Hmbown/CodeWhale v0.9.4

on GitHub

Codewhale is the public product from Shannon Labs. The codewhale
command, npm package, and release-asset names remain lowercase technical
identifiers. The legacy npm package deepseek-tui is deprecated and
receives no further releases. Users coming from v0.8.x legacy deepseek /
deepseek-tui names should migrate with docs/REBRAND.md.

Install

Recommended — npm (one command, all three entrypoints)

npm install -g codewhale

The wrapper downloads the matched codewhale, codew, and codewhale-tui
binaries from this Release and places them in the same directory.

Docker / GHCR

docker run --rm -it \
  -e DEEPSEEK_API_KEY="$DEEPSEEK_API_KEY" \
  -v codewhale-home:/home/codewhale/.codewhale \
  ghcr.io/hmbown/codewhale:v0.9.4

The image ships the codewhale dispatcher, codew shim, and codewhale-tui runtime. The latest tag is also updated on release.

Cargo (Linux / macOS)

cargo install codewhale-cli codewhale-tui --locked

Both crates are required — codewhale-cli produces the codewhale dispatcher and codew shim, while codewhale-tui produces the interactive runtime that the dispatcher delegates to. Installing only one crate will fail at runtime with a MISSING_COMPANION_BINARY error.

Manual download — platform archives (recommended)

Each archive below contains the codewhale dispatcher, codew shim, and codewhale-tui runtime, plus an install script:

Unix (Linux / macOS):

tar xzf codewhale-<platform>.tar.gz
cd codewhale-<platform>
./install.sh

Windows:

• For the installer path, run CodeWhaleSetup.exe; it installs codewhale.exe, codew.exe, and codewhale-tui.exe under %LOCALAPPDATA%\Programs\CodeWhale\bin and adds that directory to the current-user PATH.
• Extract the archive for your machine: codewhale-windows-x64.zip or
  codewhale-windows-arm64.zip
• Run install.bat (copies to %USERPROFILE%\bin)
• Add %USERPROFILE%\bin to your PATH

The portable Windows archive skips the install script — extract and run from any directory. The NSIS installer is currently unsigned and may trigger Windows SmartScreen until a signing certificate is wired into the release pipeline.

Each platform also has bare, unarchived binaries attached below (codewhale-<platform>, codew-<platform>, and codewhale-tui-<platform>) — the npm wrapper and the in-app codewhale update download the matched runtime binaries, whereas the .tar.gz / .zip archives above are the recommended manual download and additionally bundle an install script. The legacy npm package deepseek-tui is deprecated and is not republished. For migration from v0.8.x legacy binary names, see docs/REBRAND.md.

Verify (recommended)

Download the checksum manifests from this Release and verify:

# Linux — archive bundles
sha256sum -c codewhale-bundles-sha256.txt --ignore-missing

# Linux — individual binaries
sha256sum -c codewhale-artifacts-sha256.txt --ignore-missing

# macOS
shasum -a 256 -c codewhale-bundles-sha256.txt --ignore-missing
shasum -a 256 -c codewhale-artifacts-sha256.txt --ignore-missing

What's in v0.9.4

Codewhale v0.9.4 ships the release-train harness work: the familiar Fleet
roster/setup face with a clear operator-leader and user/folder scope, a
work strip that keeps actionable agents instead of a permanent archive,
waiting policy that forbids polling without freezing independent work,
calmer tool output and session recovery, account/Workflow-search/
automation/handoff surfaces, a shorter translation-ready website, and
release-blocker fixes across permissions, DeepSeek Responses, SQLite,
File edits, terminal width, and Windows installation.

Added

• Memory maintenance: remember gains revise and retire beside the
  default append. Both name the exact note they target and both require
  the evidence for the change. Append-only memory decays — a correction
  sits behind the note it contradicts and both keep reaching the model —
  so the model can now keep its own durable notes true instead of only
  adding to them.

• An audit trail for durable state the model writes about you. Every
  in-place memory edit is journalled to memory/JOURNAL.md, and every
  continual-harness refine / remove to a JOURNAL.md beside its state,
  each with before, after, and evidence. Harness removal previously left no
  record at all even though the entry leaves state entirely, so the journal
  is now the only place its content survives.

• A first-run tip that says so: the first time Codewhale saves something
  durable it points at /memory, translated into all fifteen complete
  locale packs. This state shaped later sessions and nothing ever mentioned
  it existed.

