[Release] 2026-09-24: v2.6.0 — phone remote bridge over LAN and Tailscale HTTPS, phone microphone, safer goal mode, invented-URL guard

[nibor1896]

Summary

Release v2.6.0: 91 commits since origin/main (v2.5.0). Crow now runs on the phone. The remote bridge (#249) pairs over the LAN with a QR code, and away from home it runs over Tailscale HTTPS; the installers gain an optional --tailscale / -Tailscale step. The phone microphone (#290) shows a live transcript, stops on silence and appends to what is already typed. Goal mode got safer: the fresh-eyes judge rubric names the step (#266/#286), visual steps need evidence (#267), a render loop is broken off (#268/#277), and a failed step is retried once, then skipped, with /goal skip (#289). Tools: invented URLs are refused (#288), edit_file gets a syntax check and hints (#269/#276/#283), WebGL errors carry hints (#253), session_search covers rollover segments (#287). Memory: the consolidation tile says what did not land and shows the full text (#285). Also: crow.log, context clearing (#262/#263). Docs are synced.

Scope

Group	Count
Close on merge (shipped and tested; live check noted in each)	9
Resolved in part, stays open	1 (#249: stages 2–4)
Already closed on live or measured evidence, shipped by this PR	32

Resolved issues (close on merge)

• Closes delegate: a subtask stopped while its spot is failing ends 'failed' and memos the spot dead, instead of 'interrupted' (#143 E2 contract) #242 — delegate: a subtask stopped while its spot is failing ends 'failed' and memos the spot dead, instead of 'interrupted' (One chat at a time is the ceiling: concurrent sessions over remote spots, subtask delegation, local slot untouched #143 E2 contract)
• Closes Goal mode accepts 'done' without evidence: 9/9 complete over a page that draws nothing, with step notes saying 'done-with-deviation only in spirit' and a failed step flipped to done with no note #250 — Goal mode accepts 'done' without evidence: 9/9 complete over a page that draws nothing, with step notes saying 'done-with-deviation only in
• Closes The model believes the machine has no GPU (RTX 5090) and saves it to memory: nothing in its context says what the machine is, and the memory gate stores contradicting facts #270 — The model believes the machine has no GPU (RTX 5090) and saves it to memory: nothing in its context says what the machine is, and the memory
• Closes The browser panel's WebKit process segfaults in the NVIDIA driver right after each GPU render_page: 512 MiB headroom ignores Crow's own panel as a second GPU client next to serve #279 — The browser panel's WebKit process segfaults in the NVIDIA driver right after each GPU render_page: 512 MiB headroom ignores Crow's own pane
• Closes Memory Consolidation tile: 'save to memory' can drop a write without a word (refusal or 5-min expiry), and the tile shows only 160 chars and never the entry a replace removes #285 — Memory Consolidation tile: 'save to memory' can drop a write without a word (refusal or 5-min expiry), and the tile shows only 160 chars and
• Closes judge: a visual step can pass on a rubric the model picks -- step 3 'G-buffer and voxel volume' was accepted on step 2's structure/camera criteria (8/9/8) #286 — judge: a visual step can pass on a rubric the model picks -- step 3 'G-buffer and voxel volume' was accepted on step 2's structure/camera cr
• Closes render_page/fetch_url fetch URLs the model invented: 11 renders of a made-up Alibaba OSS proxy URL in step 4, error pages taken as scene frames #288 — render_page/fetch_url fetch URLs the model invented: 11 renders of a made-up Alibaba OSS proxy URL in step 4, error pages taken as scene fra
• Closes goal: a 'failed' step is sent back as the next step forever (step 4 looped ~40 min after the model failed it honestly), and robin has no /goal skip #289 — goal: a 'failed' step is sent back as the next step forever (step 4 looped ~40 min after the model failed it honestly), and robin has no /go
• Closes Phone: the 🎤 records nothing -- getUserMedia needs a secure context and the mirror is plain HTTP; record on the phone over Tailscale HTTPS and transcribe with crow_voice on the PC #290 — Phone: the 🎤 records nothing -- getUserMedia needs a secure context and the mirror is plain HTTP; record on the phone over Tailscale HTTPS a

Resolved in part (stays open)

• Refs /remote: operate a running session from the phone (LAN) #249 — stage 1 (LAN) and stage 5 (Tailscale HTTPS) are accepted live on 2026-09-24; stages 2 (TUI), 3 (Windows) and 4 (docs) are still open.

