What's new in Codron

EARLIER · page 2 of 7

1. v0.15.0Jul 24, 2026Collaborators: invite by email, engage from the web

   • · Invite a collaborator per project from Settings, COLLABORATORS. Type their email and they get an accept banner on codron.app/projects the next time they sign in (no link handover needed); leave it empty to mint an anonymous single-use link instead. Invites expire in 7 days and are revocable, and email invites only work for the account that owns that address.
   • · A collaborator gets a remote control, not the keys: they see the shared project on their codron.app/projects list (with a shared badge and your machine's online dot), queue orders, and engage or pause. Cycles run on your machine and your account, and everything they press is stamped with who pressed it. Remove them in Settings and access ends immediately.
   • · Pending orders now batch by where they came from. The next engagement consumes only the oldest run of same-source orders (this desk, the website, telegram, a collaborator); the rest queue behind and each batch gets its own cycle, in arrival order. The PENDING ORDERS panel groups the queue with a 'next' marker so you can see exactly what the next cycle will pick up.
   • · Orders from a collaborator show up in PENDING ORDERS under their own header (web with their name), so you always know whose instruction the agent is about to run.
2. v0.14.3Jul 21, 2026Codex chat picks up the promote/order call-to-action, and both engines close decision-resolving replies with a settled/open scaffold

   • · In a scratch or session chat driven by Codex (GPT), the reply now surfaces the promote/order call-to-action button when the concept is ready to build. Claude chats already did this, Codex was silently dropping the marker, so a converged conversation kept looping instead of nudging you to promote (scratch) or draft an order (session). Codex chats also stop opening with 'let me check…' before a read, and stick to a two-sentence answer when that is the whole answer, matching Claude's reply style.
   • · Both engines now close decision-resolving chat replies with a one-line 'settled: X, Y · open: Z' status in your language, a mini summary of what the conversation has locked in vs. what is still open. It feeds directly into the briefing (scratch promote) or the order draft (session draft), so the conversion has real scaffolding instead of reconstructing decisions from raw prose. If the thread stalls without progress, the assistant now names the blocking tradeoff and asks you to pick a side, instead of endlessly exploring adjacent options.
3. v0.14.2Jul 21, 2026GPT-5.6 Sol + max/ultra tiers, stop/redirect on decision cards, dev-server clears on restart

   • · The GPT model picker matches OpenAI's current Codex lineup again: GPT-5.6 Sol (the top frontier model) is back and is now the default for GPT engagements, the retired 5.4-mini is dropped, and the reasoning picker gains the new max and ultra tiers where the model supports them (Sol and Terra go up to ultra, Luna to max, 5.5 to extra high). Verified live against the Codex CLI's own model catalog.
   • · In a project session, the chat now nudges you to draft an order for that project (the composer's draft button) instead of saying 'promote'. Promotion keeps its meaning elsewhere: turning a pre-project scratch into a new project. The call-to-action badge reads ORDER in a session, PROMOTE in a scratch.
   • · When an engagement is waiting on you (an env var, a file, or a yes/no), its card now has a Stop button that drops the question and parks the project clean, plus a redirect box where you can just tell it to do something else instead, which closes the decision and re-runs the project on your new instruction right away.
   • · The Overview's Decisions list gets a per-row dismiss: close out a pending decision for a project you've stopped working on so it stops showing in the list, the needs-input badge, and the counts. It runs nothing, it just clears the nag.
   • · Fixed a dev server that stayed stuck on its last error after you fully quit and reopened Codron. A restart now clears an errored dev server back to a clean stopped state instead of surfacing a failure from the previous session.
4. v0.14.1Jul 17, 2026Connect MCP servers by URL and Codron runs the OAuth for you, keeping the token fresh; MCP now lives in Settings, Integrations and applies to every project

   • · Add a remote MCP server by pasting its URL and pressing Connect. If the server uses OAuth, it opens in your system browser once, you approve, and Codron stores the token in your OS keychain and refreshes it in the background, so the connection keeps working across engagements without you handling a token by hand. Servers that need no OAuth register straight from the URL. A Token method (paste a static bearer) and a JSON method (for stdio/command servers) are still there; you pick the method up front so the add form shows only what that method needs.
   • · MCP servers moved out of their own tab into Settings, Integrations (below the CLI tools), and now apply to every project instead of being turned on per project. The per-project MCP section is gone: whatever you connect once is available to all projects' engagements, and refreshing a token updates everywhere at once.
   • · The ngrok authtoken field now sits directly under the ngrok row in Integrations instead of floating in a separate card, so the token lives with the tool it configures.
   • · Fixed a Settings page rendering error in the desktop app, the Workspace section could mismatch on a fresh page load and blank the section.
5. v0.14.0Jul 16, 2026Pick Claude or GPT per session and scratch chat, see what the model is doing while it thinks, and every reply streams as one smooth flow

