Repository navigation
Releases
Releases · JamesMoriartyDecripto/flow-tower
v0.18.0
Security
- Squad runner: a literal plan entry whose folder holds a control path needs the plan sign-off, since the coder may write that whole folder (
tools/helper.mjsgrantedtools/, sotools/squad/) (#435).
Added
- Squad: project skills.
tools/squad/skills/<name>.mdholds know-how only some jobs need;squad.jsonskillsgives a role its skills, a task'sSkills:line adds them for one run; names only, size-capped, checked bysquad doctor; with none set, prompts are unchanged. This repo shipsbash-3-2,untyped-mjsandfile-size(#432). - Phone layout: at phone width the tower view shows one panel at a time (Layers, Info, Legend tabs) as a bottom sheet, the bottom bar scrolls sideways, and the mouse and keyboard hint is hidden; the library's hub badge no longer wraps word by word (#436).
Fixed
- e2e: the tower-card test on a phone opens the Info tab, as the phone layout does (#441).
- Squad: the step-scope tests no longer inherit
SQUAD_ADOPT, so a continued step's check is not RED for that alone (#438).
Changed
- Squad:
run.mjs --base <rev>sets thediffbase for one run, and thedifftool starts with the commit list, so a read-only role on a stacked branch judges only its own commits and can confirm them (#425). - Squad: reviewer findings carry the same severity scale as security (critical, high, medium, low;
lowis a note that should not block), so a verified critical or high reviewer finding now stops the gate even with anapproveverdict, a finding with no known severity stops it too (fails closed), and a gate of your own can key on severities instead of an empty findings list (#427).
v0.17.0
Security
- Squad runner: a glob plan entry needs the plan sign-off when a control path (
.github/,tools/squad/,package.json) starts with its text before the first*, sotools/*/runner-core.mjs,**/xor*.jsonno longer reach a control file without a person (#391).
Fixed
- Squad:
SQUAD_ADOPT=<paths>hands a red step's leftover edits to the step that continues it (they were kept out as local edits, so a continuation committed only its own change and still read green); a step that names a kept-out path exactly inFiles:is RED before the coder runs. The README says how to continue a red step and that edge cases get explicit settings, not guesses (#416). - Squad: a
finishsent in the same reply as other tool calls is refused: those calls run and return their results, andfinishmust come alone in a later reply, so a report is never written before the results it describes (#426). - Squad: the
greptool passes the pattern with-e, so a pattern that starts with a dash is searched instead of read as a git option; only git grep's no-match exit reads asno match, any other failure saysgrep failed: …(#424). - Squad:
steps.shmakesREPOand the task folder absolute before it starts, so a relative path no longer turns a green step RED at commit (#394). - Squad:
run.mjscounts a[… cut N KB]marker only when it is a whole line, aspack.mjswrites it, so a pack that embedspack.mjsitself no longer reads astrimmed: 999999 KB(#422). - Squad:
steps.sh's sweep of coder probe files skips the task folder, so a task named like a probe (01-check-probe.md) is no longer deleted before staging (#420). - Sharing: systemd units (
.service,.timer,.socket,.path,.target),.conf,.ini,.cfgandDockerfile,Containerfile,Makefile,Procfileare shared as text; secret names are still refused first, and the extension is read from the file name, not its folder (#407). - Sharing: an agent's required files (in
agents:withoutoptional) enter the loader's referenced files, soshare --filessends them and the hub opens them instead of answering 403; validation output is unchanged (#406). - Squad:
squad doctorchecks that the check command can start (python3 -m pytest: the module is found withfind_spec, never run; any other program is on PATH or, as a relative path, in the repo root) and fails naming it, before a paid step goes red on it (#405). - Squad:
squad.jsontakesstyle: { language, maxFileLines }(defaultEnglish,400); every prompt, task Rules line and review focus uses them, so a project's rules (say, Italian comments and 800 lines) no longer sit next to "English only" and "~400 lines" (#404). - Squad: an empty oneshot answer prints its
finish_reason, tokens (prompt, completion, reasoning) and generation id, and the ledger records its real cost (null when unknown, never 0); an empty answer cut at the output budget is retried once with twice the budget (#412). - Squad:
steps.shpasses each path as:(literal)instead of exportingGIT_LITERAL_PATHSPECS, so a project's pre-commit hook that globs (git ls-files -- 'dir/*') sees its files again (#403); it runs on macOS bash 3.2 (nomapfile) and its private copy gets a random name there (#396). - Scene: a live chip that would cover another one moves above it, so nearby working nodes keep readable chips (#398).
- Live feed: the header counts the scope on screen, so Tower agrees with the
LIVE ERRchip; the Tower scope says how many events matched no node, with a link to them; repeated errors on one node fold into one row (×5); a failed layer stays red in the layer list, with its count, until the feed's Clear (#397). - CLI: a mistyped option prints one line (
flow-tower: Unknown option '--x' … (flow-tower --help lists the options)) and exits 1 instead of Node's stack trace;emittakes--callto pair a start with its end (#400). - Server: inside a library folder only folders,
*.tower.yamlfiles and the files towers reference are watched (nevernode_modulesor hidden folders), and Vite no longer watches the squad's role logs or.stress. A library holding a Python venv used up the system's file watchers, so other servers loggedENOSPCfor every file (#401).
v0.16.1
Changed
- Squad: the coder changes only the lines the task needs; a file that would pass ~400 lines is reported in
Known gapsinstead of trimming other code. The README says each role keeps its tasks in its own folder, sincesteps.shruns every file as a coder step (#386).
