Redcode v0.72.9

0.72.9

Patch Changes

• Stop session execution when a vault secret request is declined, preventing repeated prompts. Add repository and global vault enable/disable configuration while preserving project .env storage and existing secret redaction.

Redcode Design design-v0.72.9

0.72.9

Patch Changes

• Stop session execution when a vault secret request is declined, preventing repeated prompts. Add repository and global vault enable/disable configuration while preserving project .env storage and existing secret redaction.

Redcode v0.72.8

0.72.8

Patch Changes

• Generate project-specific CSS token and component references with source hashes in Design manifests. Audit published CSS and style blocks for literal colors that already have project tokens and solid line heights, retaining reasoned exceptions as evidence without starting an automatic correction cycle.
• Recover interrupted Responses reasoning without replaying unfinished provider state. Keep long MCP tool names available, route MCP input requests to the calling session, and disable tool calls when generating compaction summaries. Highlight sessions waiting for permission or questions in the TUI session menus.

Redcode Design design-v0.72.8

0.72.8

Patch Changes

• Generate project-specific CSS token and component references with source hashes in Design manifests. Audit published CSS and style blocks for literal colors that already have project tokens and solid line heights, retaining reasoned exceptions as evidence without starting an automatic correction cycle.
• Recover interrupted Responses reasoning without replaying unfinished provider state. Keep long MCP tool names available, route MCP input requests to the calling session, and disable tool calls when generating compaction summaries. Highlight sessions waiting for permission or questions in the TUI session menus.