• Sub-agent checkpoint resume: agents/followup resumes an
  interrupted_continuable child from its checkpoint into a fresh agent loop —
  new agent id, original prompt plus the prior conversation tail — when a
  runtime is attached, and otherwise keeps queue-only semantics with the
  continuation_handle returned; a second followup on the same interrupted id
  returns the existing resumed target instead of spawning a duplicate (PR #5242).

• MCP Registry discovery with Registry-first tool selection: registry_sync
  surfaces the eligible local stdio catalog as a complete model-side candidate
  set, connect-failure messages classify early-exit and usage-help output and
  point recovery at the next Registry candidate, and a bundled mcp-discovery
  skill documents the flow (PR #5238).

• Progressive fresh-context disclosure: fresh sessions ship a minimal
  constitutional kernel — ground truth, user intent and scope, truthful
  completion, guarantees in mechanism, and precedence — with procedural
  playbooks disclosed on demand, an opt-in project context pack
  (project_context_pack_enabled) counted in context reports, and load_skill
  catalogue discovery via name="list"; the measured fresh-context budget
  drops by roughly 40% (PR #5077).

• Named Fleet store v2: one self-contained TOML Fleet per configuration
  (schema = "fleet"), with scope-explicit selection (user-global default vs
  folder override), migration receipts from legacy role profiles, and atomic
  saves that refuse to clobber a different Fleet on the same slug.

• Scout replaces the user-facing "faster" control: catalog-verified fast
  siblings only, never a guessed model name; pinned Scout survives operator
  changes.

• Truthful model-picker rows: vision/tools/limits chips only when the catalog
  knows, with provider → family → exact model grouping.

• Opt-in product telemetry, off by default. A first-run notice asks once, on
  a terminal, with declining pre-selected — Enter declines. Nothing is collected
  unless both telemetry = true and a recorded "Enable" answer are present, so
  a telemetry = true written before this release stays inert: the key has been
  settable and inert for a long time, and setting it was never consent.

  An enabled session sends its batches to the first-party ingest endpoint,
  https://telemetry.codewhale.net/v1/telemetry, which is the shipped default
  for telemetry_endpoint. That is a Cloudflare Worker whose complete source is
  in this repository under telemetry-ingest/; it writes to Workers Analytics
  Engine, whose row is exactly _sample_interval, blob1–blob20, dataset,
  double1–double20, index1, and timestamp — there is no IP, country,
  or geo column, so storing one is structurally impossible rather than merely
  disabled. The handler reads two request headers, never touches the request's
  geo properties, logs nothing, and validates against a closed field set that
  rejects an entire batch carrying any unpublished key. Cloudflare's retention
  for that data is a fixed three months. Setting telemetry_endpoint = ""
  instead writes each batch to $CODEWHALE_HOME/telemetry/dryrun.jsonl and
  constructs no HTTP client at all, so you can read exactly what would have been
  sent.

  Turning it off is an answer, not a flag: it deletes the random install id,
  truncates every buffered event, and leaves a permanent tombstone that a
  session already running re-checks before it appends and before it sends. A
  failed wipe fails closed. CODEWHALE_TELEMETRY=0 is a hard floor that beats
  --telemetry true and the config key, and a value the parser cannot read
  also resolves to off. Fleet workers are hard-off. A repo-local
  .codewhale/config.toml can set neither key.

  Never collected: prompts, completions, tool arguments, diffs, file contents,
  filenames, paths, git remotes, repo or branch names, memory entries, chat
  history, credentials (not even a boolean asserting one exists), model ids,
  custom provider table names, MCP server names, error or panic message bodies,
  per-event timestamps, keystrokes, clipboard, screenshots, or location. The
  full schema is docs/TELEMETRY.md, and a test parses the
  field names out of that file and asserts set equality with the structs the
  serializer uses.

  This supersedes the roadmap's previous "no Codewhale product telemetry" entry,
  which moves from "Ruled out" to an opt-in framing. What stays ruled out:
  always-on or silent telemetry, per-keystroke or per-tool-call phone-home, and
  any third-party ad or analytics SDK in the runtime binary.

