1# zutfile agents
2
3A zutfile packages an agent's behavior into one portable `.zut` file. It can contain the agent's instructions, reusable skills, static assets, and metadata describing the runtime, model, operating-system, binary, and tool permissions it needs.
4
5The current implementation supports creating, inspecting, verifying, and running local directories and `.zut` archives. It can also run an agent repository or directory directly from any public GitHub repository without keeping a clone. zut has no built-in owner or official collection. Indexed registry distribution, installation, signatures, bundled executable extensions, network permissions, and environment permissions are not implemented yet. Native `web_search` is unavailable to packaged agents until network permissions are enforceable.
6
7## Quick start
8
9Create a directory for the agent:
10
11```text
12reviewer/
13├── manifest.json
14├── AGENT.md
15└── skills/
16    └── code-review/SKILL.md
17```
18
19Add a minimal `manifest.json`:
20
21```json
22{
23  "zutfile": 1,
24  "name": "code-reviewer",
25  "version": "0.1.0",
26  "description": "Reviews a repository and reports actionable findings.",
27  "runtime": {
28    "min_zut": "0.2.76"
29  },
30  "model": {
31    "requires": ["tools"],
32    "min_context": 64000
33  },
34  "permissions": {
35    "fs": {
36      "read": ["${workspace}"],
37      "write": []
38    },
39    "bash": {
40      "mode": "none"
41    }
42  },
43  "requirements": {
44    "os": ["darwin", "linux", "windows"]
45  },
46  "entry": {
47    "greeting": "What should I review?",
48    "default_prompt": null
49  }
50}
51```
52
53Add the agent's standing instructions to `AGENT.md`:
54
55```markdown
56# Code reviewer
57
58Review the current repository without modifying it.
59
60Prioritize correctness, security, regressions, and missing tests. Report findings
61in severity order with file and line references. Do not praise the code or
62summarize files that have no findings.
63```
64
65Test the directory directly during development:
66
67```bash
68zut inspect ./reviewer
69zut run ./reviewer
70zut run ./reviewer "Review the authentication package"
71```
72
73Package it when ready:
74
75```bash
76zut pack ./reviewer
77# wrote code-reviewer.zut
78# digest sha256:...
79
80zut verify ./code-reviewer.zut
81zut run ./code-reviewer.zut
82```
83
84The first run displays the declared permissions and asks for consent before the agent starts.
85
86## Directory layout
87
88```text
89my-agent/
90├── manifest.json          # identity, requirements, and permissions, required
91├── AGENT.md               # persona and standing instructions
92├── skills/                # SKILL.md directories, optional
93│   └── my-skill/SKILL.md
94├── assets/                # static files, optional
95└── README.md              # human-facing documentation, optional
96```
97
98Only `manifest.json`. Other ordinary files and directories are included in the archive, but the runtime gives no special behavior to `assets/` or `README.md` yet. Files in `assets/` are not automatically added to the model context or granted filesystem access.
99
100Do not include executable extensions in `extensions/`. The local runtime rejects any bundled extension directory containing `extension.json` because extension subprocesses cannot yet be confined to the manifest permissions. This is a deliberate fail-closed restriction.
101
102Symlinks are not supported and cause `zut pack` to fail.
103
104## `AGENT.md`
105
106`AGENT.md` is optional. When present, it is appended to zut's system prompt by default. It should describe the agent's role, workflow, constraints, output format, and when to load bundled skills. When absent, it contributes no agent-specific text; other enabled global addenda still follow their normal rules.
107
108Keep capability and security declarations out of this file. Permissions come only from `manifest.json` and are enforced independently of agent-authored prose.
109
110Set `replace_system_prompt` to `true` in the manifest to use `AGENT.md` as the replacement base identity:
111
112```json
113{
114  "replace_system_prompt": true
115}
116```
117
118Replacement is intended for fully specialized agents. It replaces zut's built-in identity, while global append addenda such as enabled Ponytail coding guidance, project context, and skills retain their normal inclusion rules. The default layering behavior is usually preferable because it retains zut's normal identity and tool-use guidance.