Already closed, shipped by this PR

• session_search never indexes rollover segments: 0 of 6 rollover-*.json on disk are searchable, so a long goal run's earlier half cannot be found #287 — session_search never indexes rollover segments: 0 of 6 rollover-*.json on disk are searchable, so a long goal run's earl
• delegate: a retired free model's HTTP 404 ('unavailable for free') is booked transient via the 'unavailable' substring and burns 1 of 3 transient attempts in every subtask #284 — delegate: a retired free model's HTTP 404 ('unavailable for free') is booked transient via the 'unavailable' substring a
• edit_file rewrites a whole CRLF file as LF: a one-line edit turned 40 CRLF lines into 0 (universal-newline read, newline="" write) #283 — edit_file rewrites a whole CRLF file as LF: a one-line edit turned 40 CRLF lines into 0 (universal-newline read, newline
• Stop does not stop goal mode: run_turn consumes INTERRUPT before the pump asks, so the engine starts the next turn at once #282 — Stop does not stop goal mode: run_turn consumes INTERRUPT before the pump asks, so the engine starts the next turn at on
• Subtasks card has no close button: the #255 panel and its column reserve cannot be dismissed, unlike the goal card #281 — Subtasks card has no close button: the Subagent tiles scroll away with the chat: subCard appends every delegate card into #flow instead of a pinned place beside the goal and git cards #255 panel and its column reserve cannot be dismissed, unlike the goal card
• Chat bubbles do not stand on the composer's edges: the user's bubble hugs the column's left (716-780 px short of the box's right edge), Crow's text starts 40 px in #280 — Chat bubbles do not stand on the composer's edges: the user's bubble hugs the column's left (716-780 px short of the box
• #268's loop breaker never fired on 9 near-black captures in step 2 (2026-09-24): a dark frame with a lit patch is not 'blank', and turns across a mid-turn rollover or behind a #202 nudge are never counted #277 — Goal mode has no brake for a working debug loop: 7 black captures in a row (22:30-22:48) while every turn called tools; after 3 near-identical captures nudge to bisect, after 6 force a rollover that lists what was tried #268's loop breaker never fired on 9 near-black captures in step 2 (2026-09-24): a dark frame with a lit patch is not 'b
• edit_file answers a missed 'old' with one bare line: 16 of 147 diorama edits failed, 4 of them on whitespace alone, and the model retried blind (4 in a row on mesh.js, goal brake twice) #276 — edit_file answers a missed 'old' with one bare line: 16 of 147 diorama edits failed, 4 of them on whitespace alone, and
• build_bundle looks for deno's esbuild in ~/.cache/crow/deno and tells the model to set an env var it cannot set: 4 rounds lost and a foreign binary copied into the diorama project (refs #212) #274 — build_bundle looks for deno's esbuild in ~/.cache/crow/deno and tells the model to set an env var it cannot set: 4 round
• render_page after an edit is answered from the per-turn _SEEN cache ('The result was, and still is:') with the pre-edit capture; the model learned to change width/height to get a real render #273 — render_page after an edit is answered from the per-turn _SEEN cache ('The result was, and still is:') with the pre-edit
• render_page refuses a local page with a ?query or #fragment ('no such page: index.html?view=albedo'), so the diorama prompt's required screenshot mode index.html?shot=<view>&w=<px> cannot be used and the model stored 'file:// rejects a ?query' in memory #272 — render_page refuses a local page with a ?query or #fragment ('no such page: index.html?view=albedo'), so the diorama pro
• render_page says 'software (swiftshader)' but not why: 49 of 49 captures fell back while serve held the RTX 5090, and the model concluded the machine has no GPU #271 — render_page says 'software (swiftshader)' but not why: 49 of 49 captures fell back while serve held the RTX 5090, and th
• edit_file runs no syntax check (write/append do since #251), and every new format needs a code change: one extension→command table for all three file tools #269 — edit_file runs no syntax check (write/append do since write_file/append_file never parse what they wrote: 20 of 98 JS/HTML writes on 2026-09-23 did not pass node --check, and the model learnt it from renders hours later #251), and every new format needs a code change: one extension→com
• Goal mode has no brake for a working debug loop: 7 black captures in a row (22:30-22:48) while every turn called tools; after 3 near-identical captures nudge to bisect, after 6 force a rollover that lists what was tried #268 — Goal mode has no brake for a working debug loop: 7 black captures in a row (22:30-22:48) while every turn called tools;
• goal_step accepts 'done' on visual steps with no picture and no score: 9/9 closed over a small box in a black frame; a visual step must cite a capture from this goal and meet the judge's bar #267 — goal_step accepts 'done' on visual steps with no picture and no score: 9/9 closed over a small box in a black frame; a v
• No fresh-eyes judge for visual work: the model scored its own black frame 9+ on every criterion; a judge tool must score the capture against the rubric on a model that never saw the transcript #266 — No fresh-eyes judge for visual work: the model scored its own black frame 9+ on every criterion; a judge tool must score
• read_image shows a small scene at a few dozen visual tokens: the diorama filled 7-20 % of each render on a black ground (~50-120 of 527-880 tokens), the model graded it 9+ unseen #265 — read_image shows a small scene at a few dozen visual tokens: the diorama filled 7-20 % of each render on a black ground
• The window stops a running turn on every typed line: goal mode's 'a typed line has priority' (#165) has no way in from the live chat #264 — The window stops a running turn on every typed line: goal mode's 'a typed line has priority' (The goal engine: plan with research tools, execute, verify, continue after a pause #165) has no way in from t
• Long goal runs roll over with a third of the window spent on tool results the model already acted on: no tool-result clearing below the 0.9 cut #263 — Long goal runs roll over with a third of the window spent on tool results the model already acted on: no tool-result cle
• Crow's own status lines (goal brake, same-failure streak, degenerate round, cache, budget) clutter the chat and every reopened session: they belong in a log file, and Crow has none #262 — Crow's own status lines (goal brake, same-failure streak, degenerate round, cache, budget) clutter the chat and every re
• Every rollover adds its archived segment to the chat rail as a separate chat, titled by Crow's own note ('[The tool budget for this turn is …') #261 — Every rollover adds its archived segment to the chat rail as a separate chat, titled by Crow's own note ('[The tool budg
• goal-mode brake reads a budget-spent working turn as an empty loop: 157 + 52 messages of real work cut from the history, each cut a 106k-token COLD re-prefill (~3 min) #258 — goal-mode brake reads a budget-spent working turn as an empty loop: 157 + 52 messages of real work cut from the history,
• Composer and chat column sit 158 px left of the window centre while a goal/git card stands (#233 right-only reserve), 5 px left otherwise #256 — Composer and chat column sit 158 px left of the window centre while a goal/git card stands (Floating goal/git cards cover the composer, open its menus underneath them, and cover the chat column #233 right-only reserve), 5
• Subagent tiles scroll away with the chat: subCard appends every delegate card into #flow instead of a pinned place beside the goal and git cards #255 — Subagent tiles scroll away with the chat: subCard appends every delegate card into #flow instead of a pinned place besid
• write_file/append_file descriptions say 'one append per section' and never what 'large' is: the diorama run built 1-5 KB files from a head plus up to 10 appends (61 writes median 940 B, 50 appends median 319 B, 0 near the 16384 cap) #254 — write_file/append_file descriptions say 'one append per section' and never what 'large' is: the diorama run built 1-5 KB
• render_page passes 'texImage33D is not a function' and 'INVALID_ENUM: renderbufferStorage' through bare, so the model read its own misspelling and 5-argument call as 'no 3D on this machine' for hours: the result must name the real API member and the argument count vs signature #253 — render_page passes 'texImage33D is not a function' and 'INVALID_ENUM: renderbufferStorage' through bare, so the model re
• The write result counts characters as bytes and never says the write is byte-exact, so the model blamed Crow's file system for its own corrupted content #252 — The write result counts characters as bytes and never says the write is byte-exact, so the model blamed Crow's file syst
• write_file/append_file never parse what they wrote: 20 of 98 JS/HTML writes on 2026-09-23 did not pass node --check, and the model learnt it from renders hours later #251 — write_file/append_file never parse what they wrote: 20 of 98 JS/HTML writes on 2026-09-23 did not pass node --check, and
• windows-latest CI red since 2026-09-19: two YoloTurnTests use cat /etc/os-release, which Windows by design does not read as an outside path #248 — windows-latest CI red since 2026-09-19: two YoloTurnTests use cat /etc/os-release, which Windows by design does not read
• a launch over a crashed session's session.json closes the window: _probe's restore() raises on a non-fresh conversation while the rail throws this.col.insertBefore and every push floods GError 601 #209 — a launch over a crashed session's session.json closes the window: _probe's restore() raises on a non-fresh conversation
• Crow's own inference engine — decision recorded, goals open #188 — Crow's own inference engine — decision recorded, goals open
• Flash-Next beyond 959.81 / 28.60 #159 — Flash-Next beyond 959.81 / 28.60

