npm - ur-agent - Versions diffs - 1.68.16 → 1.68.18 - Mend

ur-agent 1.68.16 → 1.68.18

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (10) hide show

package/CHANGELOG.md +246 -18
package/dist/cli.js +1669 -896
package/docs/VALIDATION.md +1 -1
package/documentation/index.html +1 -1
package/extensions/jetbrains-ur/build.gradle.kts +1 -1
package/extensions/vscode-ur-inline-diffs/package.json +1 -1
package/package.json +1 -1
package/technical/01-architecture.md +3 -1
package/technical/04-tools.md +11 -0
package/technical/README.md +1 -1

package/CHANGELOG.md CHANGED Viewed

@@ -1,7 +1,53 @@
 # Changelog
+## 1.68.18
+- The shell is presentational and imports only ink and useTerminalSize. A first
+  version read live state itself (settings, provider registry, app state, git),
+  and importing it from REPL changed REPL's module graph enough to flip a
+  command from hidden to visible — caught by `commandRegistryIntegrity`, which
+  passed with the REPL edit stashed and failed with it applied. Loading a module
+  is not free here; some register on load. Session values now arrive as props
+  from a caller that already holds them.
+- Reconstructed three more stub type files, clearing 177 further errors:
+  `keybindings/types.ts` (six empty interfaces), `ink/cursor.ts`, and
+  `commands/plugin/unifiedTypes.ts`. Same defect as the MCP types in 1.68.14 —
+  an empty interface means "has no members" to TypeScript, not "shape unknown",
+  so every property access on one was an error and the consuming files carried
+  `@ts-nocheck` as a result.
+- `KeybindingBlock.bindings` is keyed by chord string, not an array. A first
+  attempt typed it `ParsedBinding[]`, which made every entry in the default
+  binding table an error — the declaration form (`'ctrl+d': 'app:exit'`) and the
+  parsed form (chord resolved into keystrokes) are different shapes, and
+  conflating them broke a file that had been fine.
+- `KeybindingAction` is a string rather than a union of known actions:
+  `defaultBindings.ts` assembles entries conditionally from feature flags, so a
+  value missing from a union would turn a working binding into an error.
+- Suppression list: 140 -> 132. Eight keybinding and ink files came off.
+- Totals for the sweep so far: 223 -> 132 files, 868 -> 563 errors, and almost
+  none of it file-by-file. Four stub files accounted for 258 errors on their
+  own.
+## 1.68.17
+- Fixed false `TaskListRequired` failures for observational Bash capability
+  checks. Task tracking now has a classification distinct from permission
+  auto-approval, so an executable probe can remain permission- and
+  sandbox-sensitive without being mislabeled as a workspace change.
+- The Bash classification is category-based rather than tailored to one
+  module: known reads, exact help/version checks, command-presence checks, and
+  import-only Python probes for any module can run without reopening a
+  completed task. Arbitrary interpreter code, output writes, background
+  execution, sandbox overrides, and unknown commands remain task-gated.
+- Both the initial call and its final post-permission input are classified.
+  Hooks or permission handlers therefore cannot rewrite an observational probe
+  into an untracked mutation.
 ## 1.68.16
 - First pieces of the UR Nexus visual identity, as terminal primitives rather
   than artwork: `constants/urPalette.ts` (obsidian/bronze/electrum/lapis tokens
   in true-colour, ANSI-256 and monochrome tiers, with depth detection),
@@ -27,27 +73,37 @@
 ## 1.68.15
-- Reconstructed three more stub type files, clearing 177 further errors:
-  `keybindings/types.ts` (six empty interfaces), `ink/cursor.ts`, and
-  `commands/plugin/unifiedTypes.ts`. Same defect as the MCP types in 1.68.14 —
-  an empty interface means "has no members" to TypeScript, not "shape unknown",
-  so every property access on one was an error and the consuming files carried
-  `@ts-nocheck` as a result.
-- `KeybindingBlock.bindings` is keyed by chord string, not an array. A first
-  attempt typed it `ParsedBinding[]`, which made every entry in the default
-  binding table an error — the declaration form (`'ctrl+d': 'app:exit'`) and the
-  parsed form (chord resolved into keystrokes) are different shapes, and
-  conflating them broke a file that had been fine.
-- `KeybindingAction` is a string rather than a union of known actions:
-  `defaultBindings.ts` assembles entries conditionally from feature flags, so a
-  value missing from a union would turn a working binding into an error.
-- Suppression list: 140 -> 132. Eight keybinding and ink files came off.
-- Totals for the sweep so far: 223 -> 132 files, 868 -> 563 errors, and almost
-  none of it file-by-file. Four stub files accounted for 258 errors on their
-  own.
+- Redesigned the fixed top shell and bottom status area. The full-width bordered
+  banner is replaced by a three-column command deck: identity and version left,
+  workspace and branch centre, clock right, with a single graphite separator
+  beneath. The bright orange frame is gone, bronze is reserved for UR branding
+  and structural accents, and lapis blue carries focus, selection and context
+  usage. The house artwork is imported unchanged, character for character; the
+  UR wordmark is kept but rendered at normal weight rather than as the headline.
+- The footer is now a two-row rail with no surrounding box. Items declare a
+  priority, two widths and three renderings, so the rail degrades before it
+  drops and never wraps — verified at 80x24, 100x30, 120x40 and 160x50.
+  Mode and context usage are the last things to go.
+- System metrics render vertically with fixed-width labels and right-aligned
+  values, so the column does not shift when 9% becomes 71%. Miniature bars
+  appear only when the column affords them. CPU is sampled on a timer rather
+  than read per render: os.cpus() reports cumulative ticks, so a single reading
+  carries no rate, and the first reading reports `--` instead of a fabricated 0%.
+- Wired by composing around FullscreenLayout rather than adding a slot to it.
+  That file is React-Compiler output with hand-numbered memo slots and
+  @ts-nocheck; inserting a slot means renumbering the sequence, and an
+  off-by-one there produces cached JSX rendered against changed props — a fault
+  that neither throws nor typechecks. REPL changes by ten lines.
+- The deck is fullscreen-only. Inline mode writes to native scrollback, where
+  there is no fixed viewport and a "fixed" region would scroll away with
+  everything else. When the terminal is too short for deck, rail and a few
+  lines of conversation, the shell steps aside and renders content alone rather
+  than letting Ink silently shrink the fixed regions.