119
120## Bundled skills
121
122Place skills under `skills/<name>/SKILL.md` using the normal zut skill format:
123
124```text
125skills/
126└── investigate-failure/
127    └── SKILL.md
128```
129
130```markdown
131---
132name: investigate-failure
133description: Diagnose a failing command from its output and relevant source files.
134---
135
136# Investigate failure
137
1381. Reproduce the failure if permissions allow it.
1392. Identify the first meaningful error rather than downstream noise.
1403. Read the implementation and nearest tests.
1414. Propose the smallest correction and validation plan.
142```
143
144Bundled skills are added to normal skill discovery while the zutfile is running. The model sees their name and description in the skill manifest and can load the full body through the `skill` tool. See [skills.md](skills.md) for the complete skill format.
145
146## Manifest reference
147
148The current manifest shape is:
149
150```json
151{
152  "zutfile": 1,
153  "name": "code-reviewer",
154  "version": "0.1.0",
155  "description": "Reviews a repository and reports actionable findings.",
156  "license": "MIT",
157  "runtime": {
158    "min_zut": "0.2.76"
159  },
160  "model": {
161    "requires": ["tools", "reasoning"],
162    "min_context": 64000,
163    "preferred": ["claude-opus-4-5", "gpt-5.5"],
164    "min_tier": ""
165  },
166  "permissions": {
167    "fs": {
168      "read": ["${workspace}", "${agent_data}"],
169      "write": ["${agent_data}"]
170    },
171    "bash": {
172      "mode": "allowlist",
173      "allow": ["git", "go"]
174    },
175    "net": {
176      "allow": []
177    },
178    "env": {
179      "read": []
180    }
181  },
182  "requirements": {
183    "bin": ["git", "go"],
184    "os": ["darwin", "linux"]
185  },
186  "entry": {
187    "greeting": "What should I review?",
188    "default_prompt": null
189  },
190  "replace_system_prompt": false
191}
192```
193
194### Top-level fields
195
196| Field | Required | Current behavior |
197|---|---:|---|
198| `zutfile` | yes | Format version. Must be `1`. |
199| `name` | yes | Local agent identity and session/data namespace. See naming rules below. |
200| `version` | no | Displayed in consent and inspection output. Semver is recommended but not currently validated. |
201| `description` | no | Displayed by `zut inspect`. |
202| `license` | no | Package metadata. Not otherwise interpreted by the local runtime. |
203| `runtime.min_zut` | no | Minimum zut version. Older binaries refuse to run the agent. Unversioned development builds cannot satisfy a non-empty minimum. |
204| `model` | no | Model capabilities, minimum context, and preferences. |
205| `permissions` | no | Filesystem and bash permission ceiling. Omitted scopes deny access. Native web search remains unavailable because network permissions are not enforced. |
206| `requirements` | no | Required operating systems and executables. |
207| `entry` | no | Initial presentation and prompt metadata. |
208| `replace_system_prompt` | no | Replaces the built-in identity rather than appending the agent body when `true`; global append addenda retain their normal inclusion rules. Defaults to `false`. |
209
210### Name rules
211
212The local runtime currently accepts a flat lowercase name containing only:
213
214- `a` to `z`
215- `0` to `9`
216- `.`, `-`, and `_`
217
218Examples:
219
220```text
221code-reviewer
222repo_onboarder
223acme.audit
224```
225
226Names containing `/`, uppercase letters, spaces, or path-like values such as `..` are rejected. Registry-style `namespace/agent` names are reserved for the future distribution runtime and are not accepted locally yet.
227
228### Model requirements
229
230`model.requires` currently recognizes:
231
232| Capability | Behavior |
233|---|---|
234| `tools` | Accepted. Every model in the current catalog is treated as tool-capable. |
235| `reasoning` | Requires a catalog model marked as supporting reasoning. |
236| `vision` | Recognized, but currently fails closed because vision support is not represented in the model catalog. No model qualifies. |
237
238Unknown capabilities are rejected. Use `tools`, not the draft spelling `tool_calling`.
239
240`model.min_context` requires a catalog model with at least that many context tokens.