Checks (head 7e7246f, Linux, crow venv)

Check	Result
test_crow_remote + test_crow_core + test_crow + test_crow_gui	2,880 OK (8 skipped)
ruff	clean
check_shared_core	83/83
check_operating_point	10/10
install.sh --selftest	39 checks PASS (fake tailscale, sudo guard)
install.ps1 -Selftest	not run (no PowerShell on this machine)
check_gui_prereqs	fails on 13 glyphs missing from Google Sans Code (12 known since 2.5.0, plus 🎤); not in CI

CHANGELOG

The phone becomes a second view of the session, and goal runs get a fresh-eyes judge. A paired phone mirrors
the open session on the home LAN and, through Tailscale serve, over HTTPS from anywhere, with its own microphone
(#249 stages 1 and 5, #290). Goal mode gains a judge that scores captures in a fresh request, an evidence gate for
visual steps, a render-loop breaker, retry-then-skip for failed steps and /goal skip (#265–#268, #277, #286,
#289). Long runs keep more of the window: old tool results are cleared below the rollover, and Crow's own status
lines move to crow.log (#262, #263). The file tools share one syntax-check table and keep line endings (#269,
#276, #283); render_page, build_bundle, memory, session_search and delegate each close a failure seen in the
2026-09-23/24 diorama goal runs (#270–#275, #278, #279, #284, #285, #287, #288).