   • · Session and scratch chat can now run on GPT (the OpenAI Codex CLI) instead of Claude. A model picker sits left of the draft/promote button in the composer; pick once per thread and it locks after the first message, so the whole conversation stays on one engine. The choice governs only the thinking chat: order drafts, promote briefings, and engagements keep the project's default engine. Every listed model is verified runnable (opus-4.8, fable-5, opus-4.7, sonnet-5 / 5.6-terra, 5.6-luna, 5.5, 5.4-mini).
   • · A GPT scratch runs on the machine the scratch is bound to, so that machine needs the codex CLI installed and signed in. A missing codex, or one too old for the gpt-5.6 family (needs 0.144+), now fails with a message that says exactly that and how to fix it, instead of a raw spawn error. Projects that run codex with no engine config move from gpt-5.5 to gpt-5.6-terra as the default.
   • · While the model works, the chat now shows what it's doing: read file X, search Y, the shell command codex just ran. This activity transcript used to exist only for Claude replies inside a project session; now GPT replies and all scratch replies (both engines) paint the same rows instead of a bare 'thinking…' until the answer lands.
   • · Replies stream as one continuous flow everywhere. Claude session replies used to paint in visible chunks, codex delivered its whole answer at the end in one drop, and the order-draft preview chunked too. A backlog-aware typewriter now paces all of them: codex prose appears block by block during the turn, long bursts accelerate instead of lagging, and nothing snaps in at the finish.
   • · The model picker popup uses the same dark panel design as the @-mention list (mono CLAUDE/GPT tag plus model name per row) instead of the operating system's dropdown, and model names read as hyphenated slugs (opus-4.8, 5.6-terra).
   • · Starting a dev server on one project now automatically stops another project's running dev servers first, one project's app at a time, so ports and the preview don't collide. The stopped project's log says which project displaced it.
6. v0.13.8Jul 13, 2026@-mentions now work in the pending-orders box, reach your on-disk reference files (SVGs, PDFs, images), and color inline as you type; plus the project list opens on the machine you're actually on

   • · The pending-orders composer (the right-panel box where you queue an engagement) now supports @-mentions, which previously only worked inside a project session. The @ picker is the same in both places. It now also lists files on disk under .codron/references/ (logos, SVGs, PDFs, screenshots), not just the text docs that sync to the cloud, so a binary reference like a design SVG is mentionable where before it silently never showed up. And a picked mention renders inline in the input, colored like a live token as you type, instead of sitting as a separate chip above the box; on send it's still attached as a referenced doc for the agent to read.
   • · The dashboard project list opens filtered to the machine you're on, instead of showing every machine's projects and making you click your own machine's chip on every visit. It resolves the current machine, caches it, and snaps the filter there before the list paints, so there's no flash of the full list. Your own chip clicks still win for the rest of the session, and it falls back to All in a plain browser tab or on a machine that hosts no projects.
   • · Session and scratch answers that run a tool mid-reply (reading a file, searching the code) no longer glue the sentence before the tool and the prose after it into one run. The model streams text, pauses to run a tool, then continues, and the two stretches now break into separate paragraphs so the answer reads with proper spacing instead of the pre-tool and post-tool sentences jammed together with no gap.
   • · Composer polish: the right-panel order box and the session chat box now open at the same height (they were different sizes before), and the 'Drafting your order' banner in the session's draft-to-order modal uses the same rounded card as the 'Drafting your briefing' banner shown when you promote a scratch, so the AI-is-drafting wait looks identical wherever it happens.
7. v0.13.7Jul 11, 2026Failed dev-server and app builds now auto-recover and say so, and hitting a token limit lets you switch engines and resume instead of just waiting

   • · Pressing Run on a mobile app (the Device tab) now recovers from a broken build the same way web dev servers already did. When the build+launch fails (a missing pod, an out-of-date native project, a stale Flutter cache), a short one-shot session diagnoses it, applies the minimal fix, and retries the run, up to twice, before giving up. Before this, an app build error just went red and you had to read the log and fix it by hand; only web (frontend/backend) had the auto-recovery.
   • · The auto-recovery is now visible instead of pretending to be a slow boot. Both web and app show a 'recovering' status while it's fixing, and the log gets a clear line ('start/build failed, auto-fixing (attempt 1/2)…', then a retry line, then a give-up line if it can't). On web this was previously masked as 'starting', and on mobile it was invisible, which was confusing since an app build and a recovery both take minutes and looked identical.
   • · When a project hits a token/quota limit it parks on a cooldown; you can now switch its engine (say from one provider to another) and press re-engage to resume right away, instead of the button being disabled until you queue a new order. The cooldown clears on re-engage and the next run uses whichever engine you just selected. (The cooldown still doesn't auto-clear on its own when you switch, this is the manual switch-and-resume path.)
8. v0.13.6Jul 11, 2026Filter the feed down to failures while debugging, see which mobile app (iOS/Android) is running right from the overview, and cleaner, aligned panel headers