Redcode v0.72.7

0.72.7

Patch Changes

• Design audits list what matters first on a noisy page: a check found at several widths becomes one line ("390/768/1440px"), errors come before review and info, the repetition and reuse findings lead their severity instead of falling past the 30-line cap, and the report says how many checks were merged and how many more are in the full report. The reuse check is linear and bounded (200 source files, 256 KB each, 4 MB in all, with an info line when a bound left files out), and a failing reuse check becomes an info line instead of failing the audit. It no longer calls a capitalized constant a re-declared component, a local ./card import a design-system import, or every color off-token when the design system only declares a radius; colors compare with alpha stripped across hex, rgb, hsl and oklch, so hsl(220 14% 96%) matches a shadcn --muted token, and native <button> or <input> elements are noted (info) when the design system has a Button or Input. Layout signatures are taken once at the widest audited viewport and kept with their width, two different layouts made of plain boxes are no longer called alike, and text width counts East Asian characters as two columns so a translated clone is still recognised. The System One judge no longer asks about accepted exceptions, reuses its verdict while an item is unchanged, sends nothing in observe mode, and sends component, prop and import names instead of source lines. An accepted exception applies to the next audit without a new revision. Jobs written by a newer version with an unknown check severity still load (read as review). The approved-signature file is written one approval at a time, an unreadable or newer file is kept instead of overwritten, and write failures are logged.
• Design now finishes what it starts. When the Design agent stops after doing some work while a Design task is still pending or the design has no published revision yet, it is sent back to the work up to three times per message. Each reminder lists the pending tasks, and the wording changes once a revision is published. A reply that only talks, such as answering a question, ends it. An unfinished feedback round is now continued up to three times while each continuation is followed by progress. Each continuation names only the next missing step, together with the open notes. With single reasoning (the default), a message typed in the terminal, the app or an editor while a published design is open is now tracked as a review note without asking System One. The anti-slop audit no longer runs automatically at the end of a round. It runs when the reviewer asks for it, or before approval when design.gate requires it, and its findings are reported rather than fixed. Every refused note status now says it can instead be recorded unresolved or accepted with a reason, and offers partial once a verify failed the note. The Design prompt states the round rules as six numbered steps, in the same words as the review message.
• A pending Send & end can now be withdrawn. A later review message that does not end the review withdraws it, since the reviewer is clearly still reviewing, and the review page shows a "Keep reviewing" action next to "Ending after this round". The terminal sidebar now shows why a note was left partial or unresolved, the app's Design tab why a note was left partial, unresolved or accepted, and the terminal sidebar counts the partial and unresolved notes of earlier rounds on one line instead of dropping them when a newer round opens. The new Design round strings in the app are English only for now; other languages fall back to English until they are translated.
• Design review notes are now verified where they were taken and need a visible change to count as resolved. The review page records the preview width and, for app designs, the phone of each note, and the agent's review message names it. A round's verify renders each note at that width or phone (at most three viewports per verify; notes from others are verified at the nearest one and the job says so) and compares the note's element on the revision the note was taken on with the new one: pixels, text, markup or computed style, position and size. resolved is refused when the verify saw no change at all in the element, with a message telling the agent to change it or record the note unresolved or accepted with a reason. With dual reasoning, the System One review of each resolved note also sees the addressed summary, the viewport and the change, and is asked in the same request whether that change plausibly carries out the note. A design_preview publish whose prototype files are byte-identical to the previous revision now says that nothing changed.
• Design audits and verifies no longer time out on first use: the render budget now starts after the job leaves the queue and after the revision build and first-use browser install, which report a "preparing" phase on the job. Verify notes on icons inside keyed buttons and on rows that repeat an id now resolve the element the reviewer picked instead of its container or the first row.
• Design audits now catch repetition and missed design-system reuse. Each direction gets a language-independent layout signature, so two variants that are the same composition are flagged, and a new design that re-creates an approved one (from any conversation in the same application) is named. The audit also flags components declared under a design-system component's name instead of imported, literal colors and fonts the design system's tokens do not declare, and reports how many design-system components the prototype imports. With dual reasoning one System One request, sent alongside the rest of the audit, confirms these findings or downgrades false positives to info; without an answer they stay marked unconfirmed. Every check now has a stable key: record a decision accept:<key> to keep a deliberate exception from being flagged again, and the report says which keys are still flagged since the previous audit. Gradient text and backdrop blur the design system declares itself are no longer called slop. The design-system block lists components per file with their import path instead of cutting off after 40 names, so a full shadcn ui folder (Dialog, Select, Table) stays visible.
• The design review gate that refuses resolved when the verify saw no change in the note's element now judges the cases it got wrong. A note whose element could not be located on the revision it was taken on is no longer counted as changed: the verify says the change could not be measured, resolved stays possible on a clean verify, and the dual-mode System One review is told in words that this is no evidence of a change. A behavior a capture cannot show (hover, focus, a script) can be resolved when a scenario added or changed for it acts on the element and reaches its state; the note's verify line names that scenario. Page-level notes are no longer compared as one element and need only a completed verify. The refusal names each way out. Notes taken at nearly the same width (such as 1187px and 1210px, or 393px next to a 390px viewport) now share one verify viewport, and a note's verify line says when it was verified at a different width or phone than it was taken on.
• The Design review page now says why the agent is waiting or stopped: compacting its context, a scheduled retry with its countdown, a permission to approve in the terminal, a failed turn with its error, or an interruption; a cancelled message no longer stays queued. Revisions carry a number (R7) on the revision chip, the line above the preview, the picker, the activity feed and the design_preview and design_history results, and the agent names it in its reply. A refused message can be discarded or reloaded and resent, Send includes the text of an open note card, Restore asks first and waits while the agent works, the phone frame switch only previews while a new Details control sets the design's platform, and Audit, Slides to PDF, Present and the slide thumbnails follow the selected variant.
• Design no longer stops for good while review notes still lack outcomes. When the agent goes idle with notes of a round still open, it gets one continuation per round that lists the notes and what is still missing (addressed marks, the publish, a verify, the outcomes). It is skipped while a verify runs, after an interrupted turn, and for ended designs. Send & end with notes no longer deadlocks the review. The review stays open for that round (endRequested on the design) and ends by itself once every note has an outcome; the agent is told it ends after the round. A plain end works as before. The Design task reminder now says that tasks cover setup, approval and anti-slop findings, and that a design_preview publish alone proves no note outcome.
• Design review rounds now look the same in the review page, the terminal sidebar and the app's Design tab. The terminal sidebar keeps the newest round after every note has an outcome as one line (Round 3 · 12 resolved · 1 partial · 1 unresolved) and lists the partial and unresolved notes under it, names the revision under review (R7) and says when the review ends after this round, and folds the Design tasks into one line while a review has rounds. The app's Design tab shows the newest round's outcome counts, each note's outcome, the revision number and a pending end. The terminal and app feedback and approval cards name the revision (R7) when the agent's publish announced it. On the review page, Send & end is offered while earlier notes are still open as long as the message carries a note, and the round line says "Ending after this round" once it was requested.
• Stop the terminal sidebar from asking for a design's revisions when the session has no design review.
• Design tasks no longer get stuck on the evidence gate. When a Design task's completion cites no usable evidence, the newest succ...

Redcode v0.72.6

0.72.6

Patch Changes

• Reconcile Design task statuses after verified prototype work before returning to the reviewer. Record completed corrections with existing evidence, preserve unresolved and approval tasks, and avoid starting another correction cycle.