Everything on local main since 2.5.0 (origin/main 0c3e392): 87 commits, 2026-09-23 to 2026-09-24. Most
measurements are against robin's 2026-09-23 diorama goal run and its 2026-09-24 follow-ups. Tickets stay open until
robin's live check unless the ticket says otherwise.

Added

• Phone mirror on the home LAN (/remote: operate a running session from the phone (LAN) #249 stage 1). /remote on (or the phone icon left of code/git/browser in the
  title bar) starts cli/crow_remote.py, a standard-library HTTP server on one LAN address and a fixed port
  (remote_port, default 8765; 0.0.0.0 is refused). The Remote dialog shows a QR code with a single-use 120 s
  pairing code; a new device waits up to 60 s for Allow on the desktop, otherwise it is denied. A paired phone
  keeps an HttpOnly, SameSite=Strict cookie for 400 days, of which only the sha256 is stored (devices in
  secrets.json); 5 failed pairings lock pairing for 10 minutes, and revoking a device ends its stream at once.
  Host (421), Origin (403) and cookie (401) guards on every request. The phone gets the session over SSE (a
  per-device replay ring of 5,000 events, snapshot otherwise, 15 s heartbeat, text coalesced to ≤ 4 updates/s) and
  calls an allowlist of 58 proxied page methods; each client has its own view, so a phone never moves the desktop
  and the other way round. The phone layer (below 700 px): a one-row composer, rail/code/git/browser as drawers, a
  44 px goal bar, pinned approvals, image upload (20 MiB), its own home-screen icon and title, and an auto-hiding
  header in the iOS home-screen app. /remote on|off|status|devices|forget <name>; settings remote_enabled,
  remote_port, remote_host. The QR encoder (byte mode, level M, versions 1–6) matches segno module for module
  on 20 URLs.
• Phone mirror over HTTPS via Tailscale (/remote: operate a running session from the phone (LAN) #249 stage 5). With a tailnet up, Crow additionally listens on
  127.0.0.1:<remote_port> for tailscale serve --bg --https=443, where only the machine's ts.net Host and
  Origin pass (a bare 127.0.0.1 Host stays 421) and the cookie is Secure. Crow never runs a Tailscale command
  that changes anything: it reads tailscale status --json and tailscale serve status --json without sudo into
  six states (missing, down, https-off, serve-missing, funnel, ready), and the Remote dialog offers
  HTTPS · <name> with one status line per state and the one-time serve command to copy (setting
  remote_https). Never plain HTTP on 0.0.0.0 or the 100.x address; with Funnel on for the name there is no
  loopback listener at all. Setup guide: docs/user-guide/remote-tailscale.md.
• Tailscale as an opt-in installer component (/remote: operate a running session from the phone (LAN) #249). install.sh --tailscale / install.ps1 -Tailscale read
  what is already done (read-only tailscale status --json / tailscale serve status --json, through the same
  state table as the Remote dialog) and print only the missing steps: the install command for the distribution
  (sudo pacman -S tailscale on Arch/Omarchy and Arch-likes, the official curl -fsSL https://tailscale.com/install.sh | sh elsewhere, winget install --id Tailscale.Tailscale -e on Windows),
  sudo systemctl enable --now tailscaled, tailscale up, Enable HTTPS in the admin console,
  tailscale serve --bg --https=443 http://127.0.0.1:<remote_port> (in an elevated shell on Windows) and the phone
  app when no iOS/Android device is in the tailnet yet. On Linux the section follows the install; on Windows
  -Tailscale prints it and exits without downloading anything. Neither installer runs sudo or elevates.
  install.sh --selftest drives every state against a fake tailscale on PATH (39 checks); install.ps1 -Selftest does the same through a fake seam.
• Phone microphone with a live transcript (Phone: the 🎤 records nothing -- getUserMedia needs a secure context and the mirror is plain HTTP; record on the phone over Tailscale HTTPS and transcribe with crow_voice on the PC #290). On the HTTPS address the phone records with MediaRecorder
  and uploads the clip with the same guards; the window transcribes it with crow_voice (same model and settings
  as the desktop's dictation) and puts the words into that phone's input, never sent. Every 1.5 s the recording so
  far is transcribed as a greyed partial; recording stops after 2 s of silence once speech was heard; the button
  shows listening / writing / loading the speech model. One transcription per phone at a time, only the newest
  partial waits, finals first. One crow.log line per final dictation. Measured 2026-09-24 (faster-whisper small
  int8, CPU): 5 s clip 1.25–1.4 s, 10 s 1.4 s, 15 s 1.5 s; model load from cache 0.85 s. Plain HTTP keeps the
  keyboard-dictation hint.
• Fresh-eyes judge for visual work (No fresh-eyes judge for visual work: the model scored its own black frame 9+ on every criterion; a judge tool must score the capture against the rubric on a model that never saw the transcript #266, judge: a visual step can pass on a rubric the model picks -- step 3 'G-buffer and voxel volume' was accepted on step 2's structure/camera criteria (8/9/8) #286). The judge tool sends images and a rubric to a vision model
  in a fresh request with no history and stores the score in goal.json; without a pin it uses the local model. The
  rubric leads with the step's own text ("delivers: "); the goal's accept: lines, PLAN.md, the caller's
  criteria or a default rubric only add to it (judge: a visual step can pass on a rubric the model picks -- step 3 'G-buffer and voxel volume' was accepted on step 2's structure/camera criteria (8/9/8) #286: step 3 "G-buffer and voxel volume" had passed on step 2's
  camera criteria). The verdict records its rubric source. Live on the 2026-09-23 "9+" final frame: 4 vision models
  scored it at most 2.