   • · The cycle feed gets a clean / issues filter so a debugging streak of failed engagements stops burying the real-development ones. 'issues' means failures only (a non-zero exit or a classified failure like token-limit); resolved and pending decisions, verify, and halt outcomes stay out of it, so parking a decision no longer reads as an error. Failed cards also carry a red tag with the failure kind, so errors are scannable at a glance instead of blending into the list.
   • · The overview now shows a running mobile app the same way it shows web dev-server ports. Web projects already surfaced a port chip on their row; app projects showed nothing, so you couldn't tell from the overview whether an iOS Simulator or Android emulator build was live. Running an app from the Device tab now mirrors its status to the cloud, so the project row shows a phone chip with the platform (ios / android), across machines, and clears when the run stops.
   • · Reference files that preview inside Codron (images, markdown, text) no longer show an 'open in default app' shortcut, so clicking a .md reference opens the in-app viewer instead of whatever the OS associates with markdown. The cooldown badge on a project row shrinks from a wrapped three-line box to a single compact chip. The lone 'new session' button in an empty split pane is no longer a giant full-width box; it sizes to its label.
   • · The three project panels (left docs, center tabs, right activity) now share their header divider lines: the right panel's status header takes the center tab-bar height, and the left panel's stats and tab rows line up with the center tabs and the feed filter, so content starts on the same line across all three.
9. v0.13.5Jul 9, 2026Chat scroll yields when you're mid-reading a stream, and re-latches only when you scroll back to the bottom

   • · Session and scratch transcripts used to slam the scroll back to the bottom on every streaming chunk, so scrolling up mid-reply to re-read something the assistant said 30 seconds ago got yanked down again every ~50ms. Now the auto-follow only engages while you're within ~100px of the bottom; the moment you scroll up it disengages and leaves you alone. Scrolling back down to the bottom re-latches, so the 'take me to the rest of the reply' recovery is the same gesture you'd use in a terminal tail. Typing + Enter (your own send) still jumps to the bottom unconditionally, since that action carries the intent 'take me to what I just sent'.
10. v0.13.4Jul 9, 2026Rename and delete recent sessions inline, edit Codron's default project location from Settings, and a sharper app icon in Windows shell surfaces

    • · Settings gets a Workspace tab so you can move Codron's default project location without hand-editing config.json. Onboarding was the only place to set this before, so a reinstall or a config reset (yesterday's crash-mid-update, for one) left you either living with `C:\Users\...\Codron\projects` or opening the file manually. The tab reuses the same folder picker the wizard uses, and the change only affects where NEW projects scaffold (existing project roots keep their absolute paths; the section says so inline so nobody expects a bulk move).
    • · Recent-session rows and the scratches list gain inline pencil-to-rename and trash-to-delete on hover. Before this the recent-sessions list only let you reopen: a session titled from its first message (auto-derived, sometimes verbose) could only be fixed by opening and closing the tab, and even then the rename didn't survive a reload. The auto title is now a placeholder you can overwrite; the change persists (localStorage cache wins over the disk-derived title when both name the same closed session), and the trash actually wipes `.codron/sessions/<id>.json` plus its cache after a confirm. The tab strip's history popover and the empty-pane recent list now share the same row component so both surfaces stay in lockstep instead of the popover being a stripped-down variant that skipped rename entirely.
    • · Project row status dot swaps to a spinner while its link is navigating, instead of the loader appearing next to the dot (which widened the leading column for the click's duration and read as two indicators for one thing). Trash icons across sessions, scratches, machines, keyring, and variables now render in red by default rather than only on hover, so a destructive action reads as destructive at rest.
    • · Overview column alignment: the machine chip in projects rows was left-aligned inside its 7rem cell while the scratches list rendered its chip flush against the age column, so the two lists' right-side columns landed at slightly different Xs. Both are now right-aligned inside their cells and the updated column widths match (4rem on both), so a glance down the overview scans one column, not two.
    • · App icon rebuilt from the 1024×1024 master. The generator was reading a 560×560 downscale as its source, so every rasterized icon (16 through 256) was a downscale of an already-downscaled logo and read soft in the Windows Start menu, taskbar hover card, and File Explorer info panel. The .ico now packs seven cleanly-downsampled sizes, and the splash-screen logo picks up the higher-res source too.