Redcode Design design-v0.72.7

0.72.7

Patch Changes

• Design audits list what matters first on a noisy page: a check found at several widths becomes one line ("390/768/1440px"), errors come before review and info, the repetition and reuse findings lead their severity instead of falling past the 30-line cap, and the report says how many checks were merged and how many more are in the full report. The reuse check is linear and bounded (200 source files, 256 KB each, 4 MB in all, with an info line when a bound left files out), and a failing reuse check becomes an info line instead of failing the audit. It no longer calls a capitalized constant a re-declared component, a local ./card import a design-system import, or every color off-token when the design system only declares a radius; colors compare with alpha stripped across hex, rgb, hsl and oklch, so hsl(220 14% 96%) matches a shadcn --muted token, and native <button> or <input> elements are noted (info) when the design system has a Button or Input. Layout signatures are taken once at the widest audited viewport and kept with their width, two different layouts made of plain boxes are no longer called alike, and text width counts East Asian characters as two columns so a translated clone is still recognised. The System One judge no longer asks about accepted exceptions, reuses its verdict while an item is unchanged, sends nothing in observe mode, and sends component, prop and import names instead of source lines. An accepted exception applies to the next audit without a new revision. Jobs written by a newer version with an unknown check severity still load (read as review). The approved-signature file is written one approval at a time, an unreadable or newer file is kept instead of overwritten, and write failures are logged.
• Design now finishes what it starts. When the Design agent stops after doing some work while a Design task is still pending or the design has no published revision yet, it is sent back to the work up to three times per message. Each reminder lists the pending tasks, and the wording changes once a revision is published. A reply that only talks, such as answering a question, ends it. An unfinished feedback round is now continued up to three times while each continuation is followed by progress. Each continuation names only the next missing step, together with the open notes. With single reasoning (the default), a message typed in the terminal, the app or an editor while a published design is open is now tracked as a review note without asking System One. The anti-slop audit no longer runs automatically at the end of a round. It runs when the reviewer asks for it, or before approval when design.gate requires it, and its findings are reported rather than fixed. Every refused note status now says it can instead be recorded unresolved or accepted with a reason, and offers partial once a verify failed the note. The Design prompt states the round rules as six numbered steps, in the same words as the review message.
• A pending Send & end can now be withdrawn. A later review message that does not end the review withdraws it, since the reviewer is clearly still reviewing, and the review page shows a "Keep reviewing" action next to "Ending after this round". The terminal sidebar now shows why a note was left partial or unresolved, the app's Design tab why a note was left partial, unresolved or accepted, and the terminal sidebar counts the partial and unresolved notes of earlier rounds on one line instead of dropping them when a newer round opens. The new Design round strings in the app are English only for now; other languages fall back to English until they are translated.
• Design review notes are now verified where they were taken and need a visible change to count as resolved. The review page records the preview width and, for app designs, the phone of each note, and the agent's review message names it. A round's verify renders each note at that width or phone (at most three viewports per verify; notes from others are verified at the nearest one and the job says so) and compares the note's element on the revision the note was taken on with the new one: pixels, text, markup or computed style, position and size. resolved is refused when the verify saw no change at all in the element, with a message telling the agent to change it or record the note unresolved or accepted with a reason. With dual reasoning, the System One review of each resolved note also sees the addressed summary, the viewport and the change, and is asked in the same request whether that change plausibly carries out the note. A design_preview publish whose prototype files are byte-identical to the previous revision now says that nothing changed.
• Design audits and verifies no longer time out on first use: the render budget now starts after the job leaves the queue and after the revision build and first-use browser install, which report a "preparing" phase on the job. Verify notes on icons inside keyed buttons and on rows that repeat an id now resolve the element the reviewer picked instead of its container or the first row.
• Design audits now catch repetition and missed design-system reuse. Each direction gets a language-independent layout signature, so two variants that are the same composition are flagged, and a new design that re-creates an approved one (from any conversation in the same application) is named. The audit also flags components declared under a design-system component's name instead of imported, literal colors and fonts the design system's tokens do not declare, and reports how many design-system components the prototype imports. With dual reasoning one System One request, sent alongside the rest of the audit, confirms these findings or downgrades false positives to info; without an answer they stay marked unconfirmed. Every check now has a stable key: record a decision accept:<key> to keep a deliberate exception from being flagged again, and the report says which keys are still flagged since the previous audit. Gradient text and backdrop blur the design system declares itself are no longer called slop. The design-system block lists components per file with their import path instead of cutting off after 40 names, so a full shadcn ui folder (Dialog, Select, Table) stays visible.
• The design review gate that refuses resolved when the verify saw no change in the note's element now judges the cases it got wrong. A note whose element could not be located on the revision it was taken on is no longer counted as changed: the verify says the change could not be measured, resolved stays possible on a clean verify, and the dual-mode System One review is told in words that this is no evidence of a change. A behavior a capture cannot show (hover, focus, a script) can be resolved when a scenario added or changed for it acts on the element and reaches its state; the note's verify line names that scenario. Page-level notes are no longer compared as one element and need only a completed verify. The refusal names each way out. Notes taken at nearly the same width (such as 1187px and 1210px, or 393px next to a 390px viewport) now share one verify viewport, and a note's verify line says when it was verified at a different width or phone than it was taken on.
• The Design review page now says why the agent is waiting or stopped: compacting its context, a scheduled retry with its countdown, a permission to approve in the terminal, a failed turn with its error, or an interruption; a cancelled message no longer stays queued. Revisions carry a number (R7) on the revision chip, the line above the preview, the picker, the activity feed and the design_preview and design_history results, and the agent names it in its reply. A refused message can be discarded or reloaded and resent, Send includes the text of an open note card, Restore asks first and waits while the agent works, the phone frame switch only previews while a new Details control sets the design's platform, and Audit, Slides to PDF, Present and the slide thumbnails follow the selected variant.
• Design no longer stops for good while review notes still lack outcomes. When the agent goes idle with notes of a round still open, it gets one continuation per round that lists the notes and what is still missing (addressed marks, the publish, a verify, the outcomes). It is skipped while a verify runs, after an interrupted turn, and for ended designs. Send & end with notes no longer deadlocks the review. The review stays open for that round (endRequested on the design) and ends by itself once every note has an outcome; the agent is told it ends after the round. A plain end works as before. The Design task reminder now says that tasks cover setup, approval and anti-slop findings, and that a design_preview publish alone proves no note outcome.
• Design review rounds now look the same in the review page, the terminal sidebar and the app's Design tab. The terminal sidebar keeps the newest round after every note has an outcome as one line (Round 3 · 12 resolved · 1 partial · 1 unresolved) and lists the partial and unresolved notes under it, names the revision under review (R7) and says when the review ends after this round, and folds the Design tasks into one line while a review has rounds. The app's Design tab shows the newest round's outcome counts, each note's outcome, the revision number and a pending end. The terminal and app feedback and approval cards name the revision (R7) when the agent's publish announced it. On the review page, Send & end is offered while earlier notes are still open as long as the message carries a note, and the round line says "Ending after this round" once it was requested.
• Stop the terminal sidebar from asking for a design's revisions when the session has no design review.
• Design tasks no longer get stuck on the evidence gate. When a Design task's completion cites no usable evidence, the newest succ...