Fixed
- Squad runner: an answer is the first json object block that validates, so an echoed template or a quoted draft no longer wins; a role answer is read without
run.mjsheader lines, so a verifier PASS is read (it was always held); an unreadable triage answer is asked for once more;file:~200citations resolve; a person can release one issue's plan sign-off withSQUAD_APPROVE=<issue number>, posted on the issue (#369). - CLI:
reportandexportrun from a checkout path with spaces (the entry guard compared a percent-encoded URL with a raw path), the update command quotes the folder, and the tests pass there (#307). - dev-squad example: the README counts nine layers (the unattended runner is the ninth) and the GLM node no longer claims the architect, which defaults to GPT 6.1 Terra (#371).
- Squad: when
pack.mjstrimmed a review or security task to fit,run.mjsprints atrimmed: N KBheader andgate.mjsadds a note that the pass did not see the whole change, so READY no longer reads as full coverage (#383). - Squad:
squad.jsonrefuses a command with shell syntax (&&, pipes, redirects,$(…), quotes) when it is loaded, andsquad doctorreports it: the runner runs commands without a shell, so such a command could only fail at every step (#382).
Security
- Squad runner: a plan file entry with a comma or a line break is refused, so
src/a.js, .github/xcan no longer pass the control-file sign-off as one path and then reach the coder as two (#391).
v0.16.0
Added
- Unattended Dev Squad runner (#320):
node tools/squad/runner.mjsworks issues labelledsquad:goon one OpenRouter key, with no session and no SDK, one at a time. The procedure is in code (triage, worktree, plan, steps, review and security, lead-judge, fix rounds, verifier, draft PR, gate); the models only judge. It stops at a draft PR or withsquad:failedand the reason. Limits (usdPerIssue,minutes,fixRounds,pollMinutes) are insquad.json, and a lock allows one runner. New rolelead-judge(default Claude Sonnet 5.5) and architect default GPT-6.1 Terra, both from the #362 benchmark. New runner files:runner.mjs,runner-core.mjs,runner-deps.mjs,runner-gh.mjs,runner-plan.mjs,runner-judge.mjs. - Dev Squad tower, unattended runner (#320): a ninth layer shows how
runner.mjsworks labelled issues with no session: pollsquad:go, claim, a code-driven lead, the newlead-judgerole, limits, a draft-PR stop point andsquad:failed. Model choices come from the #362 benchmark: lead-judge Claude Sonnet 5.5, unattended architect GPT-6.1 Terra. Role cardroles/lead-judge.md.
Known limits
- Unattended runner (#320): five external acceptance replays each stopped safely, with the work committed and the reason on the issue; none has yet reached the draft PR on its own. The last stop: a fix that needed a test file outside the fix scope. Project checks run in the runner's process, next to its OpenRouter key and GitHub token (#375): run it only on trusted repositories.
Removed
- Dev Squad Claude Agent SDK backend (#355):
examples/dev-squad/src(the webhook pipeline on@anthropic-ai/claude-agent-sdk), its deploy files, Claude Code settings, prompts, memory and sample logs, and their tests. The squad depends on no SDK: one OpenRouter key, the user's own session as lead in any harness.
Fixed
- Unattended Squad role prompts explicitly require the machine-readable result in the finish report; the judge and verifier receive the original issue alongside the plan, preserving requirements omitted by a summary. A kept finding can name up to 5 supporting files its fix must edit; every path on a fix step's
Files:line must be a concrete regular file outside secret paths, and plans or fixes whose files reach.github/,tools/squad/orpackage.json(also through a directory, glob or+entry) need a person; mentioning those files in the issue text does not stop it (#320). The architect's plan steps hold only code changes and list only the files each must edit; checks, review, verifier and the draft PR stay with the runner (#320). A fix round'sFiles:entries are optional: they bound where the coder may write, so a finding fixed in another file than the one it quotes no longer stops the issue;(optional)entries now count for a step's write scope and commit, as their meaning implies. Findings still kept after the last fix round are listed in the draft PR instead of stopping the issue; a security finding stops it unless its severity is clearlymediumorlow(#320). - Temp files no longer fill the disk (#357). Each test run uses its own
TMPDIR, removed at the end; folders left by crashed runs are swept after 24 h. On 2026-10-31, 2205 leftover folders had filled a/tmpquota.steps.shdeletes its output folder when every step is green, and its own folders older than 7 days.run.mjsprunes role logs beyond the newest 200 that are older than 14 days (new runner filelog-prune.mjs).
Changed
- Dev Squad favours the simplest change (#367): every role prompt asks for the minimum that meets the task, with no abstraction, option or check it does not ask for; the reviewer also flags unneeded complexity and duplicated mechanisms; security names the existing guard before proposing a new check.
- Focus mode draws no cross-layer link (#360): with a layer focused in the tower view, the selected node's links to other layers are hidden too, since they ran up to layers the camera hides. The node panel lists them under "Across layers".
- Dev Squad tower (#355):
examples/dev-squadnow models the OpenRouter runner intools/squad: lead in your harness, triage, plan and task files, coder steps, packed review and security, merge gate, setup, cost ledger and models. The role node ids are unchanged. New role cards (roles/*.md), a samplesquad.example.json, the task template and a nestedrunner-looptower replace the coder and researcher towers.squad initinstalls this smaller bundle. - Squad
steps.sh(#353): each step commits only its own files: the write dirs, the folders of itsFiles:entries and the paths those entries match. It no longer stages every tracked change, and local edits from before the run are never staged; a step that modifies one of them is RED. Steps already committed are skipped on a rerun. The coder may write the folders itsFiles:line names. The script runs from a temporary copy, so a step can editsteps.sh. New runner filestep-scope.mjs. - Focus mode hides cross-layer links (#348): with a layer focused in the tower view, links that do not touch it are no longer drawn, and links that leave it show only for the selected or hovered node and its related nodes, at full strength. They used to stay at 10% and cross the camera's view of the focused layer. Overview and map view are unchanged.