 ## 1.68.14
 - Gave the MCP settings types their real shapes. `components/mcp/types.ts`
   declared seven **empty** interfaces under a "Stub: not included in leaked
   source" comment. An empty interface does not mean "unknown shape" to
@@ -66,6 +122,7 @@
 ## 1.68.13
 - Eight more files came off `@ts-nocheck` (149 -> 141) from a single fix. The
   compiled signature of `useRegisterOverlay(id, t0)` declared two required
   parameters while its own first statement reads
@@ -79,6 +136,7 @@
 ## 1.68.12
 - The Ollama request path now reports where a request's bytes actually go —
   tool definitions, system prompt, conversation — once per session in the debug
   log. The fixed cost of tools plus system prompt is what decides whether a
@@ -93,6 +151,7 @@
 ## 1.68.11
 - Investigated turning the task-list gate advisory by default and did not do
   it. The friction it causes is real, but the gate is also the final
   revalidation before a tool executes, and defaulting it off removes two
@@ -113,6 +172,7 @@
 ## 1.68.10
 - `ToolSearchTool` no longer registers on runtimes where it cannot work. Its
   purpose is fetching schemas for tools whose definitions were deferred, and
   deferral needs the runtime to expand `tool_reference` blocks — an
@@ -130,6 +190,7 @@
 ## 1.68.9
 - Narrowed that allowlist entry from `ur.com` to `ur.com/docs`. Replacing a
   docs-only host with a bare domain widened what WebFetch retrieves without
   asking, on a list whose own header warns that broad entries are dangerous
@@ -140,6 +201,7 @@
 ## 1.68.8
 - The bundled `ur-guide` agent no longer answers UR questions out of another
   product's documentation. It instructed the model to fetch
   `https://docs.claude.com/llms.txt` and present it to users as "UR SDK docs"
@@ -163,6 +225,7 @@
 ## 1.68.7
 - Corrected the ur.ai -> ur.com replacement, which had rewritten identifiers as
   well as URLs. `'ur.ai'` served two roles in this codebase: a domain in links
   such as `https://ur.ai/settings/billing`, and a discriminant value in
@@ -176,6 +239,7 @@
 ## 1.68.6
 - An oversized Ollama request now recovers instead of only explaining itself.
   1.68.3 reported "this request was 19.6 MB" clearly and then left the user to
   run /compact by hand. UR already had both halves of the fix — `isMediaSizeError`
@@ -195,6 +259,7 @@
 ## 1.68.5
 - Replaced every `ur.ai` reference in `src/` with `ur.com` (202 sites, 0 left).
   Note what these are: most are upstream URL structures carried into the fork —
   `/settings/billing`, `/settings/connectors`, `/upgrade/max`, `/chrome`,
@@ -210,6 +275,7 @@
 ## 1.68.4
 - The status line now reports subagents running in the current turn:
   `agents: 2 running`, separate from the existing background `tasks:` count.
   `isBackgroundTask()` excludes foreground entries on purpose — it was narrowed
@@ -227,6 +293,7 @@
 ## 1.68.3
 - `Ollama request failed (400): http: request body too large` now explains
   itself. That string comes from Go's `net/http` MaxBytesReader rejecting the
   payload on **byte size**, which is a different limit from the model's context
@@ -244,6 +311,7 @@
 ## 1.68.2
 - **Security: explicit file deny rules were not enforced.** `matchingRuleForInput`
   resolved which permission rule matched a path by reading `igResult.rule.pattern`
   from `ignore().test()`. That property does not exist — `TestResult` is
@@ -268,6 +336,7 @@
 ## 1.68.1
 - A detected prompt-injection attempt is now reported to the user instead of
   being refused in silence. Consolidating the scattered prompt guidance into the
   execution contract was a genuine improvement, but one clause did not survive:
@@ -280,6 +349,7 @@
 ## 1.68.0
 - Removed `@ts-nocheck` from 73 files, putting 21,503 previously unchecked lines
   under `tsc`. Every one of those files produced **zero** errors once the
   suppression was lifted — they were not suppressed because they were broken,
@@ -344,6 +414,7 @@
 ## 1.66.2
 - A long session on Ollama now says when it has run out of context instead of
   quietly getting worse. Ollama truncates an oversized prompt from the front
   rather than returning an error, and the front of the prompt is the system
@@ -359,6 +430,7 @@
 ## 1.66.1
 - Corrected the warning-state result used by reactive compaction and context
   collapse when a custom proactive threshold is configured. Their
   effective-window override can no longer report that the separate proactive
@@ -366,6 +438,7 @@
 ## 1.66.0
 - Unified proactive compaction around one model-aware threshold and one live
   token estimator. Small context windows now retain positive warning, error,
   and trigger thresholds; configured 50–95% thresholds are honored; prompt
@@ -397,6 +470,7 @@
 ## 1.65.14
 - Restored the proactive task-first behavior that was present in v1.65.0 and
   removed a prompt/gate contradiction introduced later. For any non-trivial
   state change, including a feature-rich single-file build, the model is now
@@ -424,6 +498,7 @@
 ## 1.65.13
 - Recovered otherwise valid `AskUserQuestion` calls whose UI header exceeds
   the 12-character chip width. Native structured calls, bare/wrapped JSON, and
   explicit-choice recovery now compact only that bounded presentation field;
@@ -439,6 +514,7 @@
 ## 1.65.12
 - Forced explicit weak-model decision menus through the real
   `AskUserQuestion` UI. A provider-neutral end-turn guard accepts only one
   canonical Ask object from explicit reasoning or one complete rigid Markdown
@@ -460,6 +536,7 @@
 ## 1.65.11
 - Hardened `AskUserQuestion` for weaker models without inventing user intent.
   Its model schema now exposes only a strict 1–4-question request shape with
   2–8 nested choices, compact validation explains malformed flat arrays, and
@@ -487,6 +564,7 @@
 ## 1.65.10
 - Reworked approved-plan execution into a capability-aware task graph. Plans
   now split separately completable outcomes, keep genuinely atomic work whole,
   express review-to-fix and verification dependencies, and launch only ready,
@@ -508,6 +586,7 @@
 ## 1.65.9
 - Fixed the plan-file/task-list deadlock. While plan mode is active, the exact
   normalized session plan file can be written before actionable tasks exist;
   ordinary workspace mutations still require `TaskCreate`/`TodoWrite`, and
@@ -524,6 +603,7 @@
 ## 1.65.8
 - Fixed provider API-key entry in `/model`: the masked input now uses the
   available terminal width, explicit cursor/focus state, and a one-line secret
   viewport instead of rendering one masked character per row. Narrow and
@@ -531,6 +611,7 @@
 ## 1.65.7
 - Audited all 14 technical chapters against the shipped source and generated
   CLI. The manual now distinguishes public runtime behavior from compile-time
   flags, compatibility-only state, and source helpers; it is included in the
@@ -583,6 +664,7 @@
 ## 1.65.6
 - Remote sessions no longer reconnect forever when the server repeatedly
   accepts a WebSocket upgrade and then reports that the session is missing.