Redcode Design design-v0.72.6

0.72.6

Patch Changes

• Reconcile Design task statuses after verified prototype work before returning to the reviewer. Record completed corrections with existing evidence, preserve unresolved and approval tasks, and avoid starting another correction cycle.

Redcode v0.72.5

0.72.5

Patch Changes

• Stop models from polling CI in a shell loop. When a wait cannot become a native monitor probe (a pull request check, a deploy, a job status), the refusal now hands over the same command with background: true, which releases the turn and resumes the session when it exits, instead of suggesting a for i in 1 2 3 4 5 loop. A refused to-do update now says in plain words what to fix (for example that a task has no observable acceptance criterion) instead of listing question ids.

• The Design review page's side panel is now a compact Feedback tab. The newest round's notes read as a checklist, one line per note with a status mark, and a note expands to show its element, what the agent says it changed and the recorded outcome. Settled rounds fold into one line with their tallies, "Only what is left" hides what is done, the agent's last reply stays in view with the rest of the conversation under Activity, and the message box stays pinned at the bottom.

  While the agent works, the round shows how far it has come: received, fixing (how many notes are addressed), stopped, published, verifying, outcomes missing, or ready for review. Ready for review appears only once every note has an outcome on a verified latest revision and the agent is idle. The panel also shows the agent's live activity and how long ago the round was received.

  A chip next to the revision picker tells you whether you are looking at the latest revision (Latest, N behind, Updating…, Load failed, Offline). A line above the preview says which round the revision on screen answers. When a newer revision has not replaced it yet, the line says why and offers the next step: add your open note and switch, or retry a send or a failed load. When you switch to the latest revision, your queued notes and message now move with you instead of staying on the old revision, so they can no longer be sent twice. Clicking a draft's Remove or Send in the message box no longer misses when the box grows.

• Design review notes are now a checklist the agent cannot silently skip. The agent marks each note it changed with design_document update addressed, and design_preview refuses to publish an answer to a feedback round while a note of that round has neither a mark nor an outcome, quoting those notes. Publishing a preset, tweak or restore from the review page no longer closes the open round; only the agent's design_preview answers it. Publish, verify and design_document results end with what each round still waits for, and the Design context keeps a pointer to the round's open notes after compaction.

