@withgateway/sdk 0.2.5 → 0.2.6

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 (75) hide show
  1. package/dist/cli.js +3121 -1283
  2. package/dist/cli.js.map +1 -1
  3. package/dist/cli.mjs +3185 -1347
  4. package/dist/cli.mjs.map +1 -1
  5. package/dist/files/index.js +9 -0
  6. package/dist/files/index.js.map +1 -1
  7. package/dist/files/index.mjs +9 -0
  8. package/dist/files/index.mjs.map +1 -1
  9. package/dist/hooks/index.d.mts +1 -1
  10. package/dist/hooks/index.d.ts +1 -1
  11. package/dist/hooks/index.js +280 -8
  12. package/dist/hooks/index.js.map +1 -1
  13. package/dist/hooks/index.mjs +272 -8
  14. package/dist/hooks/index.mjs.map +1 -1
  15. package/dist/{index-1xnxt0Cc.d.ts → index-BO2SNSgy.d.ts} +12 -1
  16. package/dist/{index-B0R3IgsV.d.mts → index-BaKu5ZSU.d.mts} +124 -2
  17. package/dist/{index-B0R3IgsV.d.ts → index-BaKu5ZSU.d.ts} +124 -2
  18. package/dist/{index-C1wDEwH4.d.mts → index-Cz8Q6Keo.d.mts} +12 -1
  19. package/dist/index.d.mts +4 -4
  20. package/dist/index.d.ts +4 -4
  21. package/dist/index.js +726 -167
  22. package/dist/index.js.map +1 -1
  23. package/dist/index.mjs +718 -167
  24. package/dist/index.mjs.map +1 -1
  25. package/dist/traces/index.js +9 -0
  26. package/dist/traces/index.js.map +1 -1
  27. package/dist/traces/index.mjs +9 -0
  28. package/dist/traces/index.mjs.map +1 -1
  29. package/dist/tracing/index.d.mts +1 -1
  30. package/dist/tracing/index.d.ts +1 -1
  31. package/dist/tracing/index.js +97 -61
  32. package/dist/tracing/index.js.map +1 -1
  33. package/dist/tracing/index.mjs +97 -61
  34. package/dist/tracing/index.mjs.map +1 -1
  35. package/dist/worlds/index.d.mts +154 -4
  36. package/dist/worlds/index.d.ts +154 -4
  37. package/dist/worlds/index.js +348 -97
  38. package/dist/worlds/index.js.map +1 -1
  39. package/dist/worlds/index.mjs +348 -97
  40. package/dist/worlds/index.mjs.map +1 -1
  41. package/package.json +1 -1
  42. package/runtime/gatewaysdk/environment/_bundle.py +1 -1
  43. package/runtime/gatewaysdk/environment/_materialize.py +53 -9
  44. package/runtime/gatewaysdk/environment/_toml.py +13 -0
  45. package/runtime/gatewaysdk/environment/schema/api.py +3 -7
  46. package/runtime/gatewaysdk/environment/schema/batch.py +15 -11
  47. package/runtime/gatewaysdk/environment/schema/bundle_runtime.py +24 -0
  48. package/runtime/gatewaysdk/environment/schema/compiler.py +60 -28
  49. package/runtime/gatewaysdk/environment/schema/conform.py +1 -1
  50. package/runtime/gatewaysdk/environment/schema/connector.py +1 -1
  51. package/runtime/gatewaysdk/environment/schema/connectors/google-drive/readers/merge.ingest.toml +25 -0
  52. package/runtime/gatewaysdk/environment/schema/connectors/google-drive/readers/merge.toml +17 -0
  53. package/runtime/gatewaysdk/environment/schema/extract.py +1 -5
  54. package/runtime/gatewaysdk/environment/schema/host_surfaces.py +1 -5
  55. package/runtime/gatewaysdk/environment/schema/mapping.py +1 -1
  56. package/runtime/gatewaysdk/environment/schema/platform.py +16 -4
  57. package/runtime/gatewaysdk/environment/schema/redaction.py +1 -1
  58. package/runtime/gatewaysdk/environment/schema/scaffold.py +9 -2
  59. package/runtime/gatewaysdk/environment/schema/skills/benchmarks-getting-started/SKILL.md +67 -9
  60. package/runtime/gatewaysdk/environment/schema/skills/hooks-getting-started/SKILL.md +61 -2
  61. package/runtime/gatewaysdk/environment/schema/skills/world-data-ingestion/SKILL.md +4 -1
  62. package/runtime/gatewaysdk/environment/schema/skills/worlds-getting-started/SKILL.md +11 -11
  63. package/runtime/gatewaysdk/environment/schema/skills/worlds-getting-started/references/cli-mcp-map.md +17 -1
  64. package/runtime/gatewaysdk/environment/schema/snapshot.py +1 -5
  65. package/runtime/gatewaysdk/environment/schema/tools_world.py +1 -5