- Dev Squad SDK reference (#347): the deprecated quick-fix route is gone. Triage routes to
full,needs_infoorreject; a trivial issue gets a one-step plan from the lead and the same review, verification and approval gates. The tower no longer shows thequickfixnode. The sensitive-area rule that forces a human plan sign-off now also reads the issue body, not only the title and the model-chosen areas.
v0.15.0 — portable Dev Squad
Security
- Dev Squad approval events (#333, PR #344): a plan approval label needs a current label state, a fresh application event and an actor with current repository push permission. The newer authorized label or comment decision wins; unavailable label evidence and permission errors fail closed.
- Dev Squad SDK reference (#342): test commands use credential-filtered environments and fresh reports; plan approval comments require repository push permission; webhook request bodies are bounded by raw bytes during streaming, before signature verification.
Added
- Dev Squad acceptance tests (#333, PR #344): eight boundary test files with fixture, loopback and mocked evidence. A fresh isolated install runs the nine roles through the installed runner and a trusted CLI; a wrong implementation yields a real nonzero exit; the tower validates with 0 errors and 0 warnings. Scope and limits are in docs/dev-squad-harness-trial.md; no real provider or GitHub call is claimed.
- Dev Squad acceptance evidence (#333): recorded the Codex-led local nine-role baseline with independent tests and final-head review/security; preserved Pi/Claude Code results and incomplete-role evidence. Closure targets the maintainer-confirmed portable workflow; automatic harness launch remains in #320.
- Portable Dev Squad onboarding (#332, #334):
squad doctorchecks local Node/git, runner/config and OpenRouter readiness without paid calls or mandatory Claude credentials. The user's current harness stays the lead. Shared specialist roles now include triage, architect, researcher, tester, verifier and doc-writer, with canonical tower nodes and per-role model selection. Doc-writer requires explicit writable folders. The skill documents the complete lead-driven workflow and its current safety/budget limitations; cross-harness end-to-end evidence remains tracked in #333.
Fixed
- Dev Squad SDK pipeline and quality gates (#333, PR #344): unsuccessful quickfix and quality roles stop delivery; empty or unmet verification criteria, a red or unknown suite and explicit gaps hold. Delivery returns explicit outcomes and SDK CLI exits nonzero unless it reaches
pr_open. The portable runner attempts exactly one terminal live event after each start, including provider/protocol failures and interruption; argument failures before start emit no lifecycle, and an offline viewer remains best effort. - Squad step limits (#336): exhausted tool loops now exit nonzero instead of falling through as success. The last assistant text is surfaced as bounded, control-stripped INCOMPLETE diagnostics, not a verdict. Limits are validated before paid calls and the last two calls include an honest finish reminder. Deterministic tests cover cap outcomes, final-call completion and tool-message pairing; live researcher/verifier repeats at 3/5 steps completed, without claiming guaranteed model compliance.
- Dev Squad installation (#330):
squad initnow distributes the complete canonical eight-layer tower fromexamples/dev-squad, with nested towers, agents, prompts and source. Removed the obsolete reduced tower template and placeholder substitution. Updates preserve existing towers and bundle edits; bundled example logs remain trackable. The task runner does not launch the full SDK pipeline automatically.
v0.14.0: a dev squad measured on its own runner, the PR steward, two security fences, tutorial-studio
Added
- Squad: a local PR steward (#319).
node tools/squad/prs.mjs [--pr n,m] [--force] [--keep] [--dry-run]reviews and gates the open pull requests: a detached worktree at each head, review and security packed against the PR's own base with the PR text and its linked issue as brief, both passes at once with the fallback rule (an empty or cut answer runs once more, then the security brief goes to the reviewer model), the merge gate, and one summary row per PR with verdicts, findings by severity, gate and cost. A head already reviewed is skipped unless--force; worktrees are removed, on failure too. It never merges, approves, comments or pushes. The squad tower has a PR steward layer. - Dev squad example (#315): simplicity rules and a dependency check that work in every harness (SDK pipeline, OpenRouter loop, interactive Claude Code). The learned patterns now reach the OpenRouter roles (coder, reviewer, security auditor) at the end of their system prompt; the coder gets the project's Prettier + ESLint after every write in the OpenRouter loop, as the PostToolUse hook does in Claude sessions; the reviewer scores
simplicity(Cognitive Complexity ≤ 15, comments consistent with the code, no helper without a caller); the security auditor'sosv_scanasks OSV.dev for the advisories of each new or bumped dependency, with the severity and the fix of the range that holds the installed version; child processes of the OpenRouter loop and of format + lint run without the API keys and tokens. - Example
tutorial-studio(#308): how to turn a web app into a narrated video tutorial, as a tower and as runnable scripts. Brief, outline and writer prompt; narration and browser actions per scene; one voice clip per scene through OpenRouter's speech endpoint (MP3, or PCM for Gemini TTS), a pilot rendered with three voices; a recording of a scratch instance (own port, empty user folder) in headless Chrome on the GPU at 1920x1080; captions, chapters and EBU R128 loudness through Argo; a contact sheet with one frame per scene and a review checklist.npm run dryruns every action with a silent voice, no key needed. The sample is Flow Tower's own tutorial. - Tower card in the tower view (#295). At Tower overview (no layer focused, no node selected) the right panel shows the tower's card, which until now only the library showed: description, tags, layers, nodes, agents, sub-towers, models, runtimes, errors and warnings, live usage. It uses the same facts as the library card. Inside a sub-tower it shows that sub-tower.