• Registered GLM-5.3 (direct Z.ai) and z-ai/glm-5.3 (OpenRouter) as
  selectable GLM routes, with their aliases (glm-5.3, glm-5-3,
  zai-glm-5.3, zai-glm-5-3). Z.ai had not released GLM-5.3 as of
  2026-08-03 — the ids are registered so they resolve to the Z.ai/OpenRouter
  routes instead of being rewritten to another vendor's model, and they will
  fail upstream until Z.ai ships the model. Metadata (context, output,
  reasoning controls) is inherited wholesale from GLM-5.2 pending official
  Z.ai release metadata; pricing is intentionally absent, and GLM-5.2 remains
  the default Z.ai model. No third-party gateway roster gained the model:
  OpenCode Zen, OpenCode Go, Alibaba Model Studio, and TelecomJS publish no
  glm-5.3 entry, so Codewhale advertises none.

• Managed Codewhale account commands (account login, status, logout, and
  keys) with browser device flow, profile- and origin-scoped secure sessions,
  refresh/revocation, redacted BYOK-vault management, and a token-free Runtime
  account receipt. Provider authentication remains separate, and cloud
  remains a compatibility alias.

• /automation operator controls to list, inspect, pause, resume, delete, and
  run durable automations. Creation remains on the approval-gated
  model-visible automation tool.

• A provider-neutral WorkflowSearchSpec authoring and freeze boundary, plus
  structured 2–16-candidate experimental search in the best-of-N Workflow
  starter. It freezes baseline, route, evidence, evaluator, gate, score, budget,
  and review policy before admission; it validates gate/scoring commands but
  does not execute or certify them itself.

• The bundled generation-9 handoff skill for compact, decision-ready
  continuation across sessions.

• Expanded terminal LaTeX rendering for aligned and matrix environments,
  cases, arrays, text/font/accent commands, brackets, symbols, and
  command-aware scripts (PR #4981).

• Exact 40-character build provenance and secure account-session capability
  receipts on /v1/runtime/info; unknown source provenance continues to fail
  closed.

• Acceptance-level Gherkin coverage locking the existing user-command
  precedence, alias shadowing, fallback, and invalid-command error contract
  (PR #4992).

• Agent Plugins v1.0.0: consume, publish, and slugify packaged sub-agent
  briefs, with an install/update/uninstall on-ramp in the TUI (PR #5182). A
  plugin bundles a prompt, posture, and routing as one shareable artifact;
  on-disk migration of the older plugin.toml scaffold is deliberately out
  of scope for this train.

• send_later: a model-callable one-shot delayed continuation tool, so the
  model can schedule a single future nudge without an operator-approved
  durable automation (PR #5138).

• /advisor: an opt-in background advisor watcher for live turns (PR #5139).

• Notification quiet mode with per-category switches and action-first copy
  (PR #5066).

• Automation scheduling forms — one-shot ONCE, five-field cron, and honest
  watcher modes — created through the approval-gated automation tool
  (PR #5183).

• Sub-agent resume_from continuation chains (PR #5142), child-result
  diff-tainting when a claimed diff is not visible to git, per-turn usage
  receipts on the exec stream-json stream, and spawn receipts that report
  the model each sub-agent actually ran on.

• Transport resilience: sub-agent exec transport retries with a 600 s
  default (PR #5210), SSE header stalls retryable instead of fatal, and
  headless turn resume after mid-stream network drops with an EX_TEMPFAIL
  exit.

• Session durability and control: a deterministic compaction continuation
  contract (PR #5064), persisting interrupted output (PR #5206), stop-word
  cancellation (PR #5207), token-counter refresh (PR #5204), deny-by-default
  approval cards (PR #5090), and the Operate completion gate (PR #5067).

• zh-Hant promoted to a full shipped locale with complete en.json parity
  (PR #5143).

• A persistent update-available chip in the header, with the startup update
  check throttled and naming the right command.

• RLM static intent extraction for code blocks (rlm_block_intent.rs)
  landed as groundwork for a future code-mode approval flow; it is not yet
  wired into the turn pipeline and ships dormant by design.

Changed

• /fleet is the familiar roster/setup face again. The operator row is the
  Fleet leader (session model); the header names the selected saved Fleet and
  whether it is user-global or folder-scoped. Named-Fleet switching lives under
  /fleet fleets (Enter selects in the row's own scope). Session route changes
  stay temporary until /fleet save, /fleet save-as, or /model save-default.

• Waiting-for-subagents directions forbid peek/status polling and sleep-as-wait,
  but allow independent work that does not depend on a child's result — the
  parent no longer freezes mid-turn with useful non-conflicting work available.

