@loonglab/loongcode-cli 0.1.1 → 0.1.3

package/CHANGELOG.md

@@ -4,8 +4,33 @@ All notable changes to LoongCode will be documented in this file.
 
 ## [Unreleased]
 
+## [0.1.3] - 2026-10-13
+
+A model-configuration and context-management release. The setup flow sets a
+model's context window and output ceiling again; `/model` edits and removes
+models — seeded from what is saved, never displaying a stored key, and always
+tested before anything is written; and a run whose context overflows no longer
+ends when compaction cannot help: the request is reduced instead, and every
+compaction that cannot progress is reported rather than discarded. It also
+records the rename and the command consolidation the code has carried since
+0.1.2, which are listed below.
+
 ### Added
 
+- Model configuration is complete again in the product. The setup flow's
+  unnumbered `Configure limits…` sub-step sets a new model's `contextWindow` and
+  `maxOutputTokens` (defaulting to 128000 and 8192, validated as positive
+  integers with the output ceiling no larger than the window); the values are
+  shown on the screen that confirms the connection test, and the probe tests
+  exactly the configuration that will be saved. `/model` offers `Edit model…`,
+  which runs the same sequence seeded from what is saved — the id and its
+  derived name stay put, the stored API key is never displayed and is either
+  kept or replaced through the masked prompt, and nothing is written until a
+  connection test succeeds — and `Remove model…`, which confirms by name.
+  Removing the active model reopens setup immediately instead of waiting for the
+  next task; removing any other model leaves the active one alone. All three
+  actions are refused while a task or a compaction is running, so a model change
+  can never take the workspace view out from under a run.
 - An empty transcript now opens with a large `LOONGCODE` block wordmark and the
   descriptor "an opinionated coding agent", followed by the existing keyboard
   hints. The hero is a true empty state: it is centered as a group in the
@@ -58,6 +83,11 @@ All notable changes to LoongCode will be documented in this file.
 
 ### Removed
 
+- The `settings.json` key `model`. It documented a "default active model id"
+  that nothing ever read, and its validation could fail a settings load — and
+  through it a run — over a value with no effect. The active model has always
+  been owned by `models.json`; a settings file that still carries the key keeps
+  loading, because unknown keys are ignored.
 - Slash command `/reload`. Prompt templates are discovered live, `AGENTS.md` /
   `CLAUDE.md` and agent-facing resources load on every run, so there is nothing
   left to refresh manually.
@@ -77,6 +107,115 @@ All notable changes to LoongCode will be documented in this file.
   capabilities now live inside the `/session` manager; they are not retained
   as aliases.
 