241
242`model.preferred` is an ordered list of model IDs. It is a preference, not a provider requirement. Selection proceeds as follows:
243
2441. An explicitly selected model, if present, must satisfy the requirements.
2452. The user's configured default is retained if it satisfies them.
2463. The first compatible model in `preferred` is selected.
2474. Otherwise, zut chooses a compatible active catalog model.
2485. If none qualifies, execution stops with an error.
249
250`model.min_tier` is present in the format shape but is not supported by the local runtime. Leave it empty or omit it. A non-empty value is rejected.
251
252### Runtime and host requirements
253
254Use `requirements.os` to restrict the agent to Go operating-system names:
255
256```json
257{
258  "requirements": {
259    "os": ["darwin", "linux", "windows"]
260  }
261}
262```
263
264Use `requirements.bin` for commands that must already be on `PATH`:
265
266```json
267{
268  "requirements": {
269    "bin": ["git", "python3"]
270  }
271}
272```
273
274Zut checks these requirements before requesting consent. These zutfiles do not have install or postinstall hooks, so authors must document how users can obtain missing programs.
275
276### Entry fields
277
278`entry.pre` is auto-submitted once when `zut run` starts, before the initial prompt is applied. Values that begin with `!` use the interactive shell-escape path (or `BashTool` in `--print` / `--stream` / `--json`). Shell output streams live to the TUI (interactive) or stderr (non-interactive). Other values are sent as a normal user turn; in `--stream` mode that turn's assistant text also streams to stdout. After `pre` finishes, zut reloads extensions and rediscovers skills so anything installed by that command is available for the following turn.
279
280```json
281{
282  "entry": {
283    "pre": "!npx -y skills add tt-a1i/archify --skill archify --agent cursor --copy --yes"
284  }
285}
286```
287
288`entry.default_prompt` is used when the user runs the agent without a prompt. It may be a string or `null`:
289
290```json
291{
292  "entry": {
293    "default_prompt": "Review the current repository"
294  }
295}
296```
297
298In interactive mode, when no prompt is given on the command line, `default_prompt` pre-fills the editor after `pre` finishes (or immediately when `pre` is empty). It is not auto-submitted. A prompt supplied on the command line takes precedence the same way.
299
300In `--print` / `--stream` / `--json` modes, `pre` still runs first: shell escapes abort the run on failure; plain-text `pre` is sent as a first agent turn before the main prompt (`default_prompt` or the CLI prompt).
301
302`entry.greeting` is part of the manifest, but the current local runtime does not render it yet.
303
304## Filesystem permissions
305
306Filesystem permissions are enforced by the built-in `read`, `write`, `edit`, and `create_worktree` tools. Empty or omitted scopes deny the operation.
307
308Two variables are available in filesystem scopes:
309
310| Variable | Resolves to |
311|---|---|
312| `${workspace}` | The current working directory used for the run. |
313| `${agent_data}` | The agent's persistent private data directory under `$ZUT_HOME/agents/<name>/data/`. |
314
315Relative paths are resolved beneath `${workspace}`. For example:
316
317```json
318{
319  "permissions": {
320    "fs": {
321      "read": ["src", "go.mod"],
322      "write": ["${agent_data}"]
323    }
324  }
325}
326```
327
328This allows reads under `<workspace>/src` and at `<workspace>/go.mod`, plus writes under the private agent data directory. It does not grant general workspace writes.
329
330Common permission profiles follow.
331
332Read-only repository agent:
333
334```json
335{
336  "permissions": {
337    "fs": {
338      "read": ["${workspace}"],
339      "write": []
340    },
341    "bash": { "mode": "none" }
342  }
343}
344```
345
346Repository editor with private persistent state:
347
348```json
349{
350  "permissions": {
351    "fs": {
352      "read": ["${workspace}", "${agent_data}"],
353      "write": ["${workspace}", "${agent_data}"]
354    },
355    "bash": { "mode": "none" }
356  }
357}
358```
359
360Data-only agent:
361
362```json
363{
364  "permissions": {
365    "fs": {
366      "read": ["${agent_data}"],
367      "write": ["${agent_data}"]
368    },
369    "bash": { "mode": "none" }
370  }
371}
372```
373
374Filesystem checks canonicalize existing paths, or the nearest existing parent for a new path, so symlinks cannot be used to escape a declared scope.