   Repeated `connect()` calls reuse the live client, malformed permission and
@@ -643,6 +725,7 @@
 ## 1.65.5
 - Syntax highlighting works again — in assistant messages, code previews and
   question dialogs. `cli-highlight` was imported by four rendering surfaces and
   declared in `package.json` by none of them, so the import threw on every run
@@ -669,6 +752,7 @@
   could leave a stale value on screen after a switch.
 ## 1.65.4
 - A tool call that keeps failing identically is now stopped. A 4B model refused
   once by the task-list gate answered by emitting `Write` with no arguments
   repeatedly, and nothing intervened — the trajectory grader names this pattern
@@ -684,6 +768,7 @@
 ## 1.65.3
 - Removed two tips for things that do not exist: `/mobile to use UR from the
   UR app on your phone` (no such command, no such app) and a pointer to
   `ur.ai/web` (no DNS record, same dead domain class as the `ur.com` links
@@ -701,6 +786,7 @@
 ## 1.65.2
 - Fixed `--discover-ollama` having no effect on model discovery or requests.
   `getOllamaBaseUrl` resolves the session host correctly, but three callers
   read a persisted `provider.baseUrl` *before* consulting it, so a value
@@ -739,6 +825,7 @@
 ## 1.65.1
 - Fixed `ur selftest run` reporting 0/5 anywhere but the UR repo. The drill
   runner spawned `./bin/ur.js`, a path relative to the current directory, so
   every drill failed instantly with an empty detail — which reads as five
@@ -752,6 +839,7 @@
 ## 1.65.0
 - The release gate now asks the registry whether the packed dependency ranges
   can actually be installed. This is the check that was missing when 1.61.2
   through 1.64.0 shipped uninstallable: the tarball built, the CLI started,
@@ -767,6 +855,7 @@
 ## 1.64.1
 - Fixed `npm install -g ur-agent` failing with
   `No matching version found for playwright-core@^1.64.0`. Releases 1.61.2,
   1.62.0, 1.63.0 and 1.64.0 were uninstallable: the version bump was a
@@ -784,6 +873,7 @@
 ## 1.64.0
 - Tool-result pruning now announces itself. It changed context silently, so
   there was no way to confirm it fired or to attribute a missing detail to it —
   the same defect memory suggestions had when they went to stderr. A prune now
@@ -808,6 +898,7 @@
 ## 1.63.0
 - Added size-triggered pruning of superseded tool results
   (`context.pruneToolResults`). UR already had the clearing machinery, but
   nothing external could reach it: cached microcompact is internal-only and
@@ -827,6 +918,7 @@
 ## 1.62.0
 - `ur agent-inspect --costs` now labels each row with what the agent was
   actually doing. A real 62-agent fan-out reported opaque hex ids, so you could
   see that one agent burned 810k input tokens — 14% of the session — without
@@ -838,6 +930,7 @@
 ## 1.61.2
 - Fixed `ur agent-inspect --costs` reporting nothing, always. It resolved the
   *live* session, but every `ur` invocation mints a new session id, so bare it
   pointed at a session created milliseconds earlier that had by definition
@@ -851,6 +944,7 @@
 ## 1.61.1
 - Fixed secondary model queries failing on any Ollama setup whose session model
   is not `qwen2.5-coder:7b`. `getSmallFastModel()` fell back to the compiled
   default when auto-routing was off or no model list had been discovered, so a
@@ -867,6 +961,7 @@
 ## 1.61.0
 - Added `ur selftest`, end-to-end drills for the gap that produced every
   serious defect in recent releases: a module that is correct while something
   between it and the user is not — a wire format, a CLI registration, an exit
@@ -886,6 +981,7 @@
 ## 1.60.1
 - `ur memory-integrity verify` no longer reports an empty or missing store as
   verified. It printed "verified — 0 file(s) match the recorded digests" for an
   empty directory, and that identical reassurance would have appeared for a
@@ -901,6 +997,7 @@
 ## 1.60.0
 - Added `ur memory-integrity`, tamper-evidence for the file-backed memory
   stores. Project task memory was hash-chained and could prove tampering; the
   auto-memory and team-memory directories had nothing, and their contents are
@@ -928,6 +1025,7 @@
 ## 1.59.0
 - Added `ur sources`, a claim-to-source ledger. `wrapUntrusted` already stamped
   every untrusted block with a nonce and a source label, but discarded both the
   moment the block reached the model, so there was no way to audit what web or
@@ -954,6 +1052,7 @@
 ## 1.58.1
 - Unified vision-capability detection behind
   `src/utils/model/visionCapability.ts`. Three implementations disagreed: the
   Ollama adapter's `modelCapabilityEnabled` returned `has(x) ?? true`, so a
@@ -976,6 +1075,7 @@
 ## 1.58.0
 - Added per-agent cost and token attribution: `ur agent-inspect --costs`.
   `stats.ts` already read every `{sessionId}/subagents/agent-{agentId}.jsonl`
   transcript, but only to fold those tokens into a single total, so a fan-out
@@ -990,6 +1090,7 @@
 ## 1.57.5
 - Fixed the release gate failing on `repoEditImports`. The gate runs
   `bun test --timeout 120000`, but a per-test budget silently overrides that
   global, and this test declared 15s while its body — which builds a
@@ -1001,6 +1102,7 @@
 ## 1.57.4
 - Fixed slash command arguments being silently truncated. `parseArguments`
   kept only the string tokens shell-quote returned, but shell-quote classifies
   `left?` and `src/*.ts` as globs and `&`, `>`, `(` as operators — so
@@ -1024,6 +1126,7 @@
 ## 1.57.3
 - Stopped a false diagnosis on failed tool calls. When a tool call failed
   schema validation, UR appended "this tool's schema was not sent to the API"
   and told the model to load it via `ToolSearch`. Both claims were wrong on
@@ -1040,6 +1143,7 @@
 ## 1.57.2
 - Fixed the Ollama adapter discarding images returned by tools. A tool result
   containing an image was flattened with `contentBlockToText`, which renders an
   image block as the literal string `[Image output omitted]` — so a `Computer`
@@ -1054,6 +1158,7 @@
 ## 1.57.1
 - Fixed the `Computer` tool returning a byte count instead of the screenshot.