–reduces it to its title bar and+brings it back; the choice is remembered. The scene is framed beside the open card, and the live feed sits next to it. - Example tower
agent-council: a multi-expert deliberation council with a coordinator that selects experts and runs three structured rounds of independent evaluation, critique, and final positions. Synthesiser drafts an operational plan; verifier audits it; consensus supports unanimous, majority, or dissent outcomes. Per-role models on OpenRouter with adversarial coverage from a different-provider red team.
Changed
- Dev squad kit (#322): the reviewer and the security pass answer in one request with the code packed by default (
nowandrelaxed;overnightstays a batch;--loopkeeps the tool loop), and only with the JSON verdict the merge gate reads. An answer without one fails the run (no-verdictin the ledger); in the loop it is asked again once. Measured on the squad's own runner (540 runs, results): in the tool loop the cheap reviewers answered prose instead of the verdict on clean diffs, so the gate could never say READY. The reviewer default is nowz-ai/glm-5.3-flash(12/12 at a quarter of FlashX's price). The kit's tower records the measured numbers, and the results page walks a fresh project fromsquad initto a fixed security finding for $0.0121. - Contributing: the working rules live in the repo, so every agent and harness reads them: review before merge and who merges, the issue as the live record, check before proposing, nothing private on GitHub (CLAUDE.md, AGENTS.md); machine-independent gotchas for scratch servers, e2e from a worktree and low memory; the cost skill's tie-break (cheapest per correct answer, then fastest, any provider).
- Contributing and the dev squad: worktrees, scratch servers and their folders are removed after the merge (CLAUDE.md, AGENTS.md, the
flow-tower-squadskill's new step 12,tools/squad/README.md). - Dev squad (#299): every squad prompt (tool loop and one-request) shares one style line (
STYLE_RULESintools/squad/config.mjs), which now asks for the efficient approach when it adds no complexity (batches, no repeated work, independent I/O run concurrently, no premature optimization). Theflow-tower-squadlead starts the review and the security pass together and waits for both. This repo'ssquad.jsongives every model the batched-rendering rule. The kit's squad tower shows both. - Dev squad merge gate (#299):
tools/squad/gate.mjsis a script, with no model, that says READY or NOT READY for a pull request. It checks CI, mergeable state, draft, a clean worktree at the pull request head, both review verdicts, critical or high findings, and that both passes reviewed the head commit.pack.mjswrites that commit only for a clean tree, and the gate reads it only from the run's header. It never merges or approves. Review and security findings now quote their line (pack.mjs). The gate checks each quote against the real file and lists the ones it cannot find, or whose quote is too short to prove anything, as unverified: possible hallucinations. An unverified critical or high finding still blocks. The kit's squad tower adds the pull request, CI and the gate before your merge. - Command-center kit: README § 7 Update covers pulling,
npm cionly when the lock changed, then recreating Vite's cache folders. Without them the unit fails before it starts (status=226/NAMESPACE). It also covers restarting only for server changes and checking/api/version. A note covers Tailscale Serve when a web server already holds port 443. The tower's service node says the same.
Fixed
- Squad: a
--oneshotor--batchanswer cut short (finish_reasonnotstop) passed as a normal run and was recorded asfinish; it now fails withtruncated answerand the ledger outcometruncated-answer(#316). - Dev squad example: the PostToolUse format-lint hook asked ESLint for the
unixformatter, which ESLint 9 removed, so it never reported a lint problem (#315). - Tower card: a long tower name is shown whole, wrapped on more lines, instead of cut with an ellipsis; the reduce toggle stays on its first line (#301).
Security
- Squad (#324):
run.mjsno longer runs a CLI named by a--repoit does not live in. The live events and thevalidateextra tookclifrom the--repo's squad.json, so pointing the runner at someone else's worktree by hand ran code that worktree chose; the PR steward already avoided it. Thecliof any other checkout is now ignored, with a line on stderr, and the runner's own.cli,FLOW_TOWER_CLIor PATH is used. - Dev squad example (#317): the SDK harness's worktree fence was a string prefix test, so a write to
<worktree>/../../xpassed it. It now checks the real path, with symlinks followed (src/hooks/fence.ts), refuses..segments and dangling links, checks every path field and denies a write call with no path (NotebookEdit'snotebook_pathwas never fenced); the OpenRouter loop's write tools use the same test. The children that run the coder's code (tests, checks, the formatter) get an allowlisted environment instead of the parent's minus five keys, so no other exported credential reaches them, and a scratch HOME per worktree, so tools that read~find no credentials; tsc, vitest, git, prettier, eslint and npm were checked to work that way. The real home stays readable by an absolute path: the container is the isolation.
v0.13.0: hub viewers see only the app and the shared files, a voice guided tour, the role catalogue re-checked
Added
- Voice: a guided tour of the tower on screen (#275). "panoramica guidata" / "guided tour" (also "spiegami la torre", "walk me through the tower") explains the tower layer by layer: each layer is focused, an intro says what it does, then one sentence per node while that node lights up. "avanti" / "next", "indietro" / "previous", "dettagli" / "details" (every node of the layer) and "stop" steer it, in the pause between layers or mid-sentence with Interrupt by voice; Esc, another command or leaving the tower end it. The voice agent's model writes the sentences from the tower's own descriptions, in the language of the start phrase, checked against the layer's nodes, asked again once and otherwise replaced by the tower's own text; the next layer is prepared while one plays, and each layer is asked once per tour. About $0.001 per layer (measured once, illustrative). With spoken replies off it runs in the caption. "panoramica" alone still shows the whole tower.