  66. package/runtime/gatewaysdk/environment/schema/ui.py +1 -5
  67. package/runtime/gatewaysdk/environment/schema/world_tests.py +1 -1
  68. package/runtime/gatewaysdk/environment/world.py +120 -4
  69. package/skills/benchmarks-getting-started/SKILL.md +67 -9
  70. package/skills/connector-schema-authoring/connector-worlds.md +8 -1
  71. package/skills/connector-schema-authoring/world-schema.md +1 -1
  72. package/skills/hooks-getting-started/SKILL.md +61 -2
  73. package/skills/world-data-ingestion/SKILL.md +4 -1
  74. package/skills/worlds-getting-started/SKILL.md +11 -11
  75. package/skills/worlds-getting-started/references/cli-mcp-map.md +17 -1
@@ -70,9 +70,9 @@ Expected output (JSON): `"kind": "harbor"`, `"tasks": ["hello"]`, the files, and
70
70
  code-bench/
71
71
  benchmark.json # kind, models, rollouts: the defaults `benchmarks run` reads
72
72
  pyproject.toml # [project] name + version: the slug and the version `push` records; not a Python package
73
- tasks/hello/task.toml # version, [metadata], [agent] timeout_sec, [verifier] timeout_sec, [environment] docker_image
73
+ tasks/hello/task.toml # version, [metadata], [agent] timeout_sec, [verifier] timeout_sec; docker_image only as a comment
74
74
  tasks/hello/instruction.md # what the agent is told
75
- tasks/hello/environment/Dockerfile # the task's container (FROM docker_image; WORKDIR /app)
75
+ tasks/hello/environment/Dockerfile # the task's container, built by the runner (FROM --image; WORKDIR /app): the ONE image source
76
76
  tasks/hello/tests/test.sh # the grader: runs after the agent, writes /logs/verifier/reward.json
77
77
  tasks/hello/solution/solve.sh # the reference solution; must score 1.0 (harbor's oracle agent runs it)
78
78
  ```
@@ -82,7 +82,40 @@ The reward contract is the only thing the platform reads: `tests/test.sh` writes
82
82
  other key a sub-reward — or a bare number to `/logs/verifier/reward.txt`. Any language, any method.
83
83
 
84
84
  Add a task: copy `tasks/hello` to `tasks/<name>`, edit `instruction.md`, `tests/test.sh`,
85
- `solution/solve.sh`, and the image (`docker_image` or the Dockerfile). Two tasks minimum.
85
+ `solution/solve.sh`, and the image. Two tasks minimum.
86
+
87
+ **One image source.** Harbor's rule (`should_use_prebuilt_docker_image`): when `task.toml` names
88
+ `[environment] docker_image`, the runner pulls that prebuilt image and runs it as it is —
89
+ `environment/Dockerfile` is **never built**, so anything it installs is absent on the runner (a tool the
90
+ agent needs, a pip package) while a Dockerfile-first local run passes. Without `docker_image` the
91
+ runner builds the Dockerfile. The scaffold therefore writes only the Dockerfile; put what the agent
92
+ needs in it. To run a prebuilt image instead, delete the Dockerfile and set `docker_image`.
93
+ `validate` fails a task that has both when the Dockerfile adds anything beyond `FROM`/`WORKDIR`
94
+ (a note when it does not), and `dry-run` follows the runner's choice (`image … (prebuilt: docker_image)`,
95
+ `pull`, no `build`).
96
+
97
+ **Secrets (an API key the task needs).** Declare it in the task's `environment/docker-compose.yaml`
98
+ — an overlay harbor applies on top of the image, prebuilt or built — and set it on the hosted
99
+ environment:
100
+
101
+ ```yaml
102
+ # tasks/q1/environment/docker-compose.yaml
103
+ services:
104
+ main:
105
+ environment:
106
+ EXA_API_KEY: ${EXA_API_KEY:?set with gateway bench secrets set}
107
+ ```
108
+
109
+ ```bash
110
+ gateway bench secrets set --env code-bench --key EXA_API_KEY # paste the value, then Ctrl-D; never on the command line
111
+ gateway bench secrets list --env code-bench
112
+ ```
113
+
114
+ `${NAME:?…}` (and a bare `${NAME}`) is a requirement; `${NAME:-default}` is not. `validate` lists a
115
