Changelog
Everything shipped to Agentlas, newest first.
Updates stop re-downloading the whole app
Differential updates — downloading just the changed bytes instead of the full ~340MB app — had never actually engaged on macOS, for every user, on every release. The cause was ours.
- electron-updater computes a differential download against a fixed baseline file at the cache root. Our own stale-artifact cleanup was deleting that baseline file along with old install payloads — including on the success path, so every completed update destroyed the exact file the next update needed to stay small.
- The log confirmed this three releases in a row before the fix.
- The cleanup sweep still discards everything it can't trust; it just no longer discards the baseline an accepted install proved it needed.
Desktop updates recover instead of stranding you
Desktop's updater had accumulated more safety gates than it needed, and a few of them were actively blocking healthy updates instead of protecting anyone. We removed the gates that were causing false failures while keeping the ones that matter.
- A pre-install recovery copy — meant as a convenience, not a requirement — was treated as a precondition for updating, so a transient disk error could cancel an update four times in a row on a real machine.
- Startup used to skip re-arming the update timer whenever a hold was active, and nothing else re-armed it, so one failed install meant no further updates were ever checked again.
- Holds now expire and the next check resumes from a clean state; the recovery copy, writer pause, and install journal are all best-effort instead of blocking.
- Shipped in Desktop v0.9.38.
Network staffing stops excluding good candidates over fields nobody fills in
A work order that asked for a specific authority or role could silently wipe out an entire candidate pool — not because no one qualified, but because almost no published agent card fills in fields the live catalog barely uses.
- Measured live: requiring a single authority field staffed an engineering slot with two unrelated auditors and excluded every actually relevant engineer. Lifting that one hard requirement took the candidate pool from 2 to 6, with every excluded candidate turning out to be innocent.
- Requirement dimensions that almost nothing in the catalog declares — like authority or role — now demote to a ranking signal instead of a hard filter, and report which requirement was under-supported instead of just returning nothing.
- Real enforcement of authority still happens at prepare time, through permission policy — this change only affects search-time filtering, not what an agent is actually allowed to do once selected.
- Authoring a work order slot no longer requires spelling out every optional field as an empty array; an absent list field is now treated as the empty constraint everywhere, including in the digest used to verify the order wasn't tampered with.
Publishing a package no longer gets blocked by ordinary words
Our credential scanner was matching a bare substring anywhere in a filename — so any file with "token" or "secret" in its name got flagged as a leaked credential, real or not. A flagship paid package got blocked for shipping its own file named `token-architecture.md`.
- Detection now matches a path segment, not a filename substring, so ordinary product vocabulary stops tripping the scanner.
- This costs no real coverage: everything that reaches the scan is already text content, and it's still matched against the actual secret-value patterns — the check that looks at what a file contains, not just what it's called.
- A related gap is closed too: the old pattern required a parent folder, so a credential store placed at a package's root previously matched nothing at all. Both nested and root-level stores are covered now.
Non-English requests now route to the right agent
Hub and Network routing were quietly worse for Korean (and other non-English) requests than for English ones — an identical query in Korean could rank the correct agent 144th instead of 1st. The fix wasn't a bigger model; it was making sure the request gets written in the language the candidate corpus actually speaks.
- `/hep-network` now has the connected model author the work order's searchable fields in English — faithfully translating a non-English request rather than passing the original wording through — while `languages` still tracks the language you actually want the work delivered in.
- Card matching moved off a token-hashing scorer that scored an equivalent Korean/English request pair at 0.0 (not semantically similar at all) onto the same verified sentence model already used elsewhere, which scores that same pair at 0.383.
- Measured improvement: 18 of 18 Korean test queries with a known right answer improved after the fix, with no regressions on queries that already worked.
Hephaestus keeps itself updated even inside sandboxed hosts
Claude Code, Codex, Cursor, and Antigravity all run tool commands inside a sandbox that can't write to your home directory. That meant a machine that only ever used Hephaestus through one of these hosts could never actually install a runtime update, even though the plugin itself kept updating fine through the host's own marketplace.
- The update worker now starts from the host's SessionStart hook, which runs outside the tool sandbox, instead of relying only on an in-command trigger that couldn't write the files it needed.
- Verified end to end on a completely empty machine: one session start bootstrapped a fully verified runtime with no terminal and no manual step.
- A denied cache write during an update check no longer kills the whole check — it's treated as an optimization, not a requirement, and now reports the real cause (a sandbox permission boundary) instead of blaming your network or disk.
