Skip to content

minidsh 1.2.0

Latest

Choose a tag to compare

@github-actions github-actions released this 26 Oct 03:13
· 86 commits to main since this release

Background work, a third provider, web search, skills and a real browser, built across four sessions; then the whole widened runtime was reviewed as one system and what that review found was fixed before release. Read the migration note at the end of this section before upgrading a deployment that sets a temperature, a base URL or an agent preset.

Added

  • A real browser to check a frontend. Where Chrome, Edge or Chromium is installed, the model gets five tools:

    • browser_navigate opens a page and waits for a dev server that is still starting;
    • browser_snapshot reads the page as an accessibility tree with refs;
    • browser_act clicks, hovers, types, presses keys and selects, as real input that is refused when the element is hidden, disabled or covered;
    • browser_screenshot captures the page, on a vision-capable route only;
    • browser_evaluate runs JavaScript in the page.

    Every result opens with what the page logged, failed, blocked or asked since the previous one. Each session gets its own headless browser under a private profile in ~/.minidsh/browser/, started on first use and gone when the session ends, over a pipe rather than a debugging port. It reaches only localhost and the workspace's own files, plus any allowHosts you name in the tool-browser row. This is enforced inside the browser (a proxy the harness owns, resolver rules and a file: filter) and proved each time it starts. Downloads are refused. A script still running at a call's deadline is stopped, and a page that still does not answer (a loop that re-arms itself) is replaced by a fresh one on the next call; the model is told either way. It asks no consent, so a delegated verifier child can use it: a text-only model hands the look to a subagent entry routed to a vision model.

  • Background jobs. job_start runs a long command beside the session — a dev server, a slow test run — and returns at once; job_output reads what it wrote since the last read (optionally waiting for it to end) and job_cancel stops it and what it started. A job's start and its one end are durable facts in the log; when a job ends on its own its notice arrives in the session and wakes an idle agent, at most eight times in a row before a person says something, and a stop someone asked for never wakes it. A job runs under the session's own authority, never an escalation, a durable switch of the sandbox mode or acceptance ends it, and it ends with its session. If the MiniDSH process dies, the next resume (and sessions verify before it) closes each running job as lost and tells the model so.

  • OpenAI as a third provider. --provider openai --model gpt-6-luna (or /model openai/gpt-6-luna, or session/model) routes a session to OpenAI's Responses API, mid-session included, with OPENAI_API_KEY in the environment or credentials.json. Nothing is stored at OpenAI (store: false): a reasoning model's encrypted reasoning travels back to the same model inside the session log, and is dropped, never forged, for any other. The catalog covers the current GPT-6, GPT-5.6, GPT-5.5, GPT-5.4, GPT-5.2 and GPT-5 mini/nano models with their measured effort sets; an option a model would refuse or silently ignore — an effort it lacks, a temperature with reasoning on, an output cap above its own — is refused before anything is sent.

  • Web search, opt-in. A composition.json row ({"patches": [{"insert": [{"id": "tool-web-search", "plugin": "tool-web-search"}]}]}) gives the model web_search, backed by Tavily (TAVILY_API_KEY). Every query asks first: the line you answer names the exact text that would leave the machine, where it goes and under which credential, and a session grant covers only that identical query. Results arrive as a numbered list of titles, URLs and snippets behind a fixed notice that they are untrusted content — never a provider's own summary. A delegated child is not offered the tool, and its pinned approvals would refuse every query anyway. There is no web_fetch.

  • Skills: know-how as data. A folder holding a SKILL.md (the open agentskills.io format) in ~/.minidsh/skills/, inside the package, or in a repository's .agents/skills/ or .claude/skills/ is listed to the model by name and description, and the model reads one with its ordinary file tools when a task fits. A skill is advice, never a permission: its scripts are shell commands under the session's own authority. A repository's skill never replaces yours or a bundled one (the operator is warned on stderr). Two ship, on Node alone: office builds, reads, unpacks and validly repacks .docx, .xlsx and .pptx, and pdf reads PDFs and writes simple ones. Real Word, Excel and PowerPoint open what they build.

  • Jobs on every surface. The session view lists jobs; job/cancel stops one over the protocol; the terminal has /jobs and /jobs cancel <id>; the browser shows a jobs strip with a stop button and per-session badges. A new control/subscribe sends every client a per-session summary (status, running jobs, held notices, pending approvals) whatever session it is watching, and a question stays answerable while such a client is connected.