   `mapToolResultToToolResultBlockParam` dropped the captured image, so the
   model saw only "Captured 5164460 bytes" and had to ask the user where to save
@@ -1065,6 +1170,7 @@
 ## 1.57.0
 - Connected four features that were built, tested and then left unreachable.
   Each had passing unit tests while contributing nothing to a real session.
 - `wrapUntrusted()` now runs on WebFetch and WebSearch results, at the
@@ -1093,6 +1199,7 @@
 ## 1.56.1
 - Documented the 1.52.0–1.56.0 features, which had reached `technical/03` as
   command rows but nowhere else. Doc 09 now covers the fan-out limits and how
   to run several workers at once; doc 12 covers the prompt-injection module and
@@ -1103,6 +1210,7 @@
 ## 1.56.0
 - Added a subagent fan-out governor. Agents could spawn agents with no depth or
   concurrency bound, and `/crew`, `/arena`, `/bg fanout` and `/pattern` all
   spawn several at once, so a single prompt could expand into an unbounded tree
@@ -1131,6 +1239,7 @@
 ## 1.55.0
 - Narrowed `APIProvider` to the two values `getAPIProvider()` can actually
   return, `'foundry' | 'ollama'`, and let the compiler find every dead branch.
   It surfaced 66 errors across 25 files; all are now resolved and the
@@ -1157,6 +1266,7 @@
 ## 1.54.1
 - Fixed `/speak`, `/computer`, `/memory-suggest`, `/import-session` and
   `/permission-profile` being unreachable from the shell. They were registered
   as slash commands but never wired into the Commander tree in `main.tsx`, so
@@ -1172,6 +1282,7 @@
 ## 1.54.0
 - Wired the 1.53.0 capability libraries into runnable commands. They were
   verified modules but nothing invoked them; these are the execution paths.
 - Added `/speak <text>` (alias `/say`), which drives the platform speech
@@ -1191,6 +1302,7 @@
 ## 1.53.0
 - Added automatic memory extraction (`src/memdir/extractFacts.ts`). Durable
   preferences and project conventions are proposed from user messages, deduped
   against stored memory by normalized key and containment, and ranked by
@@ -1221,6 +1333,7 @@
 ## 1.52.0
 - Implemented Ollama Cloud authentication. The Ollama client sent no
   `Authorization` header at all, so the hosted API was unreachable: local
   sessions only worked because the signed-in daemon proxies `:cloud` models on
@@ -1241,6 +1354,7 @@
 ## 1.51.0
 - Added named permission profiles. `settings.permissions.profiles` holds named
   rule sets (allow/deny/ask plus a description) and
   `settings.permissions.activeProfile` selects one; its rules are appended to
@@ -1259,6 +1373,7 @@
 ## 1.50.6
 - Implemented `--effort` on Ollama. The support predicate compared
   `getAPIProvider()` against `'firstParty'` — a value it can never return — so
   the advertised flag was silently dropped everywhere. Effort is now advertised
@@ -1294,6 +1409,7 @@
 ## 1.50.5
 - Added `.github/workflows/release.yml`. The repository had only a test
   workflow, so tags never became production releases: 18 tags existed with no
   GitHub Release and no automated publish. Pushing a `v*` tag now runs the full
@@ -1308,6 +1424,7 @@
 ## 1.50.4
 - Removed the `/install-github-app` command and its GitHub App setup flow,
   along with the startup tip suggesting `@ur` be tagged from issues and pull
   requests. The composite action at the repository root is removed with it.
@@ -1318,6 +1435,7 @@
 ## 1.50.3
 - Removed the GitHub App installation step from `/install-github-app`. UR
   authenticates in CI with the workflow's built-in `GITHUB_TOKEN` and a
   repository secret, so there is no app to install and no bot identity to
@@ -1333,6 +1451,7 @@
 ## 1.50.2
 - Fixed local-provider sessions showing "Not logged in · Run /login" with no
   account to log in to, introduced in 1.50.1. Credential ownership and URHQ
   auth applicability are separate questions: an Ollama session uses the user's
@@ -1347,6 +1466,7 @@
 ## 1.50.1
 - Fixed an always-true provider test that silently disabled a large part of the
   command surface. `isUsing3PServices()` was derived from `getAPIProvider()`,
   a request-shaping enum that never returns `'firstParty'`, so the comparison
@@ -1365,6 +1485,7 @@
 ## 1.50.0
 - Completed `/install-github-app` so `@ur <task>` works from GitHub. The
   mention now triggers on issue comments, pull-request comments, inline review
   comments, submitted reviews, and new issues; `/ur` remains accepted. Matching
@@ -1397,6 +1518,7 @@
 ## 1.49.0
 - Added cryptographically signed A2A Agent Cards: RFC 7515 detached JWS over
   the RFC 8785 canonical form of the card, using Ed25519 (`alg: "EdDSA"`).
   Verification recomputes the payload with `signatures` excluded, so a card can
@@ -1413,6 +1535,7 @@
 ## 1.48.0
 - Added managed cloud fan-out with durable, idempotent steering, owner-scoped
   mobile/A2A control, explicit PASS plus safe-branch selection, and
   cancellation-safe task transitions.
@@ -1431,6 +1554,7 @@
 ## 1.47.1
 - Hardened file downloads, filesystem permission checks, and signed skill
   trees against traversal and chained-symlink escapes.
 - Fixed task-memory integrity validation, prompt-plan file locking and change
@@ -1444,6 +1568,7 @@
 ## 1.47.0
 - Added a secure, opt-in AG-UI HTTP/SSE adapter with official schema/encoder
   integration, truthful capability discovery, ordered text/tool/state events,
   cancellation, exact CORS, bearer protection for network exposure, resource
@@ -1482,6 +1607,7 @@
 ## 1.46.0
 - Added a stable, official-SDK ACP v1 stdio agent with resumable sessions,
   client MCP transports and additional roots, streamed updates, native
   permission requests, cancellation, and private persisted session identity.
@@ -1513,6 +1639,7 @@
 ## 1.45.6
 - Deduplicated project verification approval so compile/test/lint commands are
   offered at most once per user turn. The approval marker is cleared for the
   next user task, preserving one explicit decision per task.
@@ -1522,6 +1649,7 @@
 ## 1.45.5
 - Bounded Ollama Cloud response-header and streaming phases to 120 seconds by
   default while preserving the five-minute allowance for local Ollama models.
   `API_TIMEOUT_MS` and per-request timeouts still take precedence.
@@ -1537,6 +1665,7 @@
 ## 1.45.4
 - Added mandatory provider-first model selection for the first interactive run
   in every workspace that has no project-local model. The validated provider
   and model pair is saved to `.ur/settings.local.json` before the REPL starts.