Agentlas Terminal is a ground-up rewrite, now on npm
The independent Agentlas Terminal has been rebuilt from the ground up and published to npm as `agentlas`. It now ships 56 commands as a single, coherent CLI instead of the previous generation's patchwork.
- Install with `npm install -g agentlas` — cross-platform, no separate runtime download.
- Full smoke coverage across all 56 commands passed before publish (54/54, two intentionally skipped edge cases).
- The Terminal keeps the `agentlas` command name for itself; the underlying Hephaestus engine no longer creates a conflicting alias.
- This is a separate product surface from the Hephaestus plugin used inside Claude Code, Codex, and other hosts — same engine underneath, its own release cadence.
New: Context Map keeps AI edits scoped to what actually depends on them
Hephaestus now maintains a dependency-aware map of your project instead of treating every task as a fresh, context-free request. `hephaestus context` exposes it directly.
- `context refresh` builds a canonical map of definitions, backlinks, and module dependencies across your project, kept fresh by a source fingerprint so you don't need to remember to re-run it.
- `context slice --task "..." --target <file>` returns a bounded, prioritized set of the files, definitions, and interfaces that specific task actually depends on — not a full project dump.
- `context impact --changed <file>` shows what a change actually touches downstream; `context verify` requires those impacted files to be reviewed before a task can report complete.
- Claude Code and Codex hooks now warn about reverse references immediately before an edit, and the same local implementation backs Desktop, Terminal, Stormbreaker, and Workforce — so the answer doesn't depend on which surface you're using.
- Network and Hub discovery never receive your project's source or local map paths; this stays entirely local.
Hub runtime bundles now repair legacy manifests safely
Fixed legacy Hub agent packages whose runtime manifest was missing fields such as `packageHash` or `createdBy`, so published callable agents can return runtime bundles again.
- Backfilled production Hub package records: 121 checked, 25 repaired, 0 remaining invalid.
- Added runtime compatibility normalization for older `agentlas.json` schemas and packages missing `agentlas.json`.
- Added audit and smoke scripts so future Hephaestus/runtime upgrades can catch the same failure before release.
Desktop now reads the live Hub catalog from Web
Agentlas Web now exposes a public full-list Hub catalog endpoint for trusted clients such as Agentlas Desktop and Hephaestus. Desktop can stop showing local hardcoded fallback agents and render the same published Hub entries that Web serves.
- Added `/api/marketplace/agents` with a short public cache and safe empty-list fallback.
- Reused the existing published-profile loader so Web remains the catalog source of truth.
- Documented the API in the repo README and removed unused draft screenshot assets from the source tree.
Hub search now stays online during Mongo slowdowns
Agentlas Hub search no longer depends on full marketplace profile reads during the request path. Railway now loads a precomputed `marketplace_search_index.json` from Cloudflare R2 into memory, while Mongo stays on Flex as the source of truth for publish and backfill work.
- Built and uploaded a 115-entry Hub search index with OpenAI routing embeddings on every entry.
- Production reads the R2 public index URL with a short in-memory TTL before falling back to slower stores.
- Hybrid ranking remains intact: keyword, BM25-style token evidence, CJK tokens, routing text, embeddings, trust/eval signals, and diversity gates.
- Z.ai reranking is enabled only for ambiguous cases and is capped at 1 second; timeouts fall back to deterministic ranking.
- Hephaestus Network, MCP `marketplace.search_agents`, and Desktop Hub/Cloud clients benefit through the live Hub API immediately. Desktop app binary updates are separate and still use the GitHub Releases auto-updater.
Agent Cloud now remembers your saved Network agents
Hephaestus Network now has a real workspace-scoped control plane for frequently used agents.
- `/hep-network` searches signed-in Agent Cloud first, then saved Hub bookmarks, then the public Hub. `/hep-cloud` stays Cloud-only.
- Agent Cloud has separate sections for your private Cloud packages and Hub agents you bookmarked for reuse.
- Runtime bundle calls can include a workspace personalization overlay: promoted memory summaries, promoted playbooks, plugin locks, and a retrieval receipt id.
- New MCP tools let callers save candidate memory, candidate playbooks, and self-evolution proposals without mutating the public Hub package.
- The storage model is account-scoped and indexed for collection-backed MongoDB: agent bindings, overlays, memory items, playbook cards, plugin locks, retrieval receipts, run events, and evolution proposals.