Changed

  • The default tool set grows by the five browser tools on a machine with a browser installed, and minidsh config lists the tool-browser row (flagged when a layer changes it, since allowHosts widens what it reaches). A refused image now says that a subagent starts with a conversation of its own; the image bound is the whole conversation's until a compaction.
  • deepseek-flash sees images, as DeepSeek now serves it, in a user message or a tool result; the retired deepseek-v4-flash names resolve to it and take the same facts. deepseek-v4-pro stays text-only (it accepts an image and ignores it). A DeepSeek id the catalog does not know is treated as text-only with a 131,072-token window rather than the full million.
  • Breaking: DeepSeek refuses temperature while thinking. Thinking is on unless the reasoning effort is off, and in thinking mode DeepSeek ignores the temperature without saying so; a request that sets one is now refused before anything is sent, naming the two ways out: effort off, or a route to a model that honours a temperature (a recorded temperature cannot be cleared). A settings.json with agent.temperature and no effort will see its first DeepSeek step refused.
  • Breaking: the model's shell no longer sees OPENAI_API_KEY. The OpenAI adapter is mounted by default, which makes its key one of the harness's own credentials, withheld from every shell command like the DeepSeek and Anthropic keys. A repository script that read it through MiniDSH's shell needs it passed another way.
  • A log written by this version may carry a consent subject 1.1.0 does not know (a web search's): 1.1.0's sessions verify reports such a log invalid rather than misreading it, and 1.1.0 will not resume it; 1.0.0, which records no subjects, ignores it. A home composition.json that mounts web search stops 1.1.0 before it reads anything.
  • A key no HTTP header can carry (a stray line break, a NUL) is refused by its name before any request, for every provider and for web search: the platform's own error for it quoted the value.
  • The consent line shows invisible characters escaped — zero-width and format characters, bidi controls, the Unicode tag block, private-use code points — in a shell command as in a web query, so what a person approves is every character that runs or leaves.
  • A one-shot run stops what it started. minidsh run (and a delegated child) stops a job still running when its task ends, records it, and says so on stderr; the exit code is unchanged.
  • On Windows a process is killed with its tree (taskkill /T): a timed-out or cancelled command's children no longer outlive it.
  • PowerShell one-shot commands write UTF-8, as the persistent shell always did: on a host whose console code page is not UTF-8 their non-ASCII output arrived garbled.
  • The Anthropic catalog follows Anthropic's own model listing (2026-10-25), and a live drift probe now holds it there, as the DeepSeek and OpenAI ones already were. The 5.5 models have rows of their own instead of their predecessors'. Effort off is refused where the model cannot turn thinking off (Opus 5.5, Sonnet 5.5, Fable 5 and 5.1). The 4.6–4.8 models are asked to think at a named effort (Opus 4.8, measured, did not think without it); Opus 4.5 thinks on a budget beside its effort. Sonnet 4.5's window is 200K. A temperature beside thinking is refused before anything is sent, since the API refuses it. An id the catalog does not know gets no output cap of ours.
  • DeepSeek errors are classified by their HTTP status first: a 5xx is retried as a server error even when its text mentions "context"; DeepSeek running out of capacity mid-stream is retried; its content filter ends the turn as a refusal.
  • llm-retry gives up on a provider that asks for a long wait: a retry-after past maxRetryAfterMs (default 60 s) ends the retries instead of holding the turn, and the failure says how long the provider asked for.
  • The editor reads big files in part. view of a file over 8 MiB shows its first 8 MiB and says how to read the rest with the shell; str_replace and insert refuse such a file the same way; an empty old_str is refused. Reading or writing a FIFO or a device is refused, never waited on.
  • A resume or fork composes the world its session was created in: under another agent preset it is refused, and over the wire (serve, web) an explicit --sandbox, --approval or --accept on the host now applies to a resumed session as it did on the CLI.
  • minidsh config warns about more: a provider row a layer changed (its baseURL decides where every conversation and its key go) and any plugin loaded from a module path.
  • The office helper stops early. office read stops rendering a document or deck once it is past what --max-chars prints (the paragraph counts then read N+), as it already did for a workbook, and renders a slide two entries name once; pack refuses such a deck.
  • Releases wait for the live suite. release.yml publishes only after a green check.yml and an unfiltered green live.yml for the tagged commit, runs the same install smoke as check.yml, and holds its publish credential in a job that installs no dependency. Contributors: pnpm lint fails on warnings.

Security

Found by the 1.2 checkpoint's review; each fix has a test that failed before it.

