Skip to content

v0.12.0: an audit that reads stated facts, models per node, the dev squad kit, SQLite in the viewer, share from the interface

Choose a tag to compare

@JamesMoriartyDecripto JamesMoriartyDecripto released this 21 Oct 10:45
· 665 commits to main since this release
30d03d0

[0.12.0] - 2026-10-21

Added

  • Ops fields effect (read, reversible, irreversible: what running a step changes) and model_varies (why an agent pins no model: picked at run time by design), and a description on sandbox (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) and share-actions.ts, which the share CLI now calls too. share without --name keeps 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 .sql schemas open as tables (#264). A database (recognised by its header, whatever the extension) is read with node:sqlite in 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 .sql file 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 sample ambient.sqlite3 with fictitious cards.
  • Dev squad kit (#265): flow-tower squad init <project> sets the squad up in any git project (the runner copied into tools/squad/, a detected squad.json, a task template and a ready squad tower), flow-tower squad key stores the OpenRouter key in the user folder (squad key status says whether one is set), and the new flow-tower-squad skill is the lead agent's procedure. The runner reads per-project settings from tools/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 in examples/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 classes decision (fixed answers with probabilities, OpenRouter's Decisions API) and rerank; a type: decision node is now classed decision.
  • Model benchmark: bench.mjs calls every endpoint a role needs (decisions, embeddings, rerank, image input, images, videos, speech, transcriptions), with checks wer, wer-roundtrip and vision-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 in examples/model-benchmark/categories.md.

Changed

  • flow-tower audit reads facts the tower states instead of guessing, after optimizing the example towers turned up false positives (#249). AUD03 asks a gate for exactly_once and effect: irreversible; effect: read or reversible lifts 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.fail is asked only of a decision that runs one. AUD07 skips nodes whose status is not active and reached humans and guards. AUD09 accepts model_varies and 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's gemini-2.0-flash, no longer listed on OpenRouter, becomes gemini-3.8-flash; airline-support's gpt-4.1-mini, which the benchmark could not run under its no-training data policy, becomes gpt-6-luna (also its triage). The dev-squad verifier.md agent file now says Opus, as the tower and src/config.ts already did. A new guard task airline-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.mjs finds a quote in a flow-style node and in the node's agent under agents: (#249).
  • Dev squad: pack.mjs --paths a,b picks the folders the review diff covers (default: squad.json review), so a change under examples/ or docs/ gets a reviewed diff (#249).
  • Dev squad: tools/squad/cost.mjs sums the current project only (the ledger records project; --repo <path>, --all for every project; entries from before count for flow-tower), and ev.sh finds the flow-tower CLI through FLOW_TOWER_CLI, tools/squad/.cli or PATH and stays silent without one (#265).

Security

  • Dev squad: pack.mjs reads 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 --json into a pipe no longer stops at 64 KB: it exits only after stdout is flushed. The tower optimizer's snapshot.mjs reads it through a pipe, so the "before" snapshot of a big tower failed to parse (#249).
  • Model benchmark: judge.mjs runs check: command tasks on Node 24, which dropped --experimental-permission: it uses --permission when the running Node accepts it. Before, every command check failed with "bad option" and was counted as a wrong answer (#256).
  • flow-tower check-example no longer reports ft_ 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}), ../ or file:// is still reported (#229).