+### Fixed
+
+- A run no longer ends when its context overflows and compaction cannot help. In
+  a single-task run the compactable region is the task itself — everything the
+  run produced sits after the last user message — so compaction reported
+  `no-progress / empty-tail` and the run stopped with `context overflow detected
+  but compaction made no progress: empty-tail`. The request projection is now
+  reduced instead: tool output the model can get back is withheld from the
+  request — output a cap spilled to a captured artifact (`spillPath`), or a
+  `read` that names the file it read (`sourcePath`, plus `resumeOffset` when the
+  read was capped) — while the two most recent recoverable results, failed
+  executions, and observed-failure evidence are kept, and each withheld result
+  states what it held and how to get back to it. The two kinds of route do not
+  promise the same thing: a spill artifact holds the output as it was, while a
+  path re-reads the file's current contents — so a read whose file a later write
+  or edit in the same history names is kept, rather than offered a route that
+  would return something the model never read. Only the request is reduced:
+  durable history is untouched, and a reference that no longer resolves fails as
+  a normal read error rather than returning invented content. Recovery is
+  bounded like compaction retries (five per run), never sends a reduced request
+  identical to one a provider has already rejected, and ends the run naming the
+  failed stage and carrying the provider's own diagnostic when nothing
+  recoverable remains.
+- Compaction that cannot progress is no longer silent. The proactive check used
+  to discard a `no-progress` outcome entirely, so the only trace of it was the
+  failure it eventually caused; every attempt now emits a `compaction_skipped`
+  event naming the reason (`empty-tail`, `no-compactable-region`,
+  `empty-summary`) and the trigger (`proactive`, `reactive`, `provider`), a
+  recovery emits `projection_reduced`, and the run record carries
+  `compactionSkips` and `projectionRecoveries` so a finished run still explains
+  itself. A proactive verdict is reported once per unchanged history: the
+  attempt is not repeated — with its checkpoint and its event — on every
+  iteration that changes nothing compaction could act on, and is retried as soon
+  as the history moves. `/compact` reports which reason applied instead of a
+  bare "nothing to compact".
+
+## [0.1.2] - 2026-10-12
+
+A cross-platform compatibility release. Windows gets a real choice of shell —
+Windows PowerShell by default, CMD on request — and the tool layer no longer
+assumes a POSIX userland underneath: file search runs in-process instead of
+shelling out to `grep`, command output is decoded by contract rather than
+assumed to be UTF-8, cancelling a command always returns, and `@file`
+references work with spaces and Windows paths.
+
+Verification remains Linux: the automated suite runs there, the Windows paths
+are covered by simulated tests, and native Windows and macOS runs are still
+outstanding.
+
+### Added
+
+- **Windows shell selection.** `settings.shell` chooses which shell commands
+  run in on Windows: `"powershell"` (Windows PowerShell, the default) or
+  `"cmd"` (CMD, taken from `%ComSpec%`). It is global with a per-project
+  override, and an explicit choice is never replaced by the default. It applies
+  to Windows only — other platforms always run commands in `/bin/sh`, with the
+  prompts and tool descriptions they have always had. Nothing switches shells
+  automatically: if the selected shell cannot start, the command fails with an
+  error naming it instead of running somewhere else, and PowerShell 7 (`pwsh`)
+  is never chosen for you. CMD and PowerShell are different languages, so the
+  system prompt and the tool description state which one is active.
+
+### Changed
+
+- **Windows PowerShell is the default shell on Windows.** An installation that
+  never set `settings.shell` now runs commands in `powershell.exe` instead of
+  CMD; one that already chose `"cmd"` — globally or per project — keeps CMD.
+  The change is Windows only.
+- **File search no longer needs an external `grep` binary.** The `grep` tool
+  searches in-process, so it works where no `grep` exists — on Windows it
+  previously failed with "Executable not found in $PATH", which the agent read
+  as "no results". The interface, the output format (matching lines with
+  `file:line` prefixes, the 100-match cap, `No matches found.`) and the search
+  scope (`.git` and `node_modules` are still searched) are unchanged. Regular
+  expressions are JavaScript's rather than POSIX ERE: `\d`, lookahead and
+  non-greedy quantifiers now work, while `[[:alpha:]]`, `\<` and `\>` are
+  rejected with an error naming the replacement instead of silently matching
+  something else. This is not full GNU grep compatibility — the dialect, the
+  `include` glob (now the same dialect as `find`, where `{a,b}` alternates) and
+  the handling of unreadable paths all differ as described here.
+- **A search that cannot read part of the tree says so, and still returns what
+  it found.** Previously a single unreadable path failed the whole search and
+  discarded the matches; the result now carries them with a "(Search
+  incomplete: …)" note. Special files such as FIFOs and devices are skipped
+  rather than read, which removes a way for a search to block indefinitely.
+- **Subprocess output is decoded as UTF-8 unless the whole stream proves
+  otherwise.** On Windows, output that is not valid UTF-8 — `cmd.exe`
+  built-ins and many native tools write the console code page — is decoded
+  with the console code page and the result says which one was used, instead
+  of arriving as mojibake. That fallback is a labelled best guess, not
+  detection of what a program intended: the console code page describes the
+  console, not the program, and a stream can be valid UTF-8 while having been
+  written in another encoding. When no code page can be determined, the output
+  says that instead of pretending. POSIX decoding is unchanged.
+
+### Fixed
+
+- **Cancelling a command always returns.** Aborting a run (Esc) settles the
+  tool call within a bounded grace period even when the command refuses to
+  die; previously a Windows cancellation whose `taskkill` was missing or failed
+  could leave the call waiting forever, with no error and no diagnostic.
+  Termination is still best-effort — a process or descendant that escapes the
+  process tree can survive — and the result says explicitly when cleanup could
+  not be confirmed, rather than implying the tree is gone.
+- **`@file` references now work with spaces and Windows paths.** `@"my
+  notes.md"` and `@C:\src\a.ts` (either slash) expand to the file's content
+  like any other reference; they used to be left as literal text. Unresolved
+  references still stay literal, so `@` in ordinary prose is never mangled.
+
 ## [0.1.1] - 2026-10-12
 