375
376`create_worktree` bootstraps a repository without an existing worktree root in two calls. Its first branch-only call returns no-write guidance for the agent to ask the user whether to use `.worktrees` or an external directory. The reply is sent as `bootstrap_root`: `.worktrees` creates `<repository-root>/.worktrees/<branch>` and adds `/.worktrees/` to the repository root's `.gitignore`; an absolute external path creates `<external-root>/<branch>` without changing repository files. The selected root is stored privately as `zut.worktrees.path` in local Git config, so later calls only need `branch`.
377
378The tool needs read and write scopes covering the repository root, Git metadata, and the selected worktree root. It also requires Bash permission for `git`, even though it invokes Git directly without a shell:
379
380```json
381{
382  "permissions": {
383    "fs": {
384      "read": ["${workspace}"],
385      "write": ["${workspace}"]
386    },
387    "bash": { "mode": "allowlist", "allow": ["git"] }
388  }
389}
390```
391
392The tool disables Git hooks, but a repository checkout can still use configured Git filters. Treat `git` permission for `create_worktree` as permission to run repository-controlled checkout behavior; jail mode remains an accident-prevention guardrail, not a security sandbox.
393
394If the working directory changes during a run, `${workspace}` permissions are expanded again for the new directory.
395
396## Bash permissions
397
398`permissions.bash.mode` supports three values.
399
400### `none`
401
402```json
403{
404  "bash": { "mode": "none" }
405}
406```
407
408All bash tool calls are denied. This is also the default when `mode` is omitted.
409
410### `ask`
411
412```json
413{
414  "bash": { "mode": "ask" }
415}
416```
417
418Bash is available after the user approves the agent at launch. Consent for an agent using `ask` is requested on every launch and is not cached.
419
420In the current local runtime, `ask` is launch-time capability consent. It does not open a separate confirmation dialog for every individual bash invocation.
421
422### `allowlist`
423
424```json
425{
426  "bash": {
427    "mode": "allowlist",
428    "allow": ["git", "go", "grep"]
429  }
430}
431```
432
433Every command in a shell expression must be listed. Zut checks commands separated by `;`, `&&`, `||`, and pipes. Paths are reduced to their base command name, so `/usr/bin/git` is checked as `git`. A leading environment assignment is skipped when identifying the command.
434
435The allowlist rejects shell substitution, newlines, redirection, backticks, `$()`, `<`, and `>`. This intentionally supports simple command pipelines rather than arbitrary shell programs.
436
437An `allowlist` mode with an empty `allow` array is invalid.
438
439Only declare commands the agent genuinely needs. `requirements.bin` checks that a command exists, while `permissions.bash.allow` controls whether the agent may invoke it. Agents that execute a required command normally need it in both places.
440
441## Network and environment permissions
442
443The manifest shape includes `permissions.net.allow` and `permissions.env.read`, but the local runtime does not enforce them yet. To avoid presenting unenforced declarations as security controls, it rejects manifests containing any network host or environment variable. `permissions.net` is rejected rather than ignored; it does not authorize the built-in `web_search` tool. Packaged agents receive no native web search until a future runtime design enforces network permissions.
444
445Use empty arrays or omit these sections:
446
447```json
448{
449  "permissions": {
450    "net": { "allow": [] },
451    "env": { "read": [] }
452  }
453}
454```
455
456Bundled executable extensions are rejected for the same reason. Network filtering, scrubbed extension environments, and extension confinement belong to a later runtime milestone.
457
458## Consent and persistent data
459
460Before the first run, zut prints the agent identity and expanded filesystem and bash permissions, then asks:
461
462```text
463Agent code-reviewer@0.1.0 wants to run.
464
465  fs read: /path/to/workspace
466  fs write: none
467  bash: none
468
469Allow? [y/N]
470```
471
472For modes other than `bash: ask`, approval is cached for the exact artifact digest under:
473
474```text
475$ZUT_HOME/agents/<name>/consents/<digest>.json
476```
477
478Any change to the packaged artifact produces a different digest and requires consent again. `bash: ask` always requires fresh launch consent.