@@ -1549,6 +1678,7 @@
 ## 1.45.3
 - Made slash-command resolution deterministic across bundled skills, plugins,
   project skills, workflows, and built-ins. Duplicate canonical tokens are
   rejected by source priority and conflicting aliases are removed; registry
@@ -1569,6 +1699,7 @@
 ## 1.45.2
 Correctness and containment release completing the runtime audit.
 - Made sandbox, security-scope, WebFetch, API, browser, database, test-runner,
@@ -1601,6 +1732,7 @@ Correctness and containment release completing the runtime audit.
 ## 1.45.1
 Completes the three partially-delivered 1.45.0 items to 100%.
 - Semantic code search is now zero-config: the CodeSearch tool auto-enables
@@ -1617,6 +1749,7 @@ Completes the three partially-delivered 1.45.0 items to 100%.
 ## 1.45.0
 Top-tier feature release — closes the gaps against 2026's leading agents.
 - `ur cloud` — detached best-of-N tasks (the local-first codex-cloud
@@ -1661,6 +1794,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.10
 - Render `AskUserQuestion` permission requests inside a `PermissionDialog` so
   multiple-choice options appear as a bordered dialog box rather than a plain
   list.
@@ -1673,6 +1807,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.9
 - Fix recurring "String to replace not found in file" Edit errors by adding
   whitespace-tolerant matching (trailing whitespace, tab/space indentation).
 - Fix AskUserQuestion "questions type expected as array" validation errors by
@@ -1680,6 +1815,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.8
 - Keep auto-memory and automatic learning on by default with explicit opt-outs.
   Automatic learning can now be disabled with `automaticLearningEnabled: false`
   or `UR_CODE_DISABLE_AUTO_LEARNING=1`.
@@ -1690,6 +1826,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.7
 - Add `autoApprove` permission mode for command/tool approval prompts. It
   auto-approves operations that would otherwise require permission approval,
   while preserving user-input dialogs and explicit denials.
@@ -1697,14 +1834,17 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.6
 - Internal permission-mode iteration superseded by `1.44.7`.
 ## 1.44.5
 - Internal permission-mode iteration superseded by `1.44.6`.
 ## 1.44.4
 - The agent now learns from every run automatically — no `/learn run` needed.
   ci-loop, arena, escalation, and test-first completions fold their pass/fail
   outcome (per task category and model) into `.ur/learning/stats.json` as a
@@ -1726,6 +1866,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.3
 - Make thinking visually distinct from answers: thinking blocks are labeled
   "model reasoning to itself — not the answer" (dim italic, left-bordered when
   expanded); answer text carries an accent-colored ⏺ marker.
@@ -1759,6 +1900,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.2
 - Fix Ollama streamed tool-call accumulation: Ollama streams each completed
   tool call in its own chunk, but the merge logic overwrote call N-1 with
   call N, collapsing multi-call turns (e.g. several `Write` calls scaffolding
@@ -1783,6 +1925,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.1
 - Fix task board rendering: finished, failed, and skipped tasks now render as
   checked instead of unchecked.
 - Deduplicate consecutive task board emissions and keep final boards clean
@@ -1790,6 +1933,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.44.0
 - Add `verifier.askBeforeGates` setting (default `false`). When enabled, UR asks
   via `AskUserQuestion` whether to run project verification commands after a
   task, instead of auto-running tests/typecheck/lint gates. Available in
@@ -1800,6 +1944,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.43.6
 - Render output from reasoning models on OpenAI-compatible providers (LM Studio,
   vLLM). The streaming and non-streaming parsers now read `reasoning_content`
   (and `reasoning`) deltas and surface them as thinking blocks. Models that emit
@@ -1808,6 +1953,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.43.5
 - Fix model discovery for OpenAI-compatible providers (LM Studio, llama.cpp,
   vLLM) when base_url omits the API version segment. Discovery and `ur provider
   doctor` now also try `/v1/models` when base_url is just `host:port`, so
@@ -1817,6 +1963,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.43.4
 - Tolerate hallucinated extra parameters on tool calls: when input validation
   fails only because of unrecognized keys (e.g. `title`/`description` on a
   `Write` call), those keys are stripped and the call is re-validated instead
@@ -1825,6 +1972,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.43.3
 - Bias the assistant toward the interactive arrow-key select menu: the
   AskUserQuestion tool guidance now instructs the model to use the selectable
   menu whenever it offers the user a choice, instead of asking a free-form
@@ -1832,6 +1980,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.43.2
 - Fix artifact pages hanging blank: diff viewer assets (diff2html,
   highlight.js theme) are now served locally from `/assets` via the new
   `diff2html` dependency instead of render-blocking CDN tags, and the viewer
@@ -1842,6 +1991,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.43.1
 - Artifacts page renders diffs VS Code-style: side-by-side/inline views with
   syntax highlighting via diff2html (plain-text fallback when offline). New
   live view `/diff` shows current working-tree changes without manual capture,
@@ -1853,6 +2003,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.43.0
 - Add `ur artifacts serve [--port 4180]`: a local web page for artifacts.
   `GET /artifacts/<id>` renders one artifact (status, summary, feedback,
   content), `/` lists all, with `/artifacts/<id>/raw` and `/api/artifacts[/<id>]`
@@ -1860,6 +2011,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.42.0
 - Project safety policy no longer hard-blocks commands. Risky or deny-matched
   commands (package installs, destructive git operations, secret access,
   sandbox-required commands when the sandbox is unavailable) now surface as
@@ -1868,6 +2020,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.41.1
 - Harden provider tests against stored API keys in the local secure storage.
 - Revert sandbox default to disabled; expose `sandbox.enabled`,
   `sandbox.failIfUnavailable`, and `sandbox.allowUnsandboxedCommands` through
@@ -1875,6 +2028,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.41.0
 - Persist the model chosen through the interactive `/model` picker to settings
   and clear saved model state when `/model default` is used.
 - Enable the sandbox by default when no explicit `sandbox.enabled` setting is
@@ -1882,6 +2036,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.40.1
 - Pin `diff` to ^7 and OpenTelemetry packages to 2.6.1/0.214.0 to match the
   source API, fixing type errors from accidental dependency bumps, and rebuild
   the shipped bundle against these versions.
@@ -1889,11 +2044,13 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.40.0
 - Version bump: align package, build macro, VS Code extension, docs eyebrow, and
   changelog for the 1.40.0 release.
 ## 1.37.5
 - Version bump: align package, build macro, VS Code extension, docs eyebrow, and
   changelog for the 1.37.5 patch release.