• Evidence gate for visual steps (goal_step accepts 'done' on visual steps with no picture and no score: 9/9 closed over a small box in a black frame; a visual step must cite a capture from this goal and meet the judge's bar #267). goal_step done on a visual step needs a capture from this goal and
  a judge minimum of at least judge_threshold (default 8; --judge-threshold, settings judge_threshold).
• Render-loop breaker (Goal mode has no brake for a working debug loop: 7 black captures in a row (22:30-22:48) while every turn called tools; after 3 near-identical captures nudge to bisect, after 6 force a rollover that lists what was tried #268, #268's loop breaker never fired on 9 near-black captures in step 2 (2026-09-24): a dark frame with a lit patch is not 'blank', and turns across a mid-turn rollover or behind a #202 nudge are never counted #277). Three captures of one page that come back the same get a bisect nudge; six
  force a rollover that carries what was tried. A capture with a luma mean ≤ 4/255 counts as stuck, and captures
  are counted in every turn, including turns cut by a same-failure nudge or a mid-turn rollover (#268's loop breaker never fired on 9 near-black captures in step 2 (2026-09-24): a dark frame with a lit patch is not 'blank', and turns across a mid-turn rollover or behind a #202 nudge are never counted #277: on
  2026-09-24, 9 near-black captures of one page, luma mean 1–2/255, got 0 nudges; the replay nudges at session
  message 13 and forces the roll at 155). A black streak gets a concrete bisect (one pass in isolation; getError,
  checkFramebufferStatus, readPixels).
• Failed goal steps are retried once, then skipped; /goal skip (goal: a 'failed' step is sent back as the next step forever (step 4 looped ~40 min after the model failed it honestly), and robin has no /goal skip #289). The first failed sends the step back
  with the failure note and asks for a different approach; the second marks it skipped and the goal moves on. A
  goal whose steps are all done or skipped ends partial ("complete with N skipped"), never done.
  /goal skip <n> [reason] works in the window, on the phone and in the terminal. The goal bar shows skipped steps
  and follows goal.json every round, before every turn and on /goal, so a hand edit shows at once.
• render_page metrics and an enlarged crop (read_image shows a small scene at a few dozen visual tokens: the diorama filled 7-20 % of each render on a black ground (~50-120 of 527-880 tokens), the model graded it 9+ unseen #265). Coverage, colours and luma per render; a small scene gets an
  enlarged content crop for read_image. 2026-09-23: the scene filled 7–22 % of 62 renders.
• Context editing below the rollover (Long goal runs roll over with a third of the window spent on tool results the model already acted on: no tool-result clearing below the 0.9 cut #263). At 0.65 of the window, tool results older than the last 5 rounds
  become a one-line stub, in batches that free at least 10 %; the originals go to session/cleared/.
  context_clear_at in settings, --context-clear-at. Replay: rollovers 55 and 88 rounds later.
• crow.log (Crow's own status lines (goal brake, same-failure streak, degenerate round, cache, budget) clutter the chat and every reopened session: they belong in a log file, and Crow has none #262). Crow's own status lines go to ~/.local/state/crow/log/crow.log with local time and offset,
  rotated, instead of into the chat: goal brake, same-failure streak, degenerate round, cache, budget, context
  clearing (Long goal runs roll over with a third of the window spent on tool results the model already acted on: no tool-result clearing below the 0.9 cut #263), the rollover line and The boundary refuses write_file, and the model reaches the same path with run_command in the next call #98's boundary alarm, chat-open and /goal setup notes, the "no folder"
  note, a browser-panel crash (The browser panel's WebKit process segfaults in the NVIDIA driver right after each GPU render_page: 512 MiB headroom ignores Crow's own panel as a second GPU client next to serve #279) and rejected memory writes (Memory Consolidation tile: 'save to memory' can drop a write without a word (refusal or 5-min expiry), and the tile shows only 160 chars and never the entry a replace removes #285). The roll card, the /goal status answer
  and the Browser panel runs foreign pages in Crow's own web context: no memory kill (kill threshold 0), no sandbox, pywebview bridge injected, logins lost on restart #226 memory-ceiling stop stay in the chat.
• One syntax-check table for all file tools (edit_file runs no syntax check (write/append do since #251), and every new format needs a code change: one extension→command table for all three file tools #269). edit_file parses what it left (node for JS/HTML) and says
  whether this edit broke the file or it was broken before; settings syntax_checks maps an extension to a command
  for write_file, append_file and edit_file. Replay of 2026-09-23: 0 of 66 JS/HTML edits broke a file —
  parity, not a measured failure.
• Machine facts in the prompt head, and memory may not contradict them (The model believes the machine has no GPU (RTX 5090) and saves it to memory: nothing in its context says what the machine is, and the memory gate stores contradicting facts #270). The head names OS, CPU, RAM, GPU
  and total VRAM (probed once per process) and says a tool's limit is not the machine's; memory refuses a
  no-GPU/CPU-only note that does not name the real card, and a "tool changes bytes" note. 2026-09-24: the diorama
  run had saved "Machine has NO GPU" on an RTX 5090.
• render_page says why it rendered in software (render_page says 'software (swiftshader)' but not why: 49 of 49 captures fell back while serve held the RTX 5090, and the model concluded the machine has no GPU #271): the card's free VRAM against the bound, that the model
  server holds it, and that the machine has the GPU. 2026-09-23: 49 of 49 captures fell back.