- Raw prompts, transcripts, secrets, credential values, and private local files are rejected from durable personalization records.
Oberon is now open source
Oberon — our long-form AI film studio — is now open source. The engine that turns a one-line brief into a full multi-shot production is on GitHub for anyone to read, run, and build on: github.com/agentlas-ai/oberon.
- Plan a whole production from a single brief — shot list, continuity, second-by-second camera choreography, dialogue, and typography — with no API key and fully reproducible results.
- Every shot's video model is chosen by a scored router across seven factors (fit, quality, control, reliability, cost, speed, consistency), and it shows its work so you can see exactly why each model was picked.
- Titles, name captions, and subtitles are composited by code, so on-screen text is always crisp and never garbled by a generation model — Korean and other scripts render perfectly.
- Run the whole pipeline from your terminal: `oberon plan`, `route`, `export`, `titles`.
- The Oberon card on the Studio page now links straight to the repository, next to the desktop app.
Public agent pages now read package metadata before falling back to generic copy
Public agent detail pages now recover better descriptions from the package itself when older uploads do not have a full public profile. The page reads the routing card and human README before using a generic fallback, so published packages show what they do, what they need, and what they produce more accurately.
- Read package routing cards for cleaner titles, summaries, inputs, and approval notes.
- Read the human README for concrete output lists when available.
- Avoided generated skill-template descriptions as a public copy source.
- Removed duplicate bottom CTAs on shorter detail pages so the page ends cleaner.
Cloud upload now has stronger package checks and richer Hub detail pages
Cloud agent upload is stricter before registration, and Hub detail pages can use more of the package's own public metadata. The result is less generic marketplace copy and fewer packages that look installable before their structure is ready.
- Added checks for owner bindings and package architecture before upload paths proceed.
- Extended the cloud registration path to carry more public package information into Hub display.
- Improved plugin and package pages so users see clearer descriptions, setup notes, and security context.
- Added scanner coverage for risky templates and leaked values before packages are presented publicly.
Hub pages now separate team credits, package labels, and credential safety
Hub surfaces now explain paid team calls and downloadable packages more plainly, and plugin manifests have an explicit credential-request policy contract. That makes it clearer what is free to browse, what consumes workspace credit, and what kind of credential access a package is allowed to ask for.
- Clarified Hub team credit wording on public pages and MCP responses.
- Cleaned up package labels so teams, agents, plugins, and downloads are not blurred together.
- Added a credential policy contract for plugin manifests.
- Extended architecture checks so packages that ask for credentials are reviewed against the declared policy.
Startup Studio can run published HQ stages from the web app
Startup Studio now has a hosted run path for its core stages. A signed-in workspace can trigger the published HQ behind each stage, while the page keeps a calmer motion profile for users who prefer reduced motion.
- Added the hosted run endpoint for Idea, Market, Business Plan, PRD, Build, and Deck stages.
- Kept stage execution behind the same sign-in and credit checks used by MCP calls.
- Added a reduced-motion-safe hero loop on the Studio landing page.
- Made the Hephaestus install prompt versionless so the public install copy does not go stale after every patch release.
Prompt Dictionary loads faster and filters junk submissions
Prompt Dictionary got two practical fixes: faster catalog reads on MongoDB-backed installs, and a submission gate that rejects obvious junk, spam, duplicate, or unsafe prompt posts before they reach the paid catalog.
- Reworked prompt seed and store reads so the catalog does not stall on large Mongo-backed data.
- Added deterministic checks for short, repeated, link-heavy, contact-funnel, and duplicate submissions.
- Added an optional LLM review step for borderline prompt quality and safety when a provider is configured.
- Upgraded the beauty and wedding prompt seeds so the starting catalog reads more like consultant-grade material.
Hephaestus search and call are clearer from the first page
The public Hephaestus surfaces now explain the core command family more directly: build an agent, search the network, call an agent, and use cloud-backed help when needed. Search results also carry clearer candidate descriptions and a fallback for broad intent.
- Updated the landing copy so the visible command examples match the current product shape.
- Improved Hephaestus Network search responses with better candidate descriptions.
- Added an intent fallback so broad search requests do not fail silently.
- Kept the install guide aligned with the same command family shown on the product pages.
Hephaestus Network now has a Claude connector path
Agentlas now has a dedicated Hephaestus Network connector surface for Claude-style MCP clients. The connector exposes the Hub search and call flow through a reviewable public endpoint, with privacy copy and verification checks kept next to the implementation.