+ task's required names as `secrets …`; `run` refuses before dispatch when one is not set on the
116
+ environment, naming the exact `bench secrets set` command. `gateway secret …` is a different store
117
+ (the platform's secret vault for agents and hooks) and does **not** feed hosted harbor runs.
118
+ `run_parameters.json` does not either. Locally, `dry-run` exports what your shell holds — see A3.
86
119
 
87
120
  MCP: no twin for a local scaffold; on an existing environment, `env_write_file({ containerId, path,
88
121
  newContent })` for each file, then `env_commit`.
@@ -97,15 +130,18 @@ Expected output:
97
130
 
98
131
  ```
99
132
  ./code-bench: harbor benchmark — runner reads it as harbor — ok
100
- ok hello: image python:3.11-slim; instruction yes; tests/test.sh yes; solution yes
133
+ ok hello: image build tasks/hello/environment/Dockerfile; instruction yes; tests/test.sh yes; solution yes
101
134
  benchmark ok (harbor): 1 task(s): hello; next: gateway benchmarks dry-run ./code-bench
102
135
  ```
103
136
 
104
- Per task: `task.toml` present, an image (`[environment] docker_image` or `environment/Dockerfile`),
105
- `instruction.md` non-empty, `tests/test.sh` present and writing the reward file, `solution/solve.sh`
106
- noted when absent. Tree-level: `pyproject.toml` (a note when absent) and that the runner's own
107
- dispatch rule (`detect_env_kind`) reads the tree as harbor — a `gateway-env.toml` at the root would
108
- make it a world to the runner. Exit 2 on any problem; notes never fail it.
137
+ Per task: `task.toml` present, ONE image source (`[environment] docker_image` → `image <ref> (prebuilt:
138
+ docker_image)`, else `environment/Dockerfile` → `image build …`; both is a problem when the Dockerfile
139
+ adds anything), the compose overlay and its `secrets` when the task has one, `instruction.md`
140
+ non-empty, `tests/test.sh` present and writing the reward file, `solution/solve.sh` noted when absent.
141
+ Tree-level: `pyproject.toml` (a note when absent) and that the runner's own dispatch rule
142
+ (`detect_env_kind`) reads the tree as harbor — a `gateway-env.toml` at the root would make it a world
143
+ to the runner. Exit 2 on any problem; notes never fail it. The closing line counts the tasks and names
144
+ the first five (`24 tasks: a, b, c, d, e, … (19 more)`); `--json` lists them all.
109
145
 
110
146
  ### A3. Dry-run one task
111
147
 
@@ -131,6 +167,16 @@ runner's `harbor run` line, the static check of `tests/test.sh`) followed by
131
167
  `not executed: … is not installed; execution needs Docker — the plan is what would run`, exit 0.
132
168
  `GATEWAY_DOCKER` names the docker binary.
133
169
 
170
+ The image is chosen by harbor's rule (A1): `docker_image` set → `docker pull`, the Dockerfile is not
171
+ built and the plan says so; otherwise `docker build`. A task with `environment/docker-compose.yaml`
172
+ runs through `docker compose` instead of `docker run` — harbor's base for `main` (the image, the
173
+ mounts, written to `<tmp>/docker-compose.base.yaml`) plus the task's file, then `compose down` — so
174
+ its extra services and `${NAME:?…}` secrets are exercised locally. The names must be exported in
175
+ your shell (`export EXA_API_KEY=…`); a missing one is reported as `not executed: the task's
176
+ docker-compose file requires EXA_API_KEY … export EXA_API_KEY here for the dry-run. Hosted runs read
177
+ it from … gateway bench secrets set --env code-bench --key EXA_API_KEY`, exit 0, nothing spawned.
178
+ Values never appear in the printed argv.
179
+
134
180
  ### A4. Push
135
181
 
136
182
  ```bash
@@ -146,6 +192,18 @@ tree under a world's slug (exit 2, nothing sent); the hub's versions door refuse
146
192
  `benchmark_substrate_mismatch`. Later pushes are new content-hashed versions (`gateway bench log`,
147
193
  `diff`, `branch`, `tag`).
148
194
 