Changed

• Version 2.6.0.
• Hosts nobody named are refused (render_page/fetch_url fetch URLs the model invented: 11 renders of a made-up Alibaba OSS proxy URL in step 4, error pages taken as scene frames #288). render_page and fetch_url refuse a remote host that appears nowhere
  in the conversation (the user's words, a tool result, the goal, PLAN.md) before any socket or browser opens, with
  a next step; loopback always passes, a subdomain of a named host counts, and it is on in every mode, yolo
  included. Across a rollover a Crow-written line carries the newest 60 hosts.
• The memory gate answers every approved write (Memory Consolidation tile: 'save to memory' can drop a write without a word (refusal or 5-min expiry), and the tile shows only 160 chars and never the entry a replace removes #285). A refusal, a duplicate or an expired entry comes back as
  "Memory: not saved -- " in the window and the terminal, and goes to crow.log. The Memory Consolidation
  tile has three states (collapsed, previews, full text) and shows a replace as "- old" above "+ new".
• session_search covers rollover segments and chats put aside (session_search never indexes rollover segments: 0 of 6 rollover-*.json on disk are searchable, so a long goal run's earlier half cannot be found #287). 0 of 6 rollover segments on disk were
  searchable, because Every rollover adds its archived segment to the chat rail as a separate chat, titled by Crow's own note ('[The tool budget for this turn is …') #261 took them out of the rail the index read. A segment hit is labelled
  " (before the cut, )", every hit prints its path, and a search in the first turn of a new chat
  no longer drops the chat just put aside.
• A line typed during a turn is queued, never a stop (The window stops a running turn on every typed line: goal mode's 'a typed line has priority' (#165) has no way in from the live chat #264, Stop does not stop goal mode: run_turn consumes INTERRUPT before the pump asks, so the engine starts the next turn at once #282). In goal mode it goes before the goal nudge.
  A line opening with / that is not a Crow command (a path) is queued too; the button reads ↑ Queue while the
  box holds a line; a Crow command waits in the box; an IME-confirming Enter is not a submit. Stop pauses goal
  mode (Stop does not stop goal mode: run_turn consumes INTERRUPT before the pump asks, so the engine starts the next turn at once #282): Stop used to end one turn while the engine started the next at once; now the engine pauses with
  a note until the next line you send.
• edit_file answers a missed old with the closest text (edit_file answers a missed 'old' with one bare line: 16 of 147 diorama edits failed, 4 of them on whitespace alone, and the model retried blind (4 in a row on mesh.js, goal brake twice) #276): line numbers, what differs, whitespace-only
  named; a uniform indentation drift with one unambiguous window is applied and said. 16 of 147 edits missed on
  2026-09-23/24; replayed, 3 now land and 11 of the other 12 get the region.
• Chat bubbles stand on the composer's edges (Chat bubbles do not stand on the composer's edges: the user's bubble hugs the column's left (716-780 px short of the box's right edge), Crow's text starts 40 px in #280): your bubble right-aligned on the input box's right border,
  Crow's text on its left border (the ● column is gone). Headless Chromium 2026-09-24, 32 layouts: 716–780 px
  (user) and 40 px (Crow) off before, 0.0 px after.
• The Subtasks card closes like the goal card (Subtasks card has no close button: the #255 panel and its column reserve cannot be dismissed, unlike the goal card #281): an × in the same place hides the card and its 306 px
  column reserve; nothing is cancelled, and a new subtask or a jump brings it back.
• Rollover archives leave the chat rail (Every rollover adds its archived segment to the chat rail as a separate chat, titled by Crow's own note ('[The tool budget for this turn is …') #261) for the archive drawer; a Crow note is never a chat title.
• goal_set carries the whole step record (goal_set drops notes, time and tokens of re-declared steps #260): notes, seconds and tokens of a re-declared step survive.
• Comments on presence_penalty (serve: a request without temperature decodes greedy, and top_k 0 is greedy / top_k > 64 silently clamped on the device - absent fields must come from the model card row of the thinking mode crow-nest#111) no longer say crow-nest reads an absent field as 1.5
  (it is 0 since crow-nest The four behaviour changes cli/crow.py took in the split, and the gate each one stands on #91). No behaviour change: Crow sends 0.0 explicitly.