 - Rename user-facing product/package branding to UR-Nexus while preserving the
@@ -1914,11 +2071,13 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.37.3
 - Version bump: align package, build macro, VS Code extension, docs eyebrow, and
   changelog for the 1.37.3 patch release.
 ## 1.37.2
 - Tightened provider reliability: API-provider calls now use a finite default
   timeout, consistent retry handling for transient network/provider failures,
   and safer OpenAI-compatible base URL normalization without changing streaming,
@@ -1940,6 +2099,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.35.1
 - Polished the bundled VS Code inline-diffs view with native toolbar icons,
   useful empty-state rows, clearer diff labels, and a cleaner review webview.
 - Fixed `ur ide status` routing so the IDE extension status action reports
@@ -1949,6 +2109,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.35.0
 - New `ur connect` CLI command (same implementation as the `/connect` slash
   command): `ur connect status`, `ur connect <provider>`,
   `ur connect <provider> --key <KEY>`, and `ur connect logout <provider>`.
@@ -1973,6 +2134,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.34.0
 - Restore the 1.30.3 subscription approach: Codex CLI, Claude Code, Gemini CLI
   and Antigravity are first-class in `/model` again — shown by default and
   usable directly (no `UR_ENABLE_EXTERNAL_APP_PROVIDERS` opt-in and no runtime
@@ -1984,6 +2146,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.33.0
 - Add API keys from inside UR while it is running: in `/model`, selecting an
   API provider (OpenAI, Anthropic, Gemini, OpenRouter) that isn't connected now
   shows a masked key-entry step. The key is stored in the OS keychain, then the
@@ -1994,6 +2157,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.32.0
 - `/model` now shows the subscription providers (Codex CLI, Claude Code, Gemini
   CLI, Antigravity) again. They are enabled the moment you `ur connect` them
   (persisted per-account opt-in) — no `UR_ENABLE_EXTERNAL_APP_PROVIDERS` env var
@@ -2009,6 +2173,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.31.0
 - Add in-app provider connection: `ur connect` / `/connect` connects a provider
   once and persists it. Subscription providers (Codex, Claude Code, Gemini,
   Antigravity) launch their official CLI login using your own account; API
@@ -2025,6 +2190,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.30.6
 - Restore a visible `subscription` access entry in provider lists without
   exposing provider app bridges as normal runtimes.
 - Keep subscription selection honest: no fake UR model IDs are listed, and the
@@ -2033,6 +2199,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.30.5
 - Hide external app bridge providers from normal `/model`, `/provider`, and
   `ur provider list` output. The default provider UX now shows only UR-native
   API, local, and OpenAI-compatible server runtimes.
@@ -2042,6 +2209,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.30.4
 - Make the default provider runtime independent of provider apps. Codex CLI,
   Claude Code, Gemini CLI, and Antigravity are now treated as explicit external
   app bridges and are blocked from normal `/model`, config save, and runtime
@@ -2052,6 +2220,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.30.3
 - Fix Codex CLI dispatch for real interactive terminals by inheriting terminal
   stdin for `codex exec`. Codex treats both `/dev/null` and closed pipes as
   piped stdin, so the previous `1.30.2` EOF approach still triggered
@@ -2059,6 +2228,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.30.2
 - Fix Codex subscription dispatch failing with `exited 1 ... Reading additional
   input from stdin`. `codex exec` reads stdin even when the prompt is an
   argument; UR now gives it a closed, empty stdin pipe (EOF) instead of
@@ -2067,6 +2237,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.30.1
 - Fix Codex CLI runtime dispatch by ignoring stdin when UR already passes the
   prompt as a command argument. This prevents `codex exec` from treating UR's
   closed pipe as extra stdin and exiting after `Reading additional input from
@@ -2074,6 +2245,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.30.0
 - IDE integration commands: `ur ide status`, `ur ide doctor`, and `ur ide config
   <editor>` for VS Code, Cursor, Windsurf, Zed, JetBrains, Neovim, and generic
   ACP clients. Status shows workspace, ACP server, provider/model, plugin count,
@@ -2094,6 +2266,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.29.1
 - Replace fabricated Claude Code and Gemini CLI static model names with
   provider-scoped CLI model aliases/names that the official CLIs can receive.
 - Reject stale subscription CLI selections such as `claude-code/sonnet-5` before
@@ -2103,6 +2276,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.29.0
 - Customer release consolidating the multi-provider selection and runtime
   dispatch work (1.27.5–1.28.1) into a single production line.
 - System-prompt identity now reflects the selected provider and runtime backend
@@ -2114,6 +2288,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.28.1
 - Keep the status bar synchronized with in-session provider/model changes from
   `/model`, `/model <model>`, and `/provider`, instead of waiting for persisted
   settings to reload.
@@ -2122,6 +2297,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.28.0
 - Subscription CLI providers (Codex, Claude Code, Gemini, Antigravity) now
   perform real dispatch: the official CLI is spawned in non-interactive mode with
   the scoped model and prompt, and its stdout becomes the response. Non-zero exit
@@ -2143,6 +2319,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.27.6
 - Route runtime requests through the selected provider/model pair instead of
   allowing stale Ollama/default-provider paths to handle non-Ollama requests.
 - Add runtime dispatch validation, backend labels, and focused mocked dispatch
@@ -2152,6 +2329,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.27.5
 - Make `/model` provider-first and provider-scoped, with clear subscription,
   API-key, local runtime, and OpenAI-compatible server labels.
 - Keep model discovery, validation, fallback, and saved config scoped to the
@@ -2162,6 +2340,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.25.3
 - Add provider alias resolution so `ur config set provider claude`,
   `ur config set provider "Claude Code"`, and `ur provider doctor agy`
   resolve to canonical provider IDs.
@@ -2170,6 +2349,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.25.2
 - Refresh public documentation so README, docs, static site, validation runbook,
   and code inventory all describe the current UR-Nexus feature set.
 - Document the recent provider auth, status bar, bundled VS Code extension,
@@ -2180,17 +2360,20 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.25.1
 - Fix VS Code extension installation to use the bundled UR-Nexus inline-diffs extension instead of the stale unpublished `urhq.ur` marketplace ID.
 - Harden AskUserQuestion normalization for description-only option objects and keep the eight-option schema in the production bundle.
 ## 1.25.0
 - Add legal multi-provider auth/provider management for subscription CLI, API-key, and local runtime access paths.
 - Add provider doctor/status/config commands, provider-aware status bar display, and explicit no-token-scraping safety policy.
 - Relax plan-mode clarification choices so professional redesign prompts do not fail when more than four options are supplied.
