Repository navigation
v0.14.0: a dev squad measured on its own runner, the PR steward, two security fences, tutorial-studio
·
409 commits
to main
since this release
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.