  • The file fence on Windows compared names case-insensitively, and so read a name with the Kelvin sign (U+212A) as the workspace's own k directory: a write could land outside the workspace. Containment now compares the canonical path exactly.
  • A write racing a link swap (another process replacing a workspace directory with a link while the harness writes) could land content outside the workspace; on Linux 13 files did in 49 writes. The harness now checks the file it actually opened, through /proc/self/fd on Linux and by inode elsewhere (best effort there), and refuses otherwise. On Linux only an empty file or directory can still be left outside; on macOS and Windows a swap timed between the checks can still let content through.
  • Keys stay out of error text. A provider or a proxy that echoes the key in an error no longer gets it quoted into the log, the model's view or stderr (all three providers and web search). A base URL carrying a username or password is refused, from config or a *_BASE_URL variable. A malformed credentials.json is reported by position, never with the parser's message, which quoted the file.
  • The session commands blank invisible characters (zero-width, bidi, tag and private-use code points) in every line they print, not only consent lines: sessions list, sessions show, inspect, verify and the audit. A consent line also escapes every blank that is not an ordinary space, which could carry hidden bits after a visible query.
  • Another user's .git above your working directory (a planted /tmp/.git, or a link from there into your own home) no longer makes their .agents/skills or AGENTS.md part of your workspace, on Linux and macOS; a project marker in a filesystem root (C:\, /) is ignored on every platform. The operator is told when one is.
  • Hostile files cannot hang or crash the host or its helpers. A repository SKILL.md linked to /proc or a FIFO is skipped instead of read forever. The office and pdf helpers refuse deeply nested or oversized documents with a coded exit, where they used to overflow the stack or exhaust the heap, and no longer take quadratic time on a long word, many short lines or a long run of digits.
  • minidsh web refuses to be framed by another page (X-Frame-Options, frame-ancestors, Sec-Fetch-Dest), so a local page cannot overlay a signed-in tab.
  • Authority switches are all-or-nothing: a session/authority call, or a host's explicit flags on a resumed session, that a delegated session's ceiling or pin refuses no longer leaves the switches before it applied. The authority invariant also checks where a delegated session's acceptance sits, every field of a recorded decision, and that a decision credited to a standing consent names one live for that exact action.

Fixed

  • A host out of file descriptors (EMFILE) refused nothing and could crash on a shell, job or browser spawn; each now refuses, under its usual code, with a message that says it is a moment, not a missing install. A confinement probe that could not run is warned about rather than passing silently for "this host has no backend"; restart once the load passes. A one-shot or job command too long to pass as an argument, or holding a NUL, is refused as the command's (SHELL_COMMAND_INVALID), with the way out.
  • A browser call stuck in page code (a looping getter in a returned value) is stopped at its deadline instead of wedging the page until it was replaced.
  • A cancel at the wrong moment could lose a job's end notice, and a step that rejected or filtered its input dropped the notices it had claimed. Both now keep them.
  • A tool result of the wrong shape left a call with no result, which crash repair later had to close; it is now an error result.
  • A compaction summary cut off at its length limit replaced the history anyway; it is declined. And a summarizer could spend its whole answer calling a tool, so compaction declined an empty summary again and again: the request now shows the tools and forbids calling them. Compaction no longer fires on pressure the system prompt and tool list alone create, which it could not relieve.
  • exit as the first shell command was reported as a broken shell (SHELL_UNAVAILABLE); it is the command's own end, reported with its output. A shell session disposed while it restarted into another sandbox mode could leave a shell running.
  • str_replace treated $& and $$ in the replacement as patterns, corrupting the text it wrote.
  • A refused agent creation leaked the session, so every later attempt in the same host failed "already exists".
  • The DeepSeek and Anthropic stream readers buffered a stream whose lines end in CRLF (a proxy's) until it failed; two DeepSeek tool calls sharing a stream index were merged into one; a base URL with a trailing slash doubled it.
  • An Anthropic tool call whose arguments were not a JSON object was replayed to the model as a call with no arguments.
  • A relative MINIDSH_HOME moved with each session's working directory; it is resolved once.
  • The WebSocket carrier could queue one large answer past its hard limit, and kept a closed socket open while its peer stopped reading. minidsh web failing to listen (a port in use) leaked everything it had started.
  • The terminal's first output could race its attach; settings that fail to parse are SETTINGS_INVALID.

Migration from 1.1

  • Sessions. The log format is unchanged: 1.2 resumes and forks 1.1 sessions. Going back, do not resume with 1.1.0 a session that ran background jobs: 1.1.0 continues an unfinished job without closing it as lost, so its model may believe a dead job is still running. 1.1.0 refuses a session in which web search ever asked for consent.
  • Now refused before anything is sent, saying why: a DeepSeek temperature while thinking and a resume or fork under a different agent preset (both used to pass silently); an Anthropic temperature beside thinking and effort off where a model cannot turn thinking off (the API refused them); a base URL carrying a password (every request failed, quoting it).
  • Credentials. The model's shell no longer sees OPENAI_API_KEY.

On versioning. Some of the changes above turn a silent degradation into a refusal (a temperature the provider would have ignored, a password every error quoted, a hostile document that used to hang a helper). This project ships such a change in a minor release: a silent degradation or a leaked secret may become a refusal without a new major version. Everything else follows Semantic Versioning.