 ## 1.24.0
 ### Added
 - Plugin marketplace capability metadata for MCP tools, executable skills,
   templates, validators, language adapters, LSP servers, hooks, agents, and
@@ -2209,11 +2392,13 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.23.3
 ### Fixed
 - `repoIndex` tests failing in CI because `listIndexableFiles` silently returned `[]` when ripgrep (`rg`) was unavailable. Added a Node.js recursive file walker fallback that applies the same extension and skip-segment filters.
 ## 1.23.2
 ### Added
 - **CI failure diagnostics** in `.github/workflows/test.yml`: environment-info step, verbose test reporter, captured `test-output.log`, and artifact upload on failure so the production test runner exposes which test fails without requiring admin log access.
@@ -2222,6 +2407,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.23.1
 ### Added
 - **CI agent (`ur ci-loop`)** with agent constitution: hard rules against hiding failures, deleting without approval, editing generated/vendor files, claiming tests passed without execution, and changing public API without warning.
 - **Plugin marketplace extensibility** for `templates`, `validators`, and `languageAdapters` alongside existing MCP tools and skills.
@@ -2231,6 +2417,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.8
 ### Added
 -
@@ -2242,6 +2429,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.7
 ### Added
 - **Benchmark mode (`ur eval`).**
   - `ur eval run <suite> [--model <m>] [--metrics]` runs eval suites with explicit model overrides and per-case metrics files.
@@ -2286,6 +2474,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.6
 ### Fixed
 - Fixed Bash tool runtime execution failing every command with
   `timeoutMs is not defined` by keeping the command hook timeout value in scope
@@ -2301,6 +2490,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.5
 ### Added
 - **Real sandbox core architecture (`ur sandbox`).** New first-class command to inspect sandbox status, run dependency checks, initialize `.ur/safety-policy.json`, and evaluate shell-command approval levels.
 - **Worktree-per-task (`ur task`).** New command surface to start, run, list, and hand off agent tasks in isolated git branches/worktrees: `task start <name> [--worktree]`, `task run <id>`, `task pr <id> [--create]`, `task list`, `task status <id>`.
@@ -2316,6 +2506,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.4
 ### Added
 - **AST-aware `ur repo-edit` (P7).** Added
   `src/services/repoEditing/ast/types.ts`,
@@ -2355,6 +2546,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.3
 ### Added
 - **Executable skill directories.** A `.ur/skills/<name>/` directory containing
   `skill.yaml` is now an executable skill that compiles into a `WorkflowSpec`.
@@ -2385,6 +2577,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.2
 ### Added
 - **Lifecycle hooks.** Added six new hook events in `src/entrypoints/sdk/coreTypes.ts`
   and `src/entrypoints/sdk/coreSchemas.ts`: `BeforeEdit`, `AfterEdit`,
@@ -2411,6 +2604,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.1
 ### Added
 - **Rich task decomposition.** `src/services/agents/decomposer.ts` splits large
   goals into atomic subtasks with `goal`, `filesTouched`, `risk` (low/medium/high),
@@ -2443,6 +2637,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.22.0
 ### Added
 - **Agent execution metrics in `ur eval`.** `ur eval run` now captures
   cost, input/output tokens, model used, API duration, files changed,
@@ -2506,6 +2701,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.21.0
 ### Added
 - **Agent skill runner (`agentSkillRunner.ts`).** Reusable helper that wraps
   `startBackgroundTask({ worktree: true, pr: true })`, polls the background
@@ -2535,6 +2731,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.20.0
 ### Added
 - **ACP server (`ur acp`).** Added an HTTP+JSON-RPC Agent Communication
   Protocol server for IDE extensions. Supports `initialize`, `tools/list`,
@@ -2570,6 +2767,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.19.0
 ### Added
 - **Permission and safety policy (`ur safety`).** Added a project shell safety
   evaluator that separates read, write, execute, and network command classes;
@@ -2597,6 +2795,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.18.0
 ### Added
 - **Test-first execution loop (`ur test-first`).** Added a P0 quality loop that
   detects the project stack, orders compile/test/lint commands, runs them as
@@ -2621,6 +2820,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.17.0
 ### Added
 - **Reliable repo editing (`ur repo-edit`).** Added a P0 repo-editing workflow
   with dependency-free file/symbol indexing, indexed search, AST-aware
@@ -2637,6 +2837,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.16.0
 ### Added
 - **Network Ollama discovery.** `ur --discover-ollama` scans active local subnets
   for Ollama servers on port 11434, verifies each via `/api/tags`, and shows an
@@ -2659,6 +2860,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.15.0
 ### Changed
 - **Version bump.** Updated from 1.14.0 to 1.15.0 across `package.json`, `bunfig.toml`, and bundled CLI.
@@ -2667,6 +2869,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.14.1
 ### Changed
 - Removed the `desktop-app` startup tip pointing to the legacy desktop URL.
@@ -2675,6 +2878,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.14.0
 ### Changed
 - **Version bump.** Updated from 1.13.9 to 1.14.0 across `package.json`, `bunfig.toml`, and bundled CLI.
@@ -2683,6 +2887,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.9
 ### Added
 - **Spec-driven development (`ur spec`).** Scaffolds `requirements.md ->
   design.md -> tasks.md` plus a phase/approval `spec.json` under `.ur/specs/`,
@@ -2718,6 +2923,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.8
 ### Fixed
 - **Image paste resize fallback.** Clipboard image paste now falls back to
   macOS `sips` when the normal Sharp/native resize path cannot process an
@@ -2733,6 +2939,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.7
 ### Added
 - **Explicit update notice.** Interactive sessions now show
   `Update available: <current> -> <latest>` when a newer published package is
@@ -2749,6 +2956,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.6
 ### Added
 - **Professional static documentation site.** Added `documentation/` with a
   full HTML/CSS/JS documentation project covering installation, architecture,
@@ -2763,6 +2971,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.5
 ### Added
 - **Headless agent crews.** Added `ur crew` for lead/worker task boards that
   split a goal into subtasks, let worker subagents claim work, and support
@@ -2787,6 +2996,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.4
 ### Added
 - **Parallel workflow execution.** The declarative workflow executor now runs
   independent ready steps concurrently. `ur workflow run --concurrency <n>`
@@ -2819,6 +3029,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.3
 ### Added
 - **Checkpointed agent workflows.** Added `ur workflow` for declaring,
   validating, graphing, planning, resuming, and dry-running multi-step agent
@@ -2843,6 +3054,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.2
 ### Added
 - **Top-level code-index and role-mode commands.** `ur code-index` and
   `ur role-mode` are now registered in the main CLI, matching the shipped
@@ -2865,6 +3077,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.13.1
 ### Added
 - **AGENTS.md as runtime context.** UR now loads `AGENTS.md` (the cross-tool
   standard) from project roots at runtime, alongside `UR.md` and `.ur/rules/`.