195
+ **Which slug.** stderr says where it came from: `slug code-bench (from benchmark.json)`. Order:
196
+ `--slug` > the dir's `.gateway/.env-metadata.json` (written by every push and pull, so a dir keeps
197
+ pushing to the slug it became) > `benchmark.json` > `pyproject.toml [project].name` > the directory
198
+ name. A tracked slug that disagrees with what the tree declares is refused with exit 2 and nothing
199
+ sent — `./vendor tracks slug websearch-bench-e2e (.gateway/.env-metadata.json, from an earlier push)
200
+ but its benchmark.json says websearch-bench` — because that is how one `--slug X` push made every
201
+ later bare push land on X. Settle it with `--slug <the one you mean>` (the dir retracks) or delete the
202
+ metadata file. The platform's collision suffix (`code-bench-2` minted for a taken `code-bench`) is
203
+ not a disagreement. A 409 `Branch "main" has moved` (another push landed on the slug since this dir's
204
+ base) names its way out: `gateway benchmarks push ./code-bench --force` moves the branch anyway;
205
+ `gateway bench pull code-bench` into a fresh dir builds on the tip.
206
+
149
207
  MCP: `create_environment_version` (or `env_write_file` + `env_commit`).
150
208
 
151
209
  ### A5. Run: models × tasks × rollouts
@@ -76,8 +76,10 @@ Kinds: `person_name`, `email`, `username`, `phone`, `company`, `domain`, `ip`, `
76
76
  `machine`, `log`; anything else gets a `REF-nnnn` stand-in. `identity` ties attributes to one
77
77
  person so they stay consistent; `role: "public"` is never replaced. Every variant of a value is
78
78
  tracked (bare surname, handle spellings, `(at)` emails, phone formats, percent-encoding, a
79
- world's `sha256:` stored form); what slips through is counted (`leak_through_inbound`), and a
80
- real value in the outgoing report fails the run (`leaks_outbound`).
79
+ world's `sha256:` stored form), and in any case: `Alexr` or `ALEXR` for a seeded `alexr` is swapped
80
+ where it stands alone, and the stand-in keeps the case it arrived in. The door swaps and never
81
+ blocks. What slips through is counted (`leak_through_inbound`), and a real value in the outgoing
82
+ report fails the run (`leaks_outbound`).
81
83
 
82
84
  ## Identities that arrive, not seeded
83
85
 
@@ -127,6 +129,63 @@ spelling back in a JSON or text file; a `[redacted …]` token stays — redacti
127
129
  value, and its raw copy is a `record` sink placed before the `redact` rule. The legend maps
128
130
  personas to people: keep it with the run key, out of version control, never beside world data.
129
131
 
132
+ ## Did anything real get through? (the audit)
133
+
134
+ The door's own meters only count values it knew about, so they read zero for anything it never
135
+ recognised. After a run, the audit checks from the other side: it compares what your tools
136
+ **returned** with what the model was **shown**, and lists every real identifier that survived.
137
+
138
+ Write both sides with two `record` rules: one beside the inbound `alias` rule (raw), one after it
139
+ (shown). The outbound `alias` rule stays as you have it, so the tool is called with real values.
140
+
141
+ ```toml
142
+ [[hooks.pre_tool_call]]
143
+ matcher = "*"
144
+ handlers = [{ type = "alias" }]
145
+
146
+ [[hooks.post_tool_call]]
147
+ matcher = "*"
148
+ handlers = [{ type = "record", sink = "{run_dir}/raw.jsonl" }, { type = "alias" }]
149
+
150
+ [[hooks.post_tool_call]]
151
+ matcher = "*"
152
+ handlers = [{ type = "record", sink = "{run_dir}/shown.jsonl" }]
153
+ ```
154
+
155
+ ```python
156
+ from gatewaysdk.hooks import RedactionAudit
157
+
158
+ audit = RedactionAudit.for_run(
159
+ legend=run.alias.legend(), # your subjects and found personas
160
+ field_kinds={"*": {"full_name": "person_name", "email": "email", "handle": "username"}},
161
+ )
162
+ report = audit.run(raw_records, shown_records) # lists of the JSONL lines
163
+ report.ok # False: something real reached the model
164
+ report.to_dict() # findings, census, detectors, coverage
165
+ report.review_sample(10, seed=1) # unflagged text for a person to read
166
+ ```
167
+
168
+ TypeScript is the same: `RedactionAudit.forRun({ legend, fieldKinds }).run(raw, shown)`.
169
+
170
+ From the command line, when the run directory holds `raw.jsonl`, `shown.jsonl` and `legend.json`:
171
+
172
+ ```bash
173
+ gateway hooks audit runs/<run id>/ --field-kinds fields.json --sample 10 # exit 1 when anything leaked
174
+ gateway hooks audit runs/<run id>/ --json # the same report, for a script
175
+ ```
176
+
177
+ `fields.json` is `{"*": {"full_name": "person_name", "email": "email"}}`, or keyed by tool name.
178
+
179
+ It knows a value is real from three places, strongest first: your legend; every value of a field
180
+ whose kind you declared (a name in a vendor row that reappears in a web snippet is caught for
181
+ certain); and what a pattern finds (`email`, `phone`, `ip`). A finding gives the kind, the tool, the
182
+ record and a 3-character prefix, never the value.
183
+
184
+ Read `ok` as "no **detected** leaks with these detectors", not as proof: a name that is in no list
185
+ and matches no pattern is invisible to it. That is what `review_sample` is for. When you find a
186
+ miss, add it to the task's `entities` or `redact` values and run again; the raw file is real data,
187
+ keep it with the legend.
188
+
130
189
  ## Wiring
131
190
 