 Fixes Backspace in the terminal environments where the key arrives as a raw

package/README.md

@@ -11,6 +11,10 @@ in an interactive terminal UI.
 LoongCode is a single platform-independent package: the same install works on
 Windows, macOS, and Linux, on both x64 and arm64.
 
+Cross-platform compatibility improvements are implemented and automated
+verification has been performed on Linux; native Windows and macOS verification
+remains pending.
+
 ```sh
 npm install -g @loonglab/loongcode-cli
 cd my-project
@@ -21,9 +25,16 @@ On first launch no model is active, and LoongCode says so: it opens the setup
 flow (protocol `openai` or `anthropic`, endpoint, model name, API key) before
 the workspace appears, and the normal interface is not reachable until a model
 is active. `/model` is that same flow — it lists the models you have configured
-and offers `Add model…` for a new one. Context and output limits are defaulted,
-and only exposed under an optional *Configure limits…* step. Cancelling setup
-exits; a task you had already typed is not run and is not saved.
+and offers `Add model…`, `Edit model…` and `Remove model…`. Context and output
+limits are defaulted to 128000 and 8192, and changeable in the flow's optional
+*Configure limits…* step. Cancelling setup exits; a task you had already typed
+is not run and is not saved.
+
+Editing a model reuses the flow, seeded from what is saved: the stored API key
+is never displayed — you choose to keep it or enter a replacement — and the
+change is saved only after a connection test succeeds. Model changes are refused
+while a task or a compaction is running, and removing the active model reopens
+setup.
 
 ## Scripting
 
@@ -44,4 +55,30 @@ Configuration and sessions live under your user config directory
 (`~/.loongcode/`) — never inside
 this package.
 
+## Windows shell
+
+On Windows, commands run in Windows PowerShell (`powershell.exe`) by default.
+To run them in CMD instead, add `shell` to `settings.json`:
+
+```json
+{
+  "shell": "cmd"
+}
+```
+
+- `"powershell"` (default) — `powershell.exe`, started with `-NoLogo
+  -NoProfile -NonInteractive`; `"cmd"` — `cmd.exe`, taken from `%ComSpec%`.
+- Global (`~/.loongcode/settings.json`) with a per-project override
+  (`.loongcode/settings.json`); the project value applies, so an explicit
+  choice is never replaced by the default.
+- Windows only: other platforms always run commands in `/bin/sh`.
+- CMD and PowerShell are different languages — their syntax, built-ins and
+  scripts are not interchangeable, and neither has POSIX commands (`ls`, `cat`,
+  `grep`, …).
+- Nothing switches shells automatically: if the selected shell is unavailable
+  the command fails with an error naming it, including when PowerShell is the
+  default. `pwsh` (PowerShell 7) is not selected for you.
+- Choose `cmd` for CMD syntax, batch built-ins, or tooling that expects
+  `%ComSpec%`.
+
 See the repository README for the full product documentation.