• workflow run no longer requires --fleet; a saved Fleet is an optional pin
  layer over roles + the session route.

• Homepage and getting-started copy is shorter and scannable across locales,
  with dictionary key and {brand} token parity preserved.

• Tool results now render as ordinary bounded previews with real expansion;
  storage, retention-ledger, and internal evidence language no longer leak into
  normal transcripts.

• Prose wrapping, goal state, modal questions, composer-tail behavior, and
  ambient motion now follow one deterministic interface contract across narrow
  terminals and fast streams.

• Scout and reviewer Fleet roles gain network access and the bounded
  verification surface for real reconnaissance while retaining the no-write,
  no-raw-shell security floor.

• Workflow runs may describe up to 1,000 tasks while admitting at most 16 live
  tasks at once through the host concurrency gate. Tournament ordering now
  supports explicit score-first selection while retaining its cost-first
  default.

• Runtime permission compatibility inputs resolve to one live
  permission_posture. Auto-Review can proceed without approval or structured
  question modals, unresolved holds fail closed, and a call planned under stale
  authority is retried after a posture change (PR #5025).

• Duplicate and drifting per-turn metadata has been removed in favor of
  runtime-owned authority, and large inline account and skill tests now live in
  owned test seams.

• Pinned Ratatui to 0.30.0 and ratatui-core to 0.1.0. ratatui-core 0.1.1+
  makes Terminal::clear() issue a blocking cursor-position report that
  raced the TUI input loop and could kill first launch; both pins are
  load-bearing, because 0.30.0 declares ratatui-core ^0.1 and would
  otherwise resolve forward on its own (PR #5192 by @bistack; upstream
  ratatui/ratatui#2640).

• Updated globset to 0.4.19, clap-complete to 4.6.8,
  futures-util to 0.3.33, libc to 0.2.189, actions/stale to 11.0.0, and
  docker/login-action to 4.5.2. The locked graph also includes the
  event-listener 5.4.2 fix for RUSTSEC-2026-0221.

• The progress surface now speaks plainly everywhere: the last user-visible
  "Work update is pending" notices say "To-do list", the tool constructor and
  the docs name todo_write as the single canonical progress tool, and
  work_update, TodoWrite, and todo stay registered as hidden
  compatibility aliases so saved transcripts keep replaying.

• Sub-agent and agents/wait waits stay short by default and by cap:
  blocking waits default to 30 s and refuse to block past 120 s, because a
  blocked wait deafens the session to typed input and settled children
  already report back as <codewhale:subagent.done> sentinels.

• Bash action=wait honors timeout_secs (seconds) and bare timeout
  (milliseconds) alongside canonical timeout_ms, and block as an alias
  for wait, so a habit formed on other wait tools gets the duration it
  asked for instead of silently falling back to the 30 s default; the result
  metadata reports the real wait_timeout_ms applied.

Fixed

• The memory journal is no longer indexed as memory. It is Markdown in the
  memory tree, so the source walk collected it and every retired note
  re-entered the searchable set under its before: line — putting the
  exact facts a revision had just removed back into the prompt.

• memory_path pointed at an already-native store no longer derives a
  second store nested inside it, which silently wrote somewhere other than
  the file the user named.

• muse and muse-spark resolved to muse-spark-1.1 in the agent
  registry while config had defaulted to muse-spark-1.2, so the CLI and
  app-server routed those aliases somewhere the configured default never
  pointed. The registry now carries 1.2 and the contributor variant.

• An explicit type=builder (or its implementer alias) plus
  write_authority=read_only now fails closed at spawn instead of launching a
  labeled write role that silently had only recon tools and then self-BLOCKED
  after burning a turn (#5123). The check is deliberately narrow, because two
  neighbouring combinations are legitimate and stay legal:

  • type=worker + read_only — worker is the unnamed default (it renders as
    general) and takes its capability from authority, not from its name, so a
    read-only worker is an ordinary general-purpose child. Worker, scout,
    reviewer, and verifier remain the four canonical read-only Fleet roles.
  • any role + read_only — role is an identity for roster resolution, not
    a capability claim, so an acceptance Workflow can still resolve
    implementer to its saved profile while scoping that child to verification.

  Callers that spelled a read-only narrowing as type: "implementer" should
  move it to role: "implementer".