132
191
  - **Agent SDK** — open the file (`from gatewaysdk.hooks import HookRun; run = HookRun("gateway-hooks.toml")`
@@ -208,7 +208,10 @@ gateway worlds describe <WORLD>; gateway worlds data query <WORLD> <entity> --li
208
208
  **The drive.** Every command here takes `drive:<path>` where it takes a local file: `gateway files
209
209
  upload rows.json --to seeds/rows.json` then `gateway worlds data import <WORLD> drive:seeds/rows.json`
210
210
  is the route from a laptop (or an agent's `request_file_upload`) to a world with nothing staged
211
- locally; the report keeps the drive path under `inputs`. `--to <path>` on `calls pull`, `extract`
211
+ locally: the platform reads the drive file itself (`.json`, `.jsonl`, `.ndjson`, optionally `.gz`;
212
+ convert CSV first) and stages it as one chunk — nothing comes down to your machine and nothing is
213
+ re-uploaded, so `drive:` needs a world slug and no `--dry-run`. The report keeps the drive path under
214
+ `inputs` and the manifest records the digest, size and rows the platform derived. `--to <path>` on `calls pull`, `extract`
212
215
  and `session export` lands the output in the drive (`--from drive:<path>` on `calls pull` and
213
216
  `ingest` reads it back). `gateway traces export drive:<path> --observations --type TOOL` puts a
214
217
  trace export there for the same route. `gateway files --help` lists the drive commands.
@@ -132,7 +132,7 @@ definitions and handler bodies are hand-written:
132
132
 
133
133
  ```bash
134
134
  gateway worlds schema entity add ./x crm_accounts --fields "id:string,name:string,tier:string?" --key id
135
- gateway worlds schema entity add ./x crm_contacts --ref CrmContact --key id --relationship account_id->crm_accounts
135
+ gateway worlds schema entity add ./x crm_contacts --ref CrmContact --key id --relationship 'account_id->crm_accounts' # quote it: a bare > is a shell redirect
136
136
  gateway worlds connector route add ./x list_accounts --path /v1/accounts --param page --entity crm_accounts --results data
137
137
  gateway worlds connector route add ./x get_account --path /v1/accounts/{id} --entity crm_accounts --filter id=id --single
138
138
  gateway worlds connector route add ./x create_account -X POST --path /v1/accounts --entity crm_accounts --action create --body name=name
@@ -140,8 +140,8 @@ gateway worlds connector route add ./x create_account -X POST --path /v1/account
140
140
 
141
141
  Each refuses (and writes nothing) on a duplicate name, an undeclared entity or `$defs`
142
142
  reference, a key that is not a required string/integer field, or a filter on an unknown
143
- argument. A custom world's first `route add` creates `connector.toml`; `schema init` prints `kind: "schema"`, `api` and
144
- the `next` commands in order (custom: entity add → route add → task add → compile → check → import → serve → `worlds create`).
143
+ argument. A create route mints its single-field primary key (`uuid` for strings, `int` for integers) unless `--body` maps it;
144
+ `--generate field=uuid|int|now` and `--default field=value` fill the rest. A custom world's first `route add` creates `connector.toml`; `schema init` prints `kind: "schema"`, `api` and the `next` commands in order (custom: entity add → route add → task add → compile → check → import → serve → `worlds create`).
145
145
  Manifest `kind` is `schema` (older trees: `mock-tools`). Handlers: `def name(ctx, args, query, body)`; `ctx` reads and writes the store and keeps the clock.
146
146
 
147
147
  MCP: `get_connector_authoring_skill({})` returns the authoring skill and conventions.
@@ -446,10 +446,13 @@ gateway bench fork x --into x-experiment # a new container sh
446
446
  gateway bench diff <base-version-id> <head-version-id> # per-file diff
447
447
  gateway bench propose ./x --title t -m why # a change proposal instead of a push
448
448
  gateway bench log x
449
+ gateway worlds sandbox open x@main # no local checkout: edit the source tree server-side,
450
+ gateway worlds sandbox write|exec|check|commit|close <id> # nothing on disk; `bench sandbox` is the same set
449
451
  ```
450
452
 