Changed
- Role catalogue: the models new on OpenRouter since 2026-10-16 re-checked on the same task sets (#281, 84 runs, $0.03). The research pick is now
deepseek/deepseek-v4.1-flash-1017(6/6, cheaper and faster than the id it replaces);minimax/minimax-m3.1-flash:freepasses coding, data and orchestrator but stays a note (free tier);perplexity/pplx-decider-flash-v1.2joins the models to avoid for decisions. - The voice-commands example tower maps the guided tour, and its voice guide lists
live_statusamong its 20 tools (#275). - Skills: tower first, then code (#291). The
flow-tower-squadlead updates the tower that models the change (validated, committed) before the first task file, and checks it still matches after verifying; the kit's squad tower draws it.flow-towerand the full procedure end a new tower withflow-tower audit(findings fixed from the code or listed in the report) and offerflow-tower-optimize; a code change the user asks for afterwards starts from the tower, through the squad skill with the dev squad. - Docs: onboard a teammate by sharing only the hub machine (PR #288).
docs/command-center.md§ Connect to it covers sharing one machine through Tailscale instead of the whole tailnet, the email Tailscale accepts, the login inviewers, the identity headers for shared-in users, the slow first page and that files show only withshare --files; the satellite skill now says a teammate signs in with their own account and accepts the share (it said "the same account").
Fixed
- Hub: a viewer opening a shared tower's file that its satellite did not send now reads
this file was not sent by the satellite: share the tower with --files(404) instead offile not shared with viewers, which said nothing about the fix (#289). - Library cards on a hub: the title no longer shares its row with the chips (machine, updated, MINE/EXAMPLE, live dot), which left it a few letters wide and split words mid-word; the chips sit on their own line under it, and a word breaks only when it alone is wider than the card (#279).
Security
- Dependencies: zod 4.6.5 → 4.6.7 for GHSA-mrhx-c7xc-m3w5 (high: validating an oversized constrained collection could exhaust memory), found by
npm auditin the release gates. - Hub: a remote viewer loads only the app's own assets (#287). The guard let a listed viewer's GET through to Vite, which served any file of the flow-tower checkout (
/package.json,/NEXT.md,/src/server/hub.ts?raw,/node_modules/.vite/../../package.json). Nowsrc/server/viewer-assets.tsallows the page, Vite's client, thesrc/appandsrc/coremodules, the pre-bundled dependencies, style sheets and fonts, refuses?raw/?urlreads of a source (also when hidden behind a second?, which Vite still reads as a flag) and any path with a dot segment, backslash or doubled slash, and answers the rest 403.vite.config.tsfs.denyalso listsflow-tower.local.json, the checkout config that can hold a sender token. - Hub: a viewer listed in
viewerscan open the files a shared tower references (#286, PR #285). Before,/api/filewas refused to every remote request, so a teammate on the command center saw the map but not the prompts and scripts, even aftershare --files. A remote read is allowed only for towers a satellite shared (real root under the library'stowers/, a symlink out refused) and only for the files such a tower points to, never a secret; the hub's own towers and other projects it serves stay unreadable remotely; requests on the hub machine itself are unchanged; voice and/@fsstay refused.
Release gates: examples/release-auditor/reports/v0.13.0.md
v0.12.0: an audit that reads stated facts, models per node, the dev squad kit, SQLite in the viewer, share from the interface
[0.12.0] - 2026-10-21
Added
- Ops fields
effect(read,reversible,irreversible: what running a step changes) andmodel_varies(why an agent pins no model: picked at run time by design), and adescriptiononsandbox(what enforces it). The audit reads them (#249). - Share and unshare a tower with the hub from the interface (#266). On a satellite, each local library card has a Share with hub menu (not shared, towers only, towers and files); Settings › Sharing lists each share's folder, mode and state with Unshare, and Remove from config only when the hub cannot be reached. A local project with recent events that no share lists is named in Settings › Sharing and in the SAT chip's tooltip, since the hub lights only the towers that machine shared. Behind it, the local route
/api/sharing/towers(loopback only) andshare-actions.ts, which theshareCLI now calls too.sharewithout--namekeeps the name a folder was already shared under (its tower may have been renamed since), and refuses a name another folder's share holds, which it used to overwrite on the hub. - File viewer: SQLite databases and
.sqlschemas open as tables (#264). A database (recognised by its header, whatever the extension) is read withnode:sqlitein read-only mode: a list of its tables (with row counts) and views (columns only: a view's query is never run), and the first 100 rows of the selected table, with keys, types and NULLs marked; a.sqlfile lists each table's columns, keys and references. The Tables/Schema switch (V) shows the schema text. Bounded for big files: 200 tables, 100 columns, 20,000 cells and 2,000,000 characters of text, counts capped at 100,000; a WAL database gets no side files. The chief-of-staff example ships a sampleambient.sqlite3with fictitious cards. - Dev squad kit (#265):
flow-tower squad init <project>sets the squad up in any git project (the runner copied intotools/squad/, a detectedsquad.json, a task template and a ready squad tower),flow-tower squad keystores the OpenRouter key in the user folder (squad key statussays whether one is set), and the newflow-tower-squadskill is the lead agent's procedure. The runner reads per-project settings fromtools/squad/squad.json: the stack line of the prompt, the check and test commands, the folders the coder may write, the review folders, the base branch, extra rules and the models; this repo ships its own, so nothing changes here. - Role catalogue (#261):
skills/flow-tower-costs/catalogue.md, the measured model pick and two alternatives from other families for every role a node can have, from a 968-run benchmark of 17 task sets on 2026-10-16 (results inexamples/model-benchmark/results/2026-10-16-roles/).models.mjs --role <id>shortlists from OpenRouter's live list by what the role needs (output modality, tools, structured output, image or audio input),--families <n>keeps n per family ranked by weekly use, and it prints the catalogue's pick first. New role classesdecision(fixed answers with probabilities, OpenRouter's Decisions API) andrerank; atype: decisionnode is now classeddecision. - Model benchmark:
bench.mjscalls every endpoint a role needs (decisions, embeddings, rerank, image input, images, videos, speech, transcriptions), with checkswer,wer-roundtripandvision-judge(a judge from another family than the candidate), and 11 new task sets (orchestrator, data, guard, decision, vision, image-gen, video-gen, tts, stt, embeddings, rerank). - Example
ai-commercialista: an AI accountant for Italian companies that cites every source, checks each citation against the text in force for the tax period, scans 2026 levers per profile (SRL, startup, startup innovativa, photovoltaic, employer, forfettario) and leaves every memo to a professional's signature (#228). It ships four benchmark task sets for its model roles (triage, abuse-of-law label, citation support, answers quoted verbatim from Normattiva text), listed inexamples/model-benchmark/categories.md.