@@ -2898,6 +3111,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.12.3
 ### Added
 - **Agent feature expansion commands.** Added `ur agent-features`,
   `ur agent-templates`, `ur automation`, `ur agent-task`, `ur model-doctor`,
@@ -2926,6 +3140,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.12.2
 ### Changed
 - **Ziggurat of Ur spinner.** Replaced the canoe spinner with the Ziggurat of
   Ur catching light: an up-pyramid whose lit face sweeps across (`△ ◭ ▲ ◮`)
@@ -2937,6 +3152,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.12.1
 ### Changed
 - **Mashoof spinner.** The activity spinner is now a Mashoof (مشحوف) — the
   marsh canoe — bobbing on the water. It cycles boat-hull arcs (`⌣ ⏝ ‿`) into
@@ -2945,6 +3161,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.12.0
 ### Added
 - **Agent trend coverage.** New `ur agent-trends` CLI command and
   `/agent-trends` slash command report how UR maps to current agent trends:
@@ -2964,6 +3181,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.11.3
 ### Changed
 - **Read-only web browsing.** `WebSearch` and `WebFetch` now run without
   prompting by default, while still respecting explicit deny or ask rules.
@@ -2972,6 +3190,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.11.2
 ### Fixed
 - **Clarification dialogs.** `AskUserQuestion` is now loaded without a
   `ToolSearch` round trip and accepts common question text aliases such as
@@ -2980,6 +3199,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.11.1
 ### Changed
 - **Npm publication docs.** README installation guidance now reflects that
   `ur-nexus` is published on npm, while keeping the GitHub install path for
@@ -2987,6 +3207,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.11.0
 ### Changed
 - **Ollama model selection now lets routing work by default.** The launcher no
   longer forces `OLLAMA_MODEL` when neither `OLLAMA_MODEL` nor `UR_MODEL` is
@@ -3017,11 +3238,13 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.10.2
 ### Fixed
 - **Clipboard image paste — the fix that actually ships.** The 1.10.1 change edited the native NSPasteboard branch, which is dead-code-eliminated from the bundle (its feature gate compiles out), so it never ran. The live path is osascript, whose `saveImage` reused a fixed temp file (`ur_cli_latest_screenshot.png`) opened `with write permission` but never truncated — so a smaller image pasted over a previously larger one kept the old trailing bytes, producing a corrupt PNG ("found in clipboard but not attached"). Added `set eof fp to 0` to truncate before writing.
 ## 1.10.1
 ### Fixed
 - **Clipboard image paste.** An image the clipboard reported as present but the native reader couldn't decode was silently dropped — "found in clipboard but not attached." `getImageFromClipboard` now falls back to the osascript path instead of treating a native `null` as authoritative.
 - **Token truncation on Ollama Cloud models.** Cloud models (the `-cloud` / `:cloud` suffix) now default to a 128K-token context floor for both `num_ctx` and auto-compaction, instead of the small or missing value `/api/show` reports for them — so prompts are no longer silently truncated, with no env vars required. The reported value is still used when it is larger, and `UR_OLLAMA_NUM_CTX` (no longer capped to the detected value) / `OLLAMA_CONTEXT_TOKENS` still override.
@@ -3032,6 +3255,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.10.0
 ### Added
 - **`skill-forge` plugin** in the `ur-plugins-official` marketplace — have the agent author skills for you. `/forge-skill <description>` runs on the active session model: it designs the skill (name, `when_to_use` triggers, arguments, minimal `allowed-tools`, inline vs fork, and steps that each carry a success criterion), shows the `SKILL.md` for a single confirmation, then saves it to `~/.ur/skills/<name>/` (or `./.ur/skills/` with `--project`) without clobbering an existing one. `/skill-refine <name> : <change>` improves an existing skill, and a bundled `skill-authoring` skill encodes the conventions. Complements the built-in `/create-skill`, which only scaffolds an empty template.
@@ -3040,6 +3264,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.9.0
 ### Added
 - **Seven first-party integration plugins** in the `ur-plugins-official` marketplace. Each bundles an official MCP server, curated slash commands, and a methodology skill, and falls back to a CLI or local library so the commands still work before any token is configured:
   - **`obsidian`** — operate a vault as a second brain: `/second-brain`, `/daily-note`, `/moc`, `/backlinks`, `/vault-search`. Direct vault file edits or the Obsidian Local REST API MCP server, plus a Zettelkasten/PARA/MOC skill.
@@ -3056,6 +3281,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.8.0
 ### Added
 - **`/create-skill` command.** Scaffold a new skill without leaving the REPL: `/create-skill <name> [: <description>] [--project]` writes a ready-to-edit `SKILL.md` (with frontmatter) to `~/.ur/skills/<name>/` — or `.ur/skills/` with `--project` — refuses to clobber an existing skill, and clears caches so it shows up immediately (alias `/new-skill`).
 - **Game Designer mode.** A new built-in output style (`/output-style`) that makes UR reason like a game designer — core loops, player fantasy, game feel, and tunable balance constants — while it writes working code.
@@ -3066,6 +3292,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.7.0
 ### Added
 - **Adaptive model routing (Ollama).** The agent auto-selects the best installed model per tier — the strongest coder model for the main loop, the smallest fast model for light internal work (titles, classification, session search, hooks). Honors `OLLAMA_MODEL` / `OLLAMA_SMALL_FAST_MODEL`; gated by `UR_OLLAMA_AUTO_ROUTE`.
 - **Per-model context auto-tuning.** Each request sets `num_ctx` from the model's real context window and the prompt size (floored at 32K for agent work, bucketed so the KV cache stays warm), fixing silent truncation at Ollama's 4096 default. Override with `UR_OLLAMA_NUM_CTX`.
@@ -3077,6 +3304,7 @@ Top-tier feature release — closes the gaps against 2026's leading agents.
 ## 1.6.0
 ### Added
 - **Proactive clarification & planning prompts.** The agent now uses the `AskUserQuestion` multiple-choice popup before significant or ambiguous work and at key planning decisions. Options are navigated with arrow keys and submitted; the last "Other" entry always lets you type a custom answer.
 - **Smarter prompt handling.** New always-on guidance makes the agent resolve ambiguity before acting, work in verifiable steps and check each step's output against the request before continuing, verify work actually runs before reporting done, report outcomes faithfully, and keep changes precisely scoped and professional.