- Added a Hephaestus Network connector page and privacy page.
- Added the MCP endpoint used by the connector review flow.
- Added verification scripts and review evidence docs so connector status is checked instead of guessed.
- Fixed MCP tool names so Claude-compatible clients can list and call them without dotted-name issues.
Hephaestus Stormbreaker: an operational control layer for coding agents
Hephaestus Stormbreaker is now documented as the control layer around serious coding-agent work. It is not positioned as a new model or a benchmark leaderboard feature; it is the execution protocol that keeps an agent run scoped, auditable, and tied to real verification evidence.
- Added `scope lock` so substantial work starts with an explicit boundary for what the agent is allowed to change.
- Added failure-memory checks so repeated mistakes and dead-end approaches can be carried into the next repair loop instead of being rediscovered.
- Added verifier-first planning: the agent names the tests, build checks, browser checks, or review gates before it edits.
- Added an evidence loop that separates public progress from final proof, making it harder to stop after a plausible-looking partial fix.
- Added final-gate language for Hephaestus surfaces: work is only done when the implementation, verification result, and remaining risk are all visible.
Hephaestus v0.4.9 and routing-indexed publishing are live
Hephaestus is now released as v0.4.9 and the live Agentlas install guide points at the new one-touch installer URL. The command shape did not change; only the version pin moved forward.
- Agentlas Web now writes routing indexes from draft publish, legacy publish, and marketplace import paths, so new public agents enter search with normalized description, tags, capabilities, examples, and optional embedding fields.
- Desktop and terminal cloud publishing now require a valid `.agentlas/routing-card.json` before registration and send that routing card to the cloud package manifest.
- Agentlas Desktop v0.2.29 is published with the cloud-routing release and Oberon product work; the live desktop update endpoint reports ready and notarized artifacts.
- Hephaestus v0.4.9 includes compact Hub routing, local-inventory-aware reranking, runtime update checks, and Windows shims.
Hub routing is now compact, ranked, and ready for semantic scale
Agentlas Hub search no longer behaves like a raw candidate dump. Search now starts with compact lexical routing, uses semantic embeddings only when the match is weak, exposes trust signals, clusters near-duplicate agents, and asks a clarify question instead of returning noisy results for broad requests.
- Published agents can store a routing index with normalized trigger/capability text and an optional OpenAI embedding created at registration or backfill time.
- Search prefers small top-K results, rank fusion, trust telemetry, and cluster summaries over sending long manifest fields to the model.
- Hephaestus reranks Hub results with local inventory context without sending local agent names, paths, memory, or private card text to the server.
- The scale plan is explicit: move from in-process profile scans to a dedicated `routing_documents` collection, BM25/vector candidate generation, and second-stage reranking before the catalog reaches hundreds of thousands of agents.
Plugin install panel now tells the truth
The plugin detail page used to show a desktop deep-link button and a terminal command that did not work yet. Both are gone.
- The Desktop tab now leads with the steps that actually work: open Agentlas Desktop, search the plugin, click Install.
- The Terminal tab honestly says the CLI is coming soon instead of showing a command that fails.
- Desktop is now the default tab.
This changelog is live
From now on, every shipped change lands here as a short blog-style note — what changed and why it matters, in plain words.
- Entries are published straight to the database, so updates appear without a redeploy.
- Available in English and Korean, following your language setting.
- Find it any time from the footer under Resources.
One-minute install: a single command, a picture guide, and the classic pages back
Getting Agentlas into your AI tools used to mean per-client config files. Now there is ONE way in: a single one-touch command that installs Hephaestus and wires the Agentlas Hub MCP into Claude Code, Codex, Gemini, and Antigravity at once.
- Agent Hub and the MCP page now open a picture-first guide — four mock screens (Spotlight, terminal, checkmarks, chat) with one sentence each. If you can turn on a computer, you can follow it.
- The Hephaestus page and the MCP page share the same command; only the popup explanation differs (building agents vs calling them).
- Manual per-client setup (Claude Code, Codex, Cursor, Windsurf, VS Code, Antigravity) moved under an Advanced toggle on `/mcp`.
- The illustrated Hephaestus page and the scroll-story memz page are back: `/products/hephaestus` and `/products/memz` now route to the original designs.
- Home hero got cleaner and wider; the Product menu in the header no longer disappears while you move your mouse down to it.
- New: a paste-into-your-AI prompt. Drop one sentence into Claude Code or Codex and the AI runs the install itself — and fixes any shell errors on its own.