Changed
flow-tower auditreads facts the tower states instead of guessing, after optimizing the example towers turned up false positives (#249). AUD03 asks a gate forexactly_onceandeffect: irreversible;effect: readorreversiblelifts the gate that sensitive data or user credentials would ask, a guard or memory that only holds such data is exempt, the gate is searched upstream (a human or an approval gates, an agent ends the search, a guard upstream is info instead of warning), and a sub-tower opened behind a human gate is gated from its entry on. AUD02 warns on a fan-out wider than 20 with no limits (a target with its own limits counts once), instead of any spawn. AUD04 accepts a single edge with a condition or to a node that branches, and skips towers that run no model;decision.failis asked only of a decision that runs one. AUD07 skips nodes whose status is not active and reached humans and guards. AUD09 acceptsmodel_variesand nested towers. AUD10 and the AUD01/AUD02 limits count the budget and limits of every tower that opens a sub-tower. AUD11 checks how much the sandbox isolates (none: warning; network or filesystem open: info). A return that closes a spawn or call is not a loop for AUD01. Role classes read role words in the id or label before spawn edges, so a reviewer that spawns is a reviewer. On the example towers: 114 findings before, 33 after (the rest are real or kept by design).- Example towers use the model the per-node benchmark picked (
examples/model-benchmark/results/2026-10-16-tower-nodes.md): the 30 switch rows in 18 towers, with each example's agent files, code, README and sample logs kept in step. Planners, writers, scoped coders and the browser-worker pre-submit verifier move from Opus to Sonnet; reformatting and note-merging from Opus to Haiku; summaries, doc writers, frame QA, the social-studio repurposer and scoped coders from Sonnet to Haiku; reviewers and the other verifiers stay on Opus. co-scientist'sgemini-2.0-flash, no longer listed on OpenRouter, becomesgemini-3.8-flash; airline-support'sgpt-4.1-mini, which the benchmark could not run under its no-training data policy, becomesgpt-6-luna(also its triage). The dev-squadverifier.mdagent file now says Opus, as the tower andsrc/config.tsalready did. A new guard taskairline-jailbreak(2026-10-20) covers the jailbreak guardrail. Nodes with their own evals should confirm the switch on them. - Example towers state the facts the audit reads: read-only and reversible tools and agents say so, the benchmark's candidate and verifier say their model varies, single-edge decisions carry their condition (#249).
- Tower optimizer:
check-quotes.mjsfinds a quote in a flow-style node and in the node's agent underagents:(#249). - Dev squad:
pack.mjs --paths a,bpicks the folders the review diff covers (default: squad.jsonreview), so a change underexamples/ordocs/gets a reviewed diff (#249). - Dev squad:
tools/squad/cost.mjssums the current project only (the ledger recordsproject;--repo <path>,--allfor every project; entries from before count for flow-tower), andev.shfinds the flow-tower CLI throughFLOW_TOWER_CLI,tools/squad/.clior PATH and stays silent without one (#265).
Security
- Dev squad:
pack.mjsreads review paths literally (--literal-pathspecs: a glob or a:(magic)prefix no longer widens the diff past the checked path) and leaves out of the diff any changed file that looks like a secret (.env*,*.pem,*.key...), naming it on stderr. Before, a tracked key under a review folder went into the packed task sent to the model provider (#249). - Dev squad: the project's check and test commands no longer receive
OPENROUTER_API_KEY(the coder writes the tests they run), and the runner refuses reads through symlinks that leave the repository (#265).
Fixed
flow-tower audit --jsoninto a pipe no longer stops at 64 KB: it exits only after stdout is flushed. The tower optimizer'ssnapshot.mjsreads it through a pipe, so the "before" snapshot of a big tower failed to parse (#249).- Model benchmark:
judge.mjsrunscheck: commandtasks on Node 24, which dropped--experimental-permission: it uses--permissionwhen the running Node accepts it. Before, every command check failed with "bad option" and was counted as a wrong answer (#256). flow-tower check-exampleno longer reportsft_inside a word (draft_for_professional_review) as a token, nor a/home/or/Users/segment inside a URL path as a home folder; a home path after-(${VAR:-/home/x}),../orfile://is still reported (#229).
v0.11.0: GPU for 3D, eco without a GPU, tooltips and panels that fit, shorter tower descriptions
[0.11.0] - 2026-10-15
Changed
- The 3D figures of
flow-tower export --pdfandflow-tower reportuse the GPU when there is one (#209): one launcher (src/cli/browser.ts) starts Chromium with GPU flags, reads the WebGL renderer, and falls back to software WebGL (SwiftShader) when the renderer is software, the probe times out, or the GPU launch fails; it prints which one it used.FLOW_TOWER_GL=gpu|softwareforces the choice. On a laptop with an Intel Iris Xe the dev-squad level-3 export went from 72 s to 16 s. - e2e:
FLOW_TOWER_E2E_GPU=1runs the Playwright project with the GPU flags, for machines that have a GPU; CI stays in software (#209). validatewarns when a node description is over 400 characters or a layer description over 300: the first sentence says what the node does, details stay in its files. docs/what-goes-where.md and docs/schema.md state the limits; the example towers were shortened to fit (voice-commands, team-hub, command-center, dev-squad; the longest was 1240 characters).