• Show the MCP servers right below Context in the terminal sidebar and move Modified Files down to just above the footer, after the to-dos and the language servers.

• The terminal sidebar shows where a Design review round stands while the agent works through it: Round 9 · 8/14 addressed · 3 recorded under the to-do group, followed by up to three notes that still have neither an addressed mark nor an outcome, each on one line with the element and the reviewer's words. It refreshes on the same 5 second poll as the tasks and disappears once every note of the round has an outcome or the review has ended.

Redcode v0.72.3

0.72.3

Patch Changes

• Compaction no longer loops when the conversation outgrows the context window the model is believed to have. A long exchange is summarized down to its newest steps instead of being carried whole in every checkpoint, which sent requests of two to four times the window with 1,024 output tokens and compacted again on the next step. When no checkpoint can fit the window, the turn ends with one message naming the history size, what would remain, the window in use and where it came from, and what to do. A compaction request rejected as too long now shrinks to the window's target rather than to 70% of its own size. The sidebar, prompt footer and mini footer show the window beside the context usage, for example 321.0K / 115.2K (279%).

• Show the Design approval card in the terminal transcript. The approval handoff message carries no description, so the transcript dropped it with other internal messages and the approval, with its link back to the review, never appeared.

• Capture the approval screenshot of real-sized Design prototypes. The capture cloned every element of the document, hidden screens and variants included, with every computed style, and gave up after 8 seconds, so a prototype with a few thousand elements always ended with "Preview screenshot unavailable". It now skips what is not rendered and writes only the styles that differ from the browser defaults, which produces the same image several times faster, and waits up to 20 seconds.

  A variant root without a box of its own (display: contents, a shell of fixed-position children) is recognised as the variant on screen, and a prototype that defines module or exports globals no longer keeps the capture runtime from loading. When a capture still fails, the notice states the reason.

• Record the outcomes of Design review notes one by one. One mistyped note id or one refused status used to void the whole design_document update, so every note of the round stayed open and approval stayed blocked. Each status is now judged on its own: the valid ones are recorded, and the result tells the agent how many were recorded, why each refused one was refused, and which notes still have no outcome, round by round, with the reviewer's words.

  With dual reasoning, the System One review of those outcomes no longer receives every note and every job of the design. It is sent only the notes being recorded as resolved or partial, each with what the reviewer asked and what the cited verify saw of it, so it stops failing with "Semantic evaluation unavailable" once a design has a few rounds. A review that is unavailable or inconclusive records the status and reports it as unverified instead of leaving the note open; only a note the review contradicts is refused.

  The verify report quotes each reviewer's note under what the verify saw of it, and the message that blocks approval names the design_read call that lists a round's notes with their text.

• Show a Design review's round and the number of notes it was sent with on its transcript card, in the terminal and in the app. A review stored by an older version that cut the message short now says so on its card, for example "8 of 14 notes reached the agent", instead of listing the surviving notes as if they were the whole review.

• Deliver every note of a Design review to the agent. A review message used to be cut at 8,000 characters, so a round of 14 notes reached the agent as 7 or 8 and the rest were silently lost. The message is no longer cut: a note on an element with its own data-design-id carries no redundant locators, a long message first shortens the page text captured with each note, then leaves out backup locators, and the words the reviewer wrote are never shortened.

  The agent can read a round's notes with their status, or one note with every locator, through design_read section notes. A review too long for one message is refused with a request to send it in two parts instead of being truncated, and the review page releases a refused review so its notes can be edited and sent again. The review feed reports the number of notes that were sent, including for messages stored before this change.

• Models on an OpenAI-compatible endpoint added through the connect wizard no longer get a 115,200-token window frozen into the global configuration when the endpoint does not report one. The wizard writes only the limits the endpoint reported or you entered; everything else is resolved from the models catalog every time the provider loads, so a model the catalog knows (glm-5.3-flash, mimo-v2.6-pro, ...) gets its real window and a configuration that an older version froze the guess into heals on load. The catalog lookup now tolerates the id shapes gateways use (case, leading vendor segments, :free/:thinking/@region suffixes, dots against dashes) without ever matching a bare family name.

• A model selected through a saved RedRouter or 9Router connection no longer keeps a guessed context window. The connection's saved catalog now holds only the limits the router reported; a limit the router left out follows the current models catalog every time the model is loaded, so a catalog refresh reaches saved selections without waiting for the router's list to change, and a guess frozen by an earlier version heals on load. A limit set in configuration under providers.<id>.models.<id>.limit now also applies to a model resolved through a saved connection.