451
453
  `worlds run --path` on a clean `bench pull` checkout files runs pinned to the base version;
452
454
  a drifted directory runs unfiled. Push a new version when the change is meant to stick.
455
+ Sandbox twins (`open_world_sandbox` … `close_world_sandbox`): `references/cli-mcp-map.md`.
453
456
 
454
457
  MCP: `env_diff({ containerId })`, `env_propose({ containerId, ... })`,
455
458
  `env_list_proposals`, `env_get_proposal`, `env_apply_proposal`, `env_dismiss_proposal`,
@@ -462,9 +465,8 @@ tasks): `references/ui-map.md`.
462
465
 
463
466
  ## CLI to MCP map
464
467
 
465
- Every CLI step with its MCP twin, the steps that have none, and how the platform (console)
466
- agent reaches the same surface (`run_gateway_cli` for the local rows, `spawn_subagent` for
467
- one world per child): `references/cli-mcp-map.md`.
468
+ Every CLI step with its MCP twin, the steps that have none, and how the platform (console) agent
469
+ reaches the same surface (`run_gateway_cli`, `spawn_subagent`): `references/cli-mcp-map.md`.
468
470
 
469
471
  ## Gotchas
470
472
 
@@ -489,12 +491,10 @@ one world per child): `references/cli-mcp-map.md`.
489
491
  ## Related
490
492
 
491
493
  - `benchmarks-getting-started`: a benchmark is a world with tasks — init, check, push, run (models × tasks × rollouts), results.
492
- - `world-data-ingestion`: rows in, calls out.
493
- - `connector-schema-authoring`: the contract, fidelity shapes, conventions, definition of done.
494
+ - `world-data-ingestion`: rows in, calls out. `connector-schema-authoring`: the contract, fidelity shapes, definition of done.
494
495
  - `documentation/features/connector-worlds.md`: `connector.toml` in full, capture, conform, serving.
495
496
  - `documentation/features/workspace-connectors.md`: the connector object, sharing, MCP surface.
496
497
  - `documentation/features/world-schema.md`: the `gateway-world/1` contract and store API.
497
- - Outside the repository, `connector-worlds.md` and `world-schema.md` ship with the npm package
498
- beside the authoring skill: `gateway worlds skill connector-schema-authoring --install .` lands
499
- them under `.claude/skills/connector-schema-authoring/`.
498
+ - Outside the repository, `connector-worlds.md` and `world-schema.md` ship with the npm package beside
499
+ the authoring skill: `gateway worlds skill connector-schema-authoring --install .` installs them.
500
500
  - Consumer pages: `gateway_documentation/pages/worlds/{getting-started,data,custom-connections,slack}.mdx`.
@@ -27,16 +27,32 @@ Reference for `worlds-getting-started`: every CLI step and its MCP twin, then ho
27
27
  | `files connectors list\|sync <id>` | `list_file_connectors` / `sync_file_connector` (`GET\|POST /api/public/files/connectors[/sync]`) |
28
28
  | `traces list\|get\|export [--observations]` | `get_trace`, `get_observations` (`GET /api/public/traces[/{id}]`, `/observations`) |
29
29
  | `traces sessions list\|get [--io]` | `get_session` (`GET /api/public/sessions[/{id}?includeIO=true]`) |
30
- | `worlds data … drive:<path>` / `--to <path>` (any data command) | `get_file` + `import_world_data` (the CLI downloads, then runs the command as on a local file) |
30
+ | `worlds data import <slug> drive:<path>` | `POST /api/public/worlds/{slug}/data/batches/{id}/chunks` with `{ordinal, drive}` (the platform stages the drive file itself; nothing is downloaded or re-uploaded) |
31
+ | `worlds data … drive:<path>` / `--to <path>` (other data commands) | `get_file` + the command (the CLI downloads, then runs the command as on a local file) |
31
32
  | `environment run` / `worlds run` | `run_evaluation` |
32
33
  | `environment link-task-set` / `create-task` | `create_task_set`, `create_task`, `link_task_set` |
33
34
  | `bench diff` / `propose` | `env_diff`, `env_propose` |
35
+ | `worlds\|bench sandbox open\|status\|read\|write\|exec\|check\|commit\|close` | `open_world_sandbox`, `get_world_sandbox`, `read_world_sandbox_file`, `write_world_sandbox_file` (base64), `exec_in_world_sandbox`, `check_world_sandbox`, `commit_world_sandbox`, `close_world_sandbox` (`/api/public/benchmark-containers/sandboxes[/{id}[/file\|exec\|check\|commit]]`; the source tree only — no data snapshot, nothing on the caller's disk; `commit` lands on the recorded base) |
34
36
  | `worlds skill <name>` / `worlds connector skill` | `get_connector_authoring_skill`, `guide` |
35
37
  | `worlds connector capture --remote` / poll + download | `start_connector_capture`, `get_connector_capture` (platform agent only) |
36
38
 