Fixed
flow-tower exportkills its server and removes its temporary user folder even when the browser launch throws (#209).flow-tower serviceon a real systemd desktop (#194, found in #181): asystemctl --user stopno longer leaves the unitfailed(the server's 130/143 exits are declared as success);service uninstallalso clears a failed state, so nonot-found failedunit stays listed;service statuson a failed unit points tojournalctl --user -u flow-tower;service installsays when the library is empty (the service cannot start) and names entries gone from disk.- The unit's
ExecStartnames thenodeon PATH that is the running binary (/usr/bin/node), not the versioned file it resolves to (/usr/bin/node-24), so a Node major upgrade does not break the service; only absolute PATH directories that group and others cannot write are trusted. Theloginctl enable-lingerhint is shown only outside a graphical session (#194). - The 3D figures of the tower document never drew in Google Chrome (the browser,
flow-tower export --pdfwithout Playwright's headless shell, and the e2e suite): Chrome stops animation frames in an off-screen or hidden iframe, so every snapshot timed out and the document never got ready. The snapshot iframe now stays on screen, inert and out of the tab order, under an opaque veil; both are hidden in print (#181). - e2e: the referenced-files check skips an optional file per node, so one node's optional entry no longer hides a file another node requires (#181).
- Library: long project names on the cards are no longer cut to one line with an ellipsis (#211): they wrap, up to 3 lines, and the chips stay on the first line. At 1440 px four example names were cut ("Sales Pipeline: prospect to invoice", "A2A Purchasing Concierge", …); now none is, at any width from 390 to 1920 px.
- The HUD tooltip always fits inside the window: it measures itself and goes below, above or beside its anchor, so a long node description in the right node list is no longer cut off at the bottom or squeezed into a narrow column.
- On a remote view of a hub the MIC is hidden, and V says that voice runs on the machine serving the page (open the tower in your own Flow Tower), instead of "this page cannot reach" the server.
- Hub budget alerts below one cent show their amounts with 4 decimals instead of "$0.00 of $0.00" (#211).
- The hub's test alert is sent with rule
testinstead ofsilent, so a webhook receiver can tell it from a real alert (#211). flow-tower sharenames a new share after the tower'sname:(as a slug) when the folder holds one tower file, so towers kept in folders calledtowerno longer all share astower;--namestill wins and existing shares keep their name (#211).flow-tower token addand the Admin tab say "replace 127.0.0.1" only when nourlis configured: a hub that sets a loopbackurlon purpose (one host) no longer gets the tailnet hint (#211).flow-tower connectsuggestshooks installonly when no harness is wired on this machine; otherwise it names the wired ones and points tohooks status(#211).flow-tower guide(and the skill's procedure.md) names the real checkout command instead of a literalnode <path>/bin/flow-tower.js(#211).- The server no longer prints Node 22's
ExperimentalWarningabout SQLite when it opens the history DB; every other warning still shows (#211). - docs/getting-started.md writes the
copyandlibrarycommands asnode $FT/bin/flow-tower.js …like the rest of the page, so they work withoutnpm link(#211). - Hub › Machines shows one row for a sender id that was revoked and added again, instead of two rows with the same data (#211).
- Without a GPU (software WebGL such as SwiftShader) the 3D view starts in eco quality: post-processing drew a dark rectangle over part of the tower there, and it is the slowest path anyway; a quality set in the URL or in Settings still wins (#211). The automatic pick is not saved as a preference, so a browser that gets a GPU later starts in balanced again.
npm run simulateprints where it sends (the URL and why:--url, the running library, or the default port) and says so instead of a stack trace when nothing answers there (#211).- The white corner marks of the node list, the layer list, the legend, the validation panel and the shortcuts dialog stay on the panel's frame: these panels now scroll an inner element, so the corners no longer scroll away with the rows.
- The command-center kit's systemd unit declares
SuccessExitStatus=130 143, like the user unit ofservice install(#194): without itsystemctl stop flow-towerleft the hubfailed. The example towers that map Flow Tower itself (team-hub, voice-commands, command-center, the e2e-sweep sub-tower) and examples/README.md match this release's code.
