@spinalfluid/ultrac 0.1.6 → 0.1.11

package/README.md

@@ -35,7 +35,7 @@ For local testing before a registry release, build and install the verified tarb
 pnpm check
 pnpm verify:install
 npm pack
-npm install -g ./spinalfluid-ultrac-0.1.6.tgz
+npm install -g ./spinalfluid-ultrac-0.1.7.tgz
 ultrac login
 ```
 
@@ -141,13 +141,34 @@ ultrac "Find the main entry points and summarize the architecture"
 Inside a session:
 
 - Type to talk to Ultrac; it inspects the repository with tools as needed.
+- On supported terminals, the full-width input box stays at the bottom while tools and replies
+  run above it. The footer shows the current model, elapsed task time, reported token usage,
+  and estimated context usage. Unreported token counts appear as unknown.
+- Type `/` to browse built-in commands and available skills. Type to filter, use arrow keys
+  to select, and press Tab or Enter to fill the command. Press Enter again to run it;
+  Esc dismisses the menu. `/skill <name> [args]` runs a skill without a custom command.
+- Multiple-choice questions accept arrow keys and Enter, or a mouse click where the terminal
+  supports mouse reporting. You can also type a number or your own answer.
 - You can type another message while Ultrac is working. It is queued and inserted after the current
-  tool batch or completed model response, before the next model request. Slash commands wait for
-  the main prompt instead of entering model context.
-- Ctrl-C stops active work at a safe boundary. At an idle prompt, Ctrl-C exits and saves the session.
+  tool call finishes, before the next model request. Remaining tools in that batch are skipped.
+  Slash commands wait for the main prompt instead of entering model context.
+- Esc stops the current turn at a safe boundary and returns to the prompt. Ctrl-C exits and saves
+  the session.
 - Gated tools (file writes, shell, network) prompt for approval: `once` / `always` / `no`.
+- `/mcp` shows the connected servers in this session; `/doctor` checks local setup.
+- `/continue [focus]` continues an unfinished task using its current transcript. `/autonomous on`
+  enables up to five rounds when a task reaches its step ceiling. Checkpoints, a six-hour elapsed
+  limit, stall detection, and Esc cancellation remain active. `/autonomous off` restores one round.
+  Failed rounds are not automatically replayed; provider retry/failover handles request failures.
+- The agent can use `create_skill` in default/coding modes (or with `--tools all`) to create a
+  project playbook with full instructions and an optional slash command. It requires write
+  permission, refuses overwrites, and retains the workspace trust requirement for loading.
 - Coding commands: `/plan [task]`, `/coding`, `/review [focus]`, `/diff [--staged]`,
   `/test [instructions]`, `/init`, `/max-steps [number]`.
+- Quiet output: `/ghost [on|off|status]` hides activity headers and progress for the current
+  session. Answers, errors, permission prompts, and private diagnostics remain visible or
+  recorded. `/logs` shows the current diagnostic log path. Ghost mode does not change
+  network routing, identities, or permissions.
 - Context and sessions: `/context`, `/compact [focus]`, `/usage`, `/status`, `/save`,
   `/sessions`, `/resume <id>`, `/new`, `/clear`.
 - Controls: `/help`, `/model [name]`, `/mode [name]`, `/permissions [readonly|workspace|full]`,
@@ -168,6 +189,15 @@ Inside a session:
   rules or automatically delete session data: exclude `.ultrac/sessions/` from version control and
   manage retention yourself.
 
+Tool activity stays compact. Interactive sessions keep successful tool bodies out of the
+main conversation. Use `/output` to inspect the latest saved tool result, `/output list`
+to find its number, or `/output 3` to inspect a particular result. This display preference
+does not shorten model answers or alter the tool results stored in the session.
+
+One-shot runs and `tools run` show at most six preview lines per result, with a hidden-line
+count. Failed results retain the end of the log. Pass `--verbose-tools` for complete
+results; `run --json` and `tools run --result-json` also keep their complete structured output.
+
 Non-interactive, scriptable one-shot runs:
 
 ```bash
@@ -388,7 +418,12 @@ Successful `ask_user` answers are kept whole so a constraint at the end is not l
 count against the context window, so an oversized request fails locally before it reaches the API.
 
 **Every request is measured before it is sent.** The provider adds up the system prompt, the
-serialised tools, and the transcript. Over 60% of the window, the harness first tries to compact
+serialised tools, and the transcript. During longer tasks, the harness also attempts a
+periodic summary every 12 completed steps when eligible older evidence exceeds 12,000
+characters. Set `ULTRAC_SUMMARY_INTERVAL_STEPS` to an integer from 0 to 1000; 0 disables
+periodic summaries. Every user instruction and the latest two complete assistant/tool
+units are retained. Failed summaries leave the transcript intact. Checkpoints preserve
+successful summaries across resume, while archived tool results remain available. Over 60% of the window, the harness first tries to compact
 the oldest turns that would otherwise be dropped into a bounded `[compacted-context]` summary.
 If summarising fails or there is not enough room, it falls back to dropping whole old turns, so an
 assistant `tool_calls` entry and its `tool` replies are removed together. The system prompt, pinned
@@ -496,7 +531,7 @@ custom skills. Each skill is a folder with a `SKILL.md` (YAML frontmatter + Mark
 ```
 
 ```bash
-ultrac skills create open-pr --description "Open a pull request" \
+ultrac skills create open-pr --description "Open a pull request" --command open-pr \
   --when "the user asks to open a PR" --permissions shell,network
 ultrac skills list
 ultrac skills show open-pr
@@ -673,3 +708,39 @@ revocable, non-transferable right to *run* the published build - nothing more. C
 modification, redistribution, and reverse engineering are reserved. The full terms are in the
 `LICENSE` file that ships inside the package, and they are the whole agreement: nothing on this
 page grants a right the licence withholds.
+
+### Recovery during unattended tasks
+
+Autonomous runs resume from completed-step checkpoints, use cancellation-aware exponential
+backoff for transient failed rounds (1 second up to 30 seconds), and stop on permanent
+provider failures or repeated unchanged progress. Recovery defaults to five rounds and a
+six-hour elapsed budget. Diagnostics record round timing, retry categories, summary events,
+and stop reasons without prompts or tool bodies. Checkpoints do not guarantee rollback or
+exactly-once execution for a tool interrupted before its completed result was saved.
+
+### Configuring MCP servers
+
+The harness connects to local stdio MCP servers, discovers their tools, and makes them
+available to agent turns after workspace approval. Add and inspect a server with:
+
+```sh
+ultrac mcp add my-server node -- /absolute/path/to/server.js
+ultrac trust
+ultrac mcp list
+ultrac mcp call my-server tool-name --input '{"query":"example"}'
+ultrac mcp remove my-server
+```
+
+Use `--env-pass TOKEN_VARIABLE` to forward a named environment variable without writing
+its value to configuration. Setup changes preserve unrelated settings and require renewed
+workspace approval. HTTP/SSE MCP servers are not supported by this transport yet.
+
+Create a reusable skill and expose it as a slash command with:
+
+```sh
+ultrac skills create project-check --description "Run project checks" --command project-check
+```
+
+Edit the generated `SKILL.md` instructions, approve the project workspace, and invoke
+`/project-check` in the terminal. The agent can also write a new `SKILL.md` with its file tools and load it with
+`skill`; normal write and tool permissions still apply.