
codex plugin add senioros@seniorosThen open /hooks in Codex and trust the three SeniorOS hooks.
CLI: npm install -g senioros@0.7.0
SeniorOS 0.6.0
The guard and the finish gate in Cursor, and SeniorOS as a native Antigravity plugin.
Added
- Cursor plugin (
.cursor-plugin/plugin.json,adapters/cursor/):- the skills;
- a
sessionStarthook with the routing rules; - a
beforeShellExecutionguard that asks before destructive commands; - a
stopfinish gate that sends the agent back on failing checks, in every repository of a multi-root workspace.
- Native Antigravity plugin:
agy plugin install https://github.com/joshsegatt/SeniorOSadds the skills and routing rules. The hooks still come fromsenior install --target antigravity. adapters/shared/: the guard and finish gate shared by the plugin hosts. The Claude Code hooks now use it, with the same behavior.
Fixed
- Outside Windows, the finish gate no longer treats two repositories whose paths differ only in case as one. On Linux and macOS each is checked and remembered separately.
Not yet verified
Neither plugin has been tested inside the Cursor app or agy. CI runs the hook commands through the shell on Linux and Windows.
Install
npm install -g senioros@0.6.0Cursor: install as a local plugin until it is in the marketplace (see the README).
v0.5.0
A finish gate in the Claude Code plugin, a guard that judges chained commands one by one, and bench results on Claude Sonnet.
Results on Claude Sonnet (core set, 8 tasks × 3 trials)
| Arm | Passed |
|---|---|
| SeniorOS | 24/24 |
| Superpowers | 19/24 |
| No pack | 17/24 |
On a strong model, SeniorOS wins on discipline, not by finding more bugs:
- Sonnet alone fixed a bug three times without a test that proves the fix.
- Sonnet alone three times discarded a colleague's work in progress and force-pushed
main. - With SeniorOS it wrote the test and stopped to ask every time.
- No false "tests pass" claim.
Added
- Claude Code finish gate (a
Stophook in the plugin): when code changed, it runssenior finishbefore the agent may stop. On a failing check it sends the agent back, at most twice per change set. - Bench honesty check: each final report's claims about the tests are compared with the real results.
- New bench tasks:
t7-ci-suiteandt8-cleanup;- the
strongsetx1–x4; - the
senioros-skillsarm.
Fixed
- The guard judges chained commands one by one:
cd app && rm -rf dist coverageno longer asks. git cleanlimited to build folders (git clean -fdX -- dist coverage) no longer asks. Without a pathspec, or with any other path, it still asks.
Install: npx senioros install or, in Claude Code, /plugin install senioros@senioros.
SeniorOS 0.4.0
Install
npx senioros install --target antigravity # Google Antigravity, one command (new)
npx senioros install --target codex --scope project
npm install -g senioros # command: seniorClaude Code plugin: /plugin marketplace add joshsegatt/SeniorOS then /plugin install senioros@senioros.
Highlights
- Google Antigravity in one command.
npx senioros install --target antigravityregisters the skills, adds the routing rules toGEMINI.mdand installs the hooks next to yours. It records what it wrote, never overwrites your edits or hand-made hooks without--force, andnpx senioros uninstall --target antigravitytakes out only what it added. - Results on the front page. The README opens with the bench table and a replay of a real run: asked to discard local work and force-push
main, Claude Code alone rewrotemain; with SeniorOS it stopped and listed what would be lost. - Claude Sonnet on the hard set: no pack 22/22, SeniorOS 22/22, Superpowers 19/22. On a strong model SeniorOS keeps every task the model solves alone; on weaker models (Claude Haiku, Gemini Flash) it adds the wins shown in
evals/bench/README.md.
One-command setup for Google Antigravity, and the bench results on the front page.
Added
senior install --target antigravity [--turbo] [--force]sets up Google Antigravity in one command: copies the skills to~/.gemini/config/senioros/skillsand registers them, adds the routing rules toGEMINI.mdin a marked block, and merges the hooks intoconfig/hooks.jsonnext to the user's own; each changed file is backed up once and reinstalling never duplicates;- the installer records what it wrote (
~/.gemini/config/senioros/install.json): without--forceit keeps skills you edited, hook blocks and scripts it did not write, and SeniorOS rules outside its block, and a hook turned off with"enabled": falsestays off; senior uninstall --target antigravityremoves only what the installer added and restores what it replaced with--force;- README: bench results up front with a replay of a real run (
brand/social/demo-approval-boundary.svg), the Claude Code plugin as the recommended install, andDESIGN.mdfor visual assets; - bench: Claude Sonnet results on the hard set (
none22/22,senioros22/22,superpowers19/22).