v0.10.0: team hub admin, alerts that reach you, one-command hooks for every harness
Added
- Hub › Admin (#93), on the hub machine only: add a sender (token and connect line shown once), pause, resume, revoke by typing its id; viewers; cadence per sender; budgets; alert switches; retention days and delete-before with a preview count. Every change writes a line to
admin-audit.jsonl(0600, never a token); remote requests get 404. The privacy profile wizard waits for #85. - Hub alerts (#94): an always-on machine gone silent, monthly spend at 80% / 100% of a budget (USD, per project or global), an error spike (over 3× the 7-day hourly median, at least 5), a token used from a new host. A
⚠ npill in the top bar and the library header opens Hub › Alerts; ack and snooze (1 h / 1 d) from the hub machine only. Rules switch off per name inconfig.jsonalerts;budgetsare documented as USD, the history's currency. - Schema: a
filesentry can be{ path, optional: true, description }for a file a run writes: while it is missing,validateand the PDF checks stay quiet and the Inspector shows it as "not generated yet". The report-studio example uses it for the tower-documentation PDF (#137). - The satellite relay recovers by itself (#188):
connect, a new token or a role change apply to a running library within 5 s (no restart); a refused or paused token keeps the queue, asks again every 30 s and delivers the queue on the first accepted answer; Send test asks at once. The hub tells a paused machine apart (401withstate: "paused", shown asheld/ "paused by the hub") from a refused token, without telling a wrong secret whether an id exists. The library prints one line per relay state change. flow-tower hooks install(#189): one command wires Claude Code, Codex, Pi and Hermes on this machine (or a project with--project), on the port of the running library; it keeps the user's own hooks, backs up each changed file, and replaces older hand-made Flow Tower entries instead of duplicating them.hooks statusshows where they are wired and flags a stale port or Node path;hooks uninstallremoves only Flow Tower's entries.connectends with the hint. Replaces the 11 hand edits the satellite setup needed.- Shared towers are validated (#190): a tower with an error is not sent and the hub keeps the last valid version (it also refuses a structurally broken upload with 422). The teammate sees
1 tower not senton the SAT chip and each shared tower's state in Settings › Sharing, including a warning when a sync removed nodes that events may still name. The hub card saysupdated <time> ago; files a share does not carry (missing, or outside the share folder) are info on the hub, not errors.share --filescarries the agent folders a tower references (.github/,.claude/agents|skills|commands/,.agents/,.codex/prompts/);shareno longer adds a folder already covered in the library. - Hub alerts reach the lead outside the tab (#191): the hub checks the rules every minute and sends each new alert, and its clearing, to an optional webhook (
notify: ntfy, Slack, Telegram or JSON; set and tested from Hub › Admin › Notifications, url masked, never logged); Hub › Alerts can show browser notifications. A newerrorsrule fires when a machine sends 3 errors in 15 minutes (alertThresholds.errors), so a new team sees its first failures without a week of history. - Harness parity (#196, #5):
tests/harness-parity.test.tsreplays one session through every adapter (Claude Code, Agent SDK, Codex, Pi, Hermes, emit), checks the normalised events and that onematch:rule lands each harness's shell call on the same node; what a harness cannot send is declared with its reason (table in docs/realtime.md). Pi was checked with a real run (extension and--mode json): JSON mode now keeps the session id on every line, drops streaming chunks, shows tool results as text and usage with the model and token detail. The Agent SDK integration also sends SessionStart, SessionEnd and Notification. - Hub › Now answers who, where and what (#192): one row per machine and session with the tower and node it works on, its state, and the last error (node or tool and age, never event text); a click opens that tower on that node. The library header shows
● n working. Consistent signals: the tower file's validation chips readYAML n ERR/WARN, a redLIVE n ERRcounts live errors, a failed tool shows ✕ in the feed,HUB n/mcounts active senders only,token addrefuses an id in use, CLI token changes are written to the admin audit log (via: cli), and the Live feed sits beside the Hub panel instead of under it. - A team without Tailscale has one documented route (#193):
integrations/hub/Caddyfile.hubputs Caddy in front of a full hub (events, heartbeats, OTLP and shared towers; the owner views over SSH), with steps in the kit README. The CLI's hints printnode <checkout>/bin/flow-tower.jswhenflow-toweris not on PATH;token addprints the connect line for the other machine; the hub warns about the kit'sexampleplaceholders left in config.json;copy --listmarks the examples that point at this repo; the library footer says how to add towers in a library or on a hub. flow-tower service install|status|uninstall(#194) keeps the local library running at login (systemd user unit on Linux, launchd agent on macOS), so a teammate's events are not lost when they forget to start it;connectsuggests it when no library runs. On a satellite,emitposts to the running library, falls back to the hub when the library refuses the connection, and posts straight to the hub when no library is recorded: never to a default port (a test event on the test server reached another server that way before the fix).- Second newcomer dry run fixes (#206): Hub › Admin PAUSE and RESUME work (they got 415); on a hub a teammate's events light only that teammate's shared towers and the hub's own, never another teammate's; the main agent is
mainin every harness, so oneagent:mainrule fits all; Hub › Now follows the session (idleafter the agent stops,endedafter the session, then dropped);Hopens the Hub panel on the first load (the library filter no longer takes the focus;/does);service installrefuses a port a library already uses andservice statustells an installed unit that is not running; the hub card saysnewer version not sentwhen a teammate's latest save was refused; the kit lists the full-hub Caddy route in step 4 and says to create a token before a hub without viewers starts.
Changed
- Example towers follow the team-UX work (#187-#196):
team-hubgains the local port file and stream ids,hooks install,service install, the relay supervisor, the share sync and the hub's shared-towers guard, the alert webhook and the Now view (36 nodes, audit clean);command-centershows the four teammate steps for any harness, Caddy for a team without Tailscale (deploy/Caddyfile.hub), the alert webhook and the token approval;dev-squad's coder never rewrites an existing test file, an empty grader answer is never a pass, and the verifier holds a user-facing change no test exercises end to end. - Docs:
docs/architecture.mdis split into four maps that each load whole: core, documents and reports (architecture.md), server and hub, CLI, app. docs-qa Q24 (which file owns the per-sender tokens) passes again, 3/3; the docs-qa report is refreshed: 28/31 (#143).
Fixed
- Hub alerts could not see hook errors (#187): history dropped every
tool.endas a replay of itstool.start(same call id), andkind: errorevents were stored without an error status, so the error spike never fired. Both now count. - Hub › Admin showed
always-onfor a sender with no cadence, while the alerts treat it ason-demand; it now shows what applies, and a new sender starts ason-demand(#187). emitandnpm run simulatepost to the server the user folder runs, on whatever port, instead of 5317 when no--urlis given: a satellite's SessionStart hook could reach another server on 5317 (#187). Thecopyhint no longer prints a 5317 URL.- Two e2e tests were wrong: the
/api/filecheck asked foroptionalfiles not generated yet (#137), and the voice hub-phrase test faked a hub without answering the alert pill route added by #94.
Release gates: examples/release-auditor/reports/v0.10.0.md