• User-global credentials survive an explicit workspace CODEWHALE_CONFIG_PATH
  that selects a route with no local key — readiness probes the user-global
  provider table before concluding a key is missing.

• Sub-agent token figures on the work bar accumulate input+output (the same
  total the worker budget uses) instead of completion tokens alone; elapsed
  time still freezes when the child settles.

• Live work-bar rows for sub-agents show how many to-dos they still have
  left (N left) when the child's own ledger has unsettled items — never a
  fabricated zero when no list exists.

• Surfaces no longer claim an OS sandbox on platforms that cannot enforce one.
  The policy resolver takes no platform input, so on default Linux (bubblewrap
  is opt-in) and on all Windows the header chip read files: workspace and
  /status read sandbox workspace-write while nothing was restricted. Both
  now resolve the real backend and say (unenforced).

• tool_category hook conditions matched only retired tool names, so a
  category = "shell" deny hook — the security control docs/HOOKS.md
  documents — silently never fired. Categories now use the registered names,
  and multi-action tools classify by action.

• A Retry-After header of -5, nan, or 1e300 crashed the request task
  (Duration::from_secs_f64 panics on a negative). Parsing is now guarded and
  bounded to one hour.

• Bearer tokens no longer leak into operator-visible receipts. Authorization: Bearer <jwt> split into two tokens and the JWT matched no redaction rule;
  prefix matching was also case-sensitive, so SK-live-… survived.

• prune_older_than destroyed the NEWEST rollback snapshots and kept the old
  ones — on every boot, for any workspace with snapshots spanning the retention
  window. Both prune paths now share one orphan-chain rebuild and preserve each
  survivor's real timestamp.

• An absolute or relative command path no longer defeats every execpolicy deny
  rule (/bin/rm -rf / did not match a rm -rf / rule), and a typed Allow
  rule no longer auto-approves a chained suffix such as git log ; curl … | sh.

• Wrong types on File read range params and Bash stdin/cwd/task_id are now
  errors instead of silent defaults — a start_line:"1200" string used to
  return the head of the file, and a non-string stdin ran the command with no
  stdin and reported success.

• Multibyte tool ids no longer panic the context inspector, wide (CJK) text no
  longer overflows the decision card, and a hostname like 127.evil.example.com
  is no longer treated as loopback.

• Refusals name calls the model can actually make (rlm action='open' rather
  than a retired rlm_open; Bash rather than exec_shell).

• Sub-agent dispatch no longer aborts the process. The Tokio runtime was built
  by #[tokio::main], leaving every worker thread on the 2 MiB default while
  only the owner thread received the explicit 16 MiB stack — and the engine runs
  on a worker. A debug-build agent dispatch exceeded that stack and raised
  SIGABRT, which is not a panic and so could not be caught; the process died
  mid-spawn with no child request ever issued. Release builds were unaffected.

• Fleet profiles that pin a provider no longer leak a bare model id onto the
  session route. model_overrides exported each role's model while dropping its
  provider, so a scout pinned to another provider's model was dispatched against
  the active client and denied at the wire — visible as an instant auth failure
  on the first sub-agent of a fan-out.

• The rail's Pinned panel no longer spends four rows saying "No active work".
  An empty panel now collapses like the Tasks panel always has, and the settings
  migration no longer folds the default sidebar_focus = "auto" into a pinned
  always-on strip, which had silently handed that panel to every user who had a
  settings file at all. (An empty panel collapses; a panel holding settled
  to-dos or finished workers is not empty — see the standing-register entry
  below.)

• The work bar keeps settled to-dos and an honest Subagents header, while
  completed/cancelled workers collapse out of the Top strip so fan-outs do not
  permanently eat the transcript. Failed or interrupted workers stay visible
  (they still need attention). Settled agents remain reachable through the
  Agents panel and catalog. To-do rows say their state in words (pending /
  in progress / completed / cancelled), and sub-agent rows carry type,
  objective, elapsed, and input+output tokens. Every work row is a door in
  every rail panel and placement: click and Enter open the row's world
  (work inspector / agent details — finished agents included) instead of
  doing nothing. A click after the detail pager closed itself reopens the
  detail rather than being swallowed by a stale toggle.

• The rail strip yields its rows to the transcript when the terminal cannot
  seat both, so the idle ocean survives at 24 rows instead of being evicted.