Fixed

• edit_file rewrote a whole CRLF file as LF (edit_file rewrites a whole CRLF file as LF: a one-line edit turned 40 CRLF lines into 0 (universal-newline read, newline="" write) #283). A one-line edit of a 40-line CRLF file left 0 CRLF
  (measured 2026-09-24 at 57ed521). The file is now read raw and only the matched span is replaced, in the ending
  it had: 40/0 stays 40/0, a 20/20 mixed file stays 20/20.
• The browser panel crashed in the NVIDIA driver next to GPU renders, and a folded panel still loaded pages
  (The browser panel's WebKit process segfaults in the NVIDIA driver right after each GPU render_page: 512 MiB headroom ignores Crow's own panel as a second GPU client next to serve #279). 2026-09-24: the panel's WebKitWebProcess segfaulted in libnvidia-eglcore 5–10 s after each of two GPU
  render_page captures (~560 MiB free), and after six crashes the window stopped repainting. render_page's GPU
  bound is now 1,536 MiB while the panel is open or holds a page (512 MiB otherwise); a folded or hidden panel
  unloads its page and loads nothing; after a crash the dead view is hidden and the window redrawn, and a watched
  page is reloaded once; during a local-model turn the panel renders without hardware acceleration.
• The visual gate refused the planning step (goal_step accepts 'done' on visual steps with no picture and no score: 9/9 closed over a small box in a black frame; a visual step must cite a capture from this goal and meet the judge's bar #267). "Think and plan: … verify findings, write PLAN.md" counted
  as visual because of "verify"; the step's leading phrase decides now, and the refusal names the planning
  exemption.
• An identical render_page after an edit replayed the old capture (render_page after an edit is answered from the per-turn _SEEN cache ('The result was, and still is:') with the pre-edit capture; the model learned to change width/height to get a real render #273). render_page joins
  run_command/build_bundle in NEVER_CACHED.
• render_page dropped a local page's ?query / #fragment (render_page refuses a local page with a ?query or #fragment ('no such page: index.html?view=albedo'), so the diorama prompt's required screenshot mode index.html?shot=<view>&w=<px> cannot be used and the model stored 'file:// rejects a ?query' in memory #272): index.html?view=albedo came back
  "no such page" (2026-09-24, session.json msg 36/37), and the model stored "file:// rejects a ?query" in memory.
  The file is checked without the suffix, which goes back onto a percent-encoded file:// URL (RFC 8089 for drive
  letters and UNC).
• build_bundle missed deno's esbuild and gave advice the model could not follow (build_bundle looks for deno's esbuild in ~/.cache/crow/deno and tells the model to set an env var it cannot set: 4 rounds lost and a foreign binary copied into the diorama project (refs #212) #274). It now looks in
  ~/.cache/deno, takes bundler in settings.json or --bundler PATH, and its error names the settings file and
  the node_modules/.bin/esbuild link and says an export does not reach Crow. Measured 2026-09-24: finds
  esbuild 0.25.5 in ~/.cache/deno/dl/ (was: none).
• The round that spent the tool budget was decoded and thrown away (Tool-budget round is decoded and then thrown away #278): 6 of 6 budget turns on 2026-09-24
  decoded one more round (298–4,290 tokens, 11,240 / 271 s in total). A budget turn is now 24 tool rounds + the
  answer (25 requests, was 26).
• goal_step running on the running step restarted its clock (goal_step 'running' resets the running step's clock and tokens #275): step 2's 73.5 min and its tokens vanished
  from goal.json on 2026-09-24. A repeated running keeps started and stores its note.
• The goal brake judged a turn by its last message (goal-mode brake reads a budget-spent working turn as an empty loop: 157 + 52 messages of real work cut from the history, each cut a 106k-token COLD re-prefill (~3 min) #258) and cut 157 + 52 messages of real work on 2026-09-23
  (cold prefills of 106k). A turn that ran tools is never "empty".
• A forced answer after the tool budget could be reasoning-only (Reasoning-only forced answer: blank window, goal brake cuts work #259) and the window showed nothing; the
  reasoning is surfaced now.
• A subtask stopped while its attempt fails closed "failed" (delegate: a subtask stopped while its spot is failing ends 'failed' and memos the spot dead, instead of 'interrupted' (#143 E2 contract) #242) and memoed the spot dead; it closes
  "interrupted" now.
• A retired free model's 404 counted as transient (delegate: a retired free model's HTTP 404 ('unavailable for free') is booked transient via the 'unavailable' substring and burns 1 of 3 transient attempts in every subtask #284). "unavailable for free" matched the retry pattern, so
  every subtask of the 2026-09-24 diorama wave spent 1 of 3 attempts on a spot that can never answer; it is a spot
  refusal now (memoed, next spot).
• Phone mirror fixes found on robin's iPhone (/remote: operate a running session from the phone (LAN) #249): pairing no longer holds a request open (Chrome for iOS gave
  up after ~6 s; POST /pair answers 202 and the page polls /pair/wait), a paired phone ignores a stale pairing
  code, a phone page load no longer re-recorded every note (1 "no folder" note became 32), and Safari's toolbar
  tint and scrolling stay as loaded after a drawer closes.