37
39
  Steps with no MCP tool: opening a session (`worlds session open`) — use the CLI or the public
38
40
  sessions route.
39
41
 
42
+ ## Size and time limits (one source: `packages/shared/src/world/limits.ts`)
43
+
44
+ Every door enforces the same number and names it in the refusal; the CLI checks the file
45
+ before it makes a request.
46
+
47
+ | Door | Limit | Over it |
48
+ |---|---|---|
49
+ | `sandbox write`, `write_world_sandbox_file`, `env_write_file`, `propose_task_fix` | 256 KB per file | refused, naming the limit; rows → `worlds data import <slug> <file>`; a big source file → `gateway files upload`, then fetch it inside the sandbox with `sandbox exec`; the whole tree → `bench push` |
50
+ | `sandbox exec --timeout`, `exec_in_world_sandbox timeoutMs` | default 60 s, at most 600 s | refused, not clamped |
51
+ | Sandbox tree at `check`/`commit`, `bench push` unpacked | 128 MB of source | refused, naming the limit; move data to `worlds data import`, delete scratch files |
52
+ | `bench push` / `worlds create --from` compressed bundle | 64 MB (16 MB from the browser) | refused, naming the limit |
53
+ | A single manifest / metadata file in a bundle | 1 MB | refused |
54
+ | `sandbox write` HTTP body | 1 MB (framework cap; only reachable by a body no schema would accept) | 413 |
55
+
40
56
  ## The platform agent
41
57
 
42
58
  The platform (console) agent has every MCP tool in this table as an agent tool of the
@@ -44,11 +44,7 @@ from datetime import datetime, timezone
44
44
  from pathlib import Path
45
45
  from typing import Any, Iterable
46
46
 
47
- try:
48
- import tomllib
49
- except ModuleNotFoundError: # pragma: no cover - python 3.10
50
- import tomli as tomllib
51
-
47
+ from .._toml import tomllib
52
48
  from .compiler import Contract
53
49
  from .store import SQLiteStore
54
50
  from .validation import ContractError, canonical, fail, read_json
@@ -47,11 +47,7 @@ from dataclasses import dataclass, field
47
47
  from pathlib import Path
48
48
  from typing import Any
49
49
 
50
- try:
51
- import tomllib
52
- except ModuleNotFoundError: # pragma: no cover - python 3.10
53
- import tomli as tomllib
54
-
50
+ from .._toml import tomllib
55
51
  from .. import _bundle
56
52
  from .._materialize import materialize_template
57
53
  from .._world import World
@@ -32,11 +32,7 @@ from dataclasses import dataclass
32
32
  from pathlib import Path
33
33
  from typing import Any
34
34
 
35
- try:
36
- import tomllib
37
- except ModuleNotFoundError: # pragma: no cover - python < 3.11
38
- import tomli as tomllib # type: ignore[no-redef]
39
-
35
+ from .._toml import tomllib
40
36
  from .validation import fail
41
37
 
42
38
  STATIC_SOURCE = Path(__file__).with_name("ui_static")
@@ -64,7 +64,6 @@ import shutil
64
64
  import subprocess
65
65
  import sys
66
66
  import tempfile
67
- import tomllib
68
67
  import urllib.error
69
68
  import urllib.parse
70
69
  import urllib.request
@@ -72,6 +71,7 @@ from dataclasses import asdict, dataclass, field
72
71
  from pathlib import Path
73
72
  from typing import Any
74
73
 
74
+ from .._toml import tomllib
75
75
  from .api import load_api_settings
76
76
  from .conform import PIN, _credentials
77
77
  from .connector import load_spec
@@ -15,14 +15,20 @@ dependencies.
15
15
 