479
480Non-interactive runs refuse to bypass consent by default. For controlled automation, pass `-y` / `--yes` or set:
481
482```bash
483zut run -y ./code-reviewer.zut --print "Review this repository"
484```
485
486```bash
487ZUT_AGENT_CONSENT=1 zut run ./code-reviewer.zut --print "Review this repository"
488```
489
490This environment variable skips the consent prompt. Only use it after independently inspecting and trusting the exact artifact being run.
491
492Persistent agent data lives under:
493
494```text
495$ZUT_HOME/agents/<name>/data/
496```
497
498The directory is created for every run, but the agent can access it only when `${agent_data}` appears in the relevant filesystem permission scope.
499
500## Agent-scoped sessions
501
502Sessions created by a zutfile are isolated from ordinary zut sessions and from other agents:
503
504```text
505$ZUT_HOME/sessions/agents/<name>/
506```
507
508Normal session flags still apply, including `--continue`, `--resume`, `--session`, and `--no-session`, but their default storage root is scoped to the active agent. The interactive `/sessions` picker and `/session` import, fork, and tree commands use that same agent-scoped root; `/fork` is the direct alias for selecting a user message and branching after that turn. `/session tree` shows the current session's family within that root (its ancestor and descendants), keeps pre-compaction user and assistant history available for checkout, and hides tool/internal rows from the fork picker. Tree-navigation branches stay hidden from the flat `/sessions` picker. Resuming honors the selected session's stored provider/model pair; a failed provider/model rebuild leaves the current session unchanged.
509
510## Commands
511
512### `zut run`
513
514Run an unpackaged directory during development:
515
516```bash
517zut run ./my-agent
518zut run -y ./my-agent
519zut run ./my-agent "Do the task"
520zut run ./my-agent -y --print "Do the task"
521```
522
523Run a packed artifact:
524
525```bash
526zut run ./my-agent.zut
527zut run ./my-agent.zut --print "Do the task"
528zut run ./my-agent.zut --stream "Do the task"
529zut run ./my-agent.zut --json "Do the task"
530```
531
532Arguments after the reference use zut's normal CLI parser, so model, provider, reasoning, cwd, session, tool, and output-mode flags remain available. The manifest still imposes its permission and compatibility ceiling. In interactive mode, enabling **show loaded resources at startup** in `/settings` displays the manifest name in an `[Agent]` section above the transcript.
533
534Local filesystem directories, local archive paths, GitHub shorthand, and public GitHub agent-directory URLs are accepted. A single-part name resolves only to a matching local directory or `.zut` archive:
535
536```bash
537zut run my-agent
538```
539
540A two-part name is treated as a GitHub `owner/repository` reference. Resolution remains local-first, then uses the repository root:
541
542```bash
543zut run frkr/zot-archify
544```
545
546Use three or more parts to select an agent directory from any public GitHub collection:
547
548```bash
549zut run acme/agents/reviewer
550zut run acme/clankers/reviewers/go
551```
552
553There is no built-in owner, official collection, collection configuration, or allowlist. Every shorthand part must use lowercase letters, digits, dots, hyphens, or underscores. Backslashes are also accepted as separators on Windows. Prefix a reference with `./`, use an absolute path, retain the `.zut` suffix, or provide a complete GitHub URL to force local or explicit remote resolution.
554
555For GitHub, zut downloads the repository archive into a temporary directory, selects the requested agent subdirectory, validates it, runs it, and removes the downloaded files when the command exits:
556
557```bash
558zut run https://github.com/acme/agents/reviewer
559```
560
561A standard GitHub tree URL is also supported:
562
563```bash
564zut run https://github.com/acme/agents/tree/main/reviewer
565```
566
567Short forms read the repository's default branch through GitHub's `HEAD` archive. A tree URL uses the branch or tag in the URL. Private repositories and branch or tag names containing `/` are not currently supported. The downloaded source is temporary, but normal agent data, consent receipts, and session transcripts remain under `$ZUT_HOME`.
568
569`owner/repository` and `owner/repository/path` are direct GitHub mappings, not an indexed or signed registry. Installed names, arbitrary URLs, and OCI references are not resolved yet.