Known limitations

• The phone on real devices (/remote: operate a running session from the phone (LAN) #249, Phone: the 🎤 records nothing -- getUserMedia needs a secure context and the mirror is plain HTTP; record on the phone over Tailscale HTTPS and transcribe with crow_voice on the PC #290): pairing, the mirror, the HTTPS address and the microphone were used
  live on robin's iPhone (iOS 27, Safari and Chrome); the auto-hiding home-screen header and the drawer fixes were
  checked in Chromium emulation only. Android is unverified. /remote is window-only; the terminal answers
  "stage 2".
• Tailscale on Windows is unverified (/remote: operate a running session from the phone (LAN) #249 stage 5): the dialog, the serve command without sudo and
  install.ps1 -Tailscale were not run on a Windows machine, and install.ps1 -Selftest was not run for this
  release (no PowerShell on the Linux release machine).
• The browser panel's WebKit process segfaults in the NVIDIA driver right after each GPU render_page: 512 MiB headroom ignores Crow's own panel as a second GPU client next to serve #279 is covered by unit tests only (red without the fix); not yet checked live.
• render_page/fetch_url fetch URLs the model invented: 11 renders of a made-up Alibaba OSS proxy URL in step 4, error pages taken as scene frames #288 counts a tool result's hosts by URL only: a host that appears there as a bare name, and nowhere else,
  is refused.
• edit_file runs no syntax check (write/append do since #251), and every new format needs a code change: one extension→command table for all three file tools #269/write_file/append_file never parse what they wrote: 20 of 98 JS/HTML writes on 2026-09-23 did not pass node --check, and the model learnt it from renders hours later #251 need node on PATH for JS/HTML; neither installer installs node.
• check_gui_prereqs (not in CI) reports 2 of 3 prerequisites: point (ii) fails on 13 glyphs missing from
  Google Sans Code (both shipped faces, 26 problems) — the 12 of 2.5.0 (0c3e392) plus U+1F3A4 🎤, new with the phone
  microphone (Phone: the 🎤 records nothing -- getUserMedia needs a secure context and the mirror is plain HTTP; record on the phone over Tailscale HTTPS and transcribe with crow_voice on the PC #290, 4d753af; cli/crow_core.py:18037, cli/crow_gui.py:8740). Measured 2026-09-24 at 0a0304b.
• Still open from 2.5.0: the Windows installer bundle (CrowSetup.exe: installer for all three operating points #196); sampling_no_thinking's presence_penalty 1.5 (presence_penalty 1.5 in the unused no-thinking sampling row #246).

🤖 Generated with Claude Code

https://claude.ai/code/session_01DmVJHThNzKH4NWHvQffwYv