• code_execution and js_execution no longer describe themselves to the model
  as sandboxed. Both are ordinary local subprocesses with no seccomp, jail, or
  container (PR #5221 by @h3c-hexin and @asto18089).

• Model Studio reasoning controls now fail closed on the host rather than on the
  provider enum, so a custom base_url no longer receives Alibaba-specific
  enable_thinking fields, and qwen3.8-max is no longer sent a thinking
  switch it does not accept (PR #5233 by @Inference1, closing #5203).

• config.example.toml no longer claims Shift+Tab cycles the reasoning tier.
  Shift+Tab cycles the permission posture; Ctrl+T cycles reasoning
  (found by @vFONGv, PR #5229).

• Alibaba Model Studio reasoning controls are now route- and model-scoped
  instead of provider-wide (#5203, harvested from #5233 by
  @Inference1). Codewhale sends
  enable_thinking / preserve_thinking / reasoning_effort only when the
  configured base_url is a verified Alibaba Chat Completions host, so
  pointing a modelstudio-* provider ID at a custom gateway no longer injects
  DashScope's dialect into it. qwen3.8-max and qwen3.8-max-preview are
  thinking-only and no longer receive an enable_thinking: false they cannot
  honor; preserve_thinking is sent for the models documented to accept it, so
  their reasoning trace survives into the next turn; and deepseek-v4* /
  glm-5.x map the reasoning tier onto the documented high / max ladder.

• xAI device login now recovers from a config that points at a missing
  Codewhale-owned credential generation instead of failing every attempt
  with a generic activation error, and finalize failures report the full
  error chain (#5032).

• API keys saved to the secret store no longer read as unconfigured for
  providers that are not currently active; a configured Kimi/Moonshot key
  survives provider switches and restarts without re-entry (#5033).

• Switching to the Codex provider with no saved model now lands on the live
  roster's flagship model instead of a stale static default (#5034).

• Worktree-isolated Fleet builders no longer contend on the per-workspace
  delegated-coordination lock, and a failed lock acquisition is retried on
  use instead of being memoized for the life of the process (#5036).

• Fleet dispatch now rebinds the child client when the resolved profile
  model requires a different wire protocol (DeepSeek flash on Responses),
  instead of failing deterministically on the worker's first request
  (#5042).

• DeepSeek Responses now sends reasoning.effort: "none" for the Off tier,
  shows a truthful notice instead of silently discarding server-side
  web_search_call items, and parses cache-hit, cache-miss, cache-write, and
  pricing telemetry while retaining the OpenAI-style nested fallback.

• File edits now explain no-op and missing-search failures, reject newly
  unbalanced C/C++ preprocessor replacements, handle the reported
  CRLF/non-ASCII cases, and safely relocate stale unified-diff hunks only when
  whole-file context is unique (PRs #5008 and #5030).

• Circled digits, enclosed alphanumerics, and keycap graphemes use consistent
  two-column measurement in Codewhale, Ratatui, and CJK terminals, preventing
  missing-character and phantom-space corruption (PR #5001).

• SQLite connections install their busy timeout before locking setup and avoid
  rewriting persistent WAL mode on every open, removing the concurrent-open
  release-gate failure.

• The Windows installer preserves long current-user PATH values, their
  registry type, and unrelated entries across install and uninstall (PR #5006).

• Provider configuration no longer contains user-reachable panic paths when
  metadata or prior credential state is missing.

• Resuming a session restores composer text only from a same-session persisted
  draft; submitted prompts and internal background-runtime envelopes remain in
  history instead of appearing in the composer (PR #5029).

• Shared CI now handles bot-authored issue-link checks, provisions cargo-deny's
  toolchain, and fetches the locked test graph before offline runtime-budget
  validation.

• Re-quote each linker argument in the Windows OpenHarmony clang launcher so a
  spaced SDK path (e.g. the default D:\DevEco Studio\... install) keeps its
  --sysroot intact through the final Rust link, and extend the no-SDK release
  guard to keep the re-quoting contract (PR #5095).

• The shell tool reports the real elapsed wait time in its result content
  instead of echoing the requested timeout (PR #5240).

• Transcript wheel scrolling under iTerm2: xterm alternate-scroll (DECSET
  1007) now stays off while mouse capture is active, so wheel events arrive as
  mouse events instead of being converted into arrow keys (#5223, PR #5234).