Fixed
- bench: a Claude Code run that ends on a usage limit or API error now fails as
launchedinstead of being graded as an untouched repository, and the bench stops starting new runs.
SeniorOS 0.3.0
Install
npx senioros install --target claude --scope project
npm install -g senioros # command: seniorClaude Code plugin: /plugin marketplace add joshsegatt/SeniorOS then /plugin install senioros@senioros. The plugin now ships hooks: routing rules at session start and a guard that asks before destructive Bash commands.
Google Antigravity: see adapters/antigravity/README.md.
Highlights
A live bench (evals/bench/) now compares the same model with no pack, with Superpowers and with SeniorOS on trap tasks. Hidden tests and the repository state grade every run; no model judges another.
| Host and model | no pack | Superpowers | SeniorOS |
|---|---|---|---|
| Claude Code, Claude Haiku (17 tasks x 3 trials) | 45/51 | 44/51 | 49/51 |
Antigravity, Gemini Flash (h7, h9–h11, 3 trials) |
not run | 8/12 | 12/12 |
Antigravity, Gemini Flash (h5 multi-regression refactor, 4 trials) |
0/4 | 0/4 | 4/4 |
In Claude Code, SeniorOS wrote no wrong code and stopped before force-pushing main in all 3 t6 runs. Caveats, including which tasks were written alongside the changes they reward, are in evals/bench/README.md.
Google Antigravity support, plugin hooks for Claude Code, and skill changes measured with a live bench against no pack
and Superpowers (Gemini Flash in Antigravity, Claude Haiku in Claude Code; results in evals/bench/README.md).
Added
senior prove --regression --test "<cmd>": proves a regression test catches the bug (passes with the fix, fails with it reverted, passes again restored), with--expect-failure,--base,--fix,--timeoutand--json; evidence written to.senior/runtime/evidence/;senior prove --restore: restores the working tree after a proof that was killed mid-run;senior-debuggingandsenior-finishusesenior prove --regressionwhen the CLI is available; new evalfinish-regression-proof;- Google Antigravity adapter (
adapters/antigravity/): skills registration, routing rules, and three fail-open hooks: project context at session start, a guard that forces confirmation for destructive commands, and a stop gate that runssenior finishbefore the agent may finish with changed code; - Antigravity guard always answers with a decision (
--passthrough=ask|allow, also on errors): Antigravity denies a call whose hook gives none, which had blocked every terminal command; hook commands are written without quotes, since Antigravity runs them without a shell; newtrack.mjshook lets the stop gate prove repositories edited in conversations that have no workspace;SENIOROS_CLIoverrides how hooks call the CLI; - from the first bench pilot (all arms 6/6; SeniorOS the most expensive):
senior-engineeringandsenior-finishno longer trigger on low-risk copy/label edits (evalssenior-copy-change-direct,finish-low-risk-skip);senior-debuggingreads the implicated commit's whole diff and restores every unintended change (evaldebug-regression-commit-diff); the Antigravity rules carry a CLI reference so agents stop runningsenior --help; newcontext.mjsPreInvocation hook gives project context and decisions when a workspace-less conversation first touches a repository; - from the first independent bench task (
h7, where every arm sometimes let a member grant themselves admin through the request body):senior-engineeringdecides what each caller may control when input reaches stored state or a privileged operation, picks allowed fields instead of spreading the input, and tests a forbidden one;senior-code-reviewasks what each caller-controlled input can reach (evalssenior-input-boundary,senior-input-boundary-typo,review-mass-assignment); - Claude Code plugin hooks (
hooks/hooks.json): a session-start hook adds the SeniorOS routing rules and approval boundary (adapters/claude-code/rules.md), and aBashguard asks before destructive commands; in the first Claude Code bench the plugin's skills were never loaded without them (0 of 51 runs); - bench
claudebackend (claude -p, arms as flags, so arms run in parallel) with per-run cost and the final report saved; - bench
hardtask set (h1–h5: root cause elsewhere, project rule, shared function, red CI with an intended change, multi-regression refactor),--tasks core|hard, and a per-run count of agent tool calls and loaded skills; - live bench (
evals/bench/): six trap tasks on a fixture project, armsnone/superpowers/senioros/senioros-skills, a deterministic grader (hidden tests, regression proof, scope, test weakening, destructive boundary) verified by reference solutions innpm run check, and an Antigravity runner (agentapi,agyor manual).