570
571### `zut pack`
572
573```bash
574zut pack [directory] [output]
575```
576
577Examples:
578
579```bash
580zut pack
581zut pack ./my-agent
582zut pack ./my-agent ./dist/reviewer.zut
583```
584
585The default directory is the current directory. When output is omitted, zut uses `<manifest-name>.zut`. If the output has no `.zut` suffix, zut adds one.
586
587Packing validates the manifest and directory, then creates a zstd-compressed tar archive. Entries are sorted and tar metadata is normalized with fixed timestamps and numeric ownership. The output archive itself is excluded when it is located inside the source directory. Zut prints the SHA-256 digest of the resulting compressed artifact.
588
589### `zut inspect`
590
591```bash
592zut inspect ./my-agent
593zut inspect ./my-agent.zut
594zut inspect https://github.com/acme/agents/code-reviewer
595```
596
597Inspection validates and prints the agent's name, version, description, digest, declared permissions, and complete file list. It does not execute the agent.
598
599For a directory, the digest is computed from its canonical uncompressed tar representation. For a `.zut` archive, the digest is computed from the archive bytes. Therefore a source-directory digest and its packed-archive digest are not expected to match.
600
601### `zut verify`
602
603```bash
604zut verify ./my-agent.zut
605```
606
607The local `verify` command validates that the archive can be safely loaded, validates the manifest and required layout, and prints its SHA-256 digest.
608
609Despite the command name, it does **not** verify a publisher signature in the current runtime. Signature and namespace verification are planned for registry distribution.
610
611## Packaging and extraction limits
612
613The archive loader applies these limits:
614
615| Limit | Value |
616|---|---:|
617| Compressed `.zut` file | 100 MiB |
618| One extracted entry | 64 MiB |
619| Total extracted content | 256 MiB |
620
621Archive extraction rejects absolute paths and parent traversal. Unsupported tar entry types are ignored. New packages use zstd; gzip-compressed tar archives are accepted only as a compatibility fallback for earlier experiments.
622
623## Authoring guidance
624
6251. Start with the narrowest permissions possible. A reviewer usually needs repository reads, not writes or bash.
6262. Keep `AGENT.md` focused on stable behavior. Put task-specific procedures in skills so the model loads them only when relevant.
6273. Declare every external command in both `requirements.bin` and the bash allowlist when the agent must execute it.
6284. Test from a directory first with `zut inspect` and `zut run`.
6295. Test denied behavior as well as allowed behavior. Ask the agent to attempt an out-of-scope read, write, and command.
6306. Pack the same source twice and compare behavior and output digests before distributing it.
6317. Inspect and verify the final archive, not only its source directory.
6328. Do not claim that network, environment, signatures, extensions, installation, or registry delivery are secured or supported by the local runtime.
633
634## Current scope
635
636Implemented now:
637
638- local directories and `.zut` archives
639- local-only resolution for single-part names
640- local-first `owner/repository` and `owner/repository/path` resolution to any public GitHub repository
641- temporary execution of agent directories from public GitHub repositories
642- `zut pack`, `zut inspect`, `zut verify`, and `zut run`
643- canonical tar creation with zstd compression
644- archive digest reporting and safe extraction limits
645- `AGENT.md` append or replacement behavior
646- bundled skills
647- runtime, OS, and binary requirements
648- model context and reasoning requirements
649- filesystem scope enforcement
650- bash `none`, `ask`, and `allowlist` modes
651- digest-scoped consent receipts
652- private agent data directories
653- agent-scoped sessions
654
655Not implemented yet:
656
657- `zut install`, `zut agents`, `zut use`, `zut update`, or `zut publish`
658- installed-name, arbitrary URL, OCI, configurable registry, or GitHub index resolution
659- network allowlist enforcement
660- environment-variable filtering
661- safely confined bundled executable extensions
662- automatic artifact theme loading
663- agent dependency composition
664- per-agent Telegram, bot, or RPC selection
665- `--no-sandbox`, `--trust-bash`, or registry trust override flags
666
667Treat the unsupported fields and commands in the broader zutfile proposal as forward-looking design, not as current runtime behavior.