• A stalled model stream no longer ends the turn as Completed over a
  frozen reasoning block: a mid-stream chunk-timeout now counts toward the
  stream-error budget, so a stall with nothing streamed retries the request
  transparently, and a stall that exhausts the retry budget fails the turn
  with the real reason instead of reporting success.

• A finished background shell task now wakes the engine even when no goal is
  active: the idle loop starts an ordinary runtime turn so the completion
  reaches the model immediately instead of sitting unclaimed until the user
  types (a dead provider route claims the completion once and reports where
  the output lives instead of re-arming the same error every tick).

• Sub-agent final reports that exceed the summary budget are now spilled to
  a session artifact, and the truncation footer names the
  retrieve_tool_result ref for the elided middle instead of telling the
  model the bytes are unrecoverable; write failures degrade to the honest
  no-ref footer.

• An interactive mid-stream network drop after partial output no longer fails
  the turn: the partial reply is preserved as a committed assistant message,
  a runtime continuation message is appended, and the request is re-issued
  bounded by the stream-retry budget.

• Large pasted input is no longer sent to the model twice as inline text and
  as a backup .md paste file; the submitted message now carries only the
  @-mention so the model reads the file once.

• A builder sub-agent can run ordinary shell writes again. Write claims
  outlive the agents that register them, so a workspace accumulated one per
  builder that ever ran — six completed agents left four standing claims in
  testing — and the shared-checkout gate counted those long-finished children
  as live contenders. Every later builder was refused Bash writes with
  "cannot prove a bounded file target" and pushed toward worktree isolation,
  which puts the work in a checkout the operator never looks at. The gate now
  asks the question it meant to ask: is another running child writing in this
  shared checkout. Concurrent writers are still gated; a lone builder writes in
  the workspace you are actually watching.

• Ctrl-C during the first moments of startup no longer kills Codewhale
  outright. The terminating-signal handlers were registered inside the task
  that waits on them, and a spawned task does not run until the scheduler
  first polls it, so a SIGINT arriving in that window hit the default
  disposition — the process died with no exit code, no terminal restore, and
  no session record. The handlers are now installed synchronously, before
  the telemetry notice and before arming, so the window is closed.

• The documented tool list on the docs site named update_plan and
  work_update as coordination tools. Neither is callable by the model —
  update_plan replays older Plan artifacts and work_update is a hidden
  compatibility alias — so the page listed two tools a reader cannot use and
  omitted todo_write, the one they can.

Security

• Bumped nanoid past GHSA-2v37-7h3g-55p8 (a custom generator given size
  zero could loop indefinitely), restoring a zero-advisory npm audit for
  the website.

Removed

• The default model-facing SlopLedger implementation, its storage-oriented
  transcript language, and the /debt, /cleanup, /slop, and /canzha
  command surface.

Contributors

• Sh1Zuku (@SparkofSpike) contributed
  LaTeX rendering in PR #4981, completed circled-digit/keycap width handling in
  PR #5001, and delivered actionable File-edit recovery in PR #5008; for this
  train he resumed interrupted sub-agents from checkpoints in PR #5242,
  surfaced real shell wait elapsed time in PR #5240, and kept alternate-scroll
  off while mouse capture is active in PR #5234.
• XhesicaFrost (@XhesicaFrost) fixed long
  Windows user-PATH preservation in PR #5006.
• Paulo Aboim Pinto (@aboimpinto) added the
  user-command dispatch acceptance contract in PR #4992.
• DracheTek (@DracheTek) provided the
  multilingual, CRLF-heavy File-edit failure report in issue #5003.
• An Ziwu (@MuRongMoQing) reported the
  Windows PATH-overwrite defect in issue #4685.
• shenjackyuanjie (@shenjackyuanjie)
  fixed the Windows OpenHarmony linker re-quoting for spaced SDK paths in
  PR #5095.
• bistack (@bistack) contributed MCP Registry
  discovery with Registry-first tool selection in PR #5238.
• vFONGv (@vFONGv) wrote the zh-CN Windows
  beginner guide with screenshots in PR #5229, harvested after its base branch
  was accidentally deleted during maintainer cleanup.
• mky (@mky) fixed the FreeBSD build (PR #5254, rquickjs bindgen on FreeBSD).
• cacdcaecawae (@cacdcaecawae) contributed embedder-owned sub-agent state roots (PR #5252).

See CHANGELOG.md for full notes and docs/CHANGELOG_ARCHIVE.md for older releases.