16
16
  from __future__ import annotations
17
17
 
18
+ import hashlib
18
19
  import os
19
20
  import re
20
- import tomllib
21
- from collections.abc import Iterator
21
+ from collections.abc import Callable, Iterator
22
22
  from dataclasses import dataclass
23
23
  from functools import cached_property
24
24
  from pathlib import Path
25
25
  from typing import ClassVar
26
+ from ._toml import tomllib
27
+
28
+ try: # optional: only consulted when a tree ships a root .gitignore
29
+ from gitignore_parser import parse_gitignore
30
+ except ImportError: # pragma: no cover - the runner and the sandbox vendor this module alone
31
+ parse_gitignore = None # type: ignore[assignment]
26
32
 
27
33
  # ── Markers (twin of packages/shared/src/world/markers.ts) ──────────────────
28
34
 
@@ -43,6 +49,10 @@ KIND_VERIFIERS = "verifiers"
43
49
  KIND_VERIFIERS_V1 = "verifiers-v1"
44
50
 
45
51
  _SKIPPED_DIRS = frozenset({"dist", "build", "outputs", "__pycache__"})
52
+ # A root Gateway marker widens the hashed root from Prime's source globs to
53
+ # every ordinary root file: those kinds keep executable contract at the root.
54
+ _BROAD_ROOT_MARKERS = frozenset({WORLD_MANIFEST, API_BUNDLE_MANIFEST})
55
+ _HASH_CHUNK = 1 << 20
46
56
  _HARBOR_TASK = re.compile(r"^(?:tasks/)?([^/]+)/task\.toml$")
47
57
 
48
58
  # The `verifiers.v1` namespace first shipped in verifiers 0.2.0, so a v1
@@ -71,6 +81,35 @@ def _skipped_dir(name: str) -> bool:
71
81
  return name.startswith(".") or name in _SKIPPED_DIRS or name.endswith(".egg-info")
72
82
 
73
83
 
84
+ def _ordinary_root_file(name: str, broad: bool) -> bool:
85
+ return broad or name in ("README.md", PYPROJECT) or name in _BROAD_ROOT_MARKERS or name.endswith(".py")
86
+
87
+
88
+ def is_hashed_path(relative: str, broad_root: bool) -> bool:
89
+ """Twin of ``worldTree.ts isHashedPath``: the path-only half of the archive
90
+ rule (hidden names, pruned directories, the root globs)."""
91
+ parts = relative.split("/")
92
+ name = parts[-1]
93
+ if name.startswith(".") or "__pycache__" in parts:
94
+ return False
95
+ directories = parts[:-1]
96
+ if not directories:
97
+ return _ordinary_root_file(name, broad_root)
98
+ return not any(_skipped_dir(component) for component in directories)
99
+
100
+
101
+ def _gitignore_matcher(root: Path) -> Callable[[str], bool] | None:
102
+ gitignore = root / ".gitignore"
103
+ if not gitignore.exists():
104
+ return None
105
+ if parse_gitignore is None:
106
+ raise RuntimeError(
107
+ "A .gitignore is present but 'gitignore_parser' is not installed. Install it (a declared "
108
+ "dependency of gatewaysdk) so the content hash stays compatible with the hub."
109
+ )
110
+ return parse_gitignore(str(gitignore), base_dir=str(root))
111
+
112
+
74
113
  def pyproject_pins_v1(path: str, text: str) -> bool:
75
114
  if path != PYPROJECT:
76
115
  return False
@@ -179,6 +218,83 @@ class WorldTree:
179
218
  def has_connector(self) -> bool:
180
219
  return self.has(self.world_path(CONNECTOR_SPEC))
181
220
 
221
+ # ── The archive: the file set a version IS ───────────────────────────────
222
+ #
223
+ # Vendored from Prime Intellect's ``prime`` CLI (MIT; THIRD_PARTY_LICENSES)
224
+ # so a container hashed here resolves to the same logical version on their
225
+ # Environments Hub: SHA-256 over ``file:<relpath>`` + raw bytes, in sorted
226
+ # path order, over README.md / pyproject.toml / *.py at the root plus every
227
+ # non-hidden nested file outside dist, build, outputs, __pycache__ and
228
+ # *.egg-info. A root Gateway marker widens the root to every ordinary file
229
+ # (the marker itself must be hashed or a manifest-only edit dedups into the
230
+ # previous version and is silently discarded). Symlinks and ``.gitignore``-d
231
+ # paths are dropped. This is the version primary key, NOT a tarball digest.
232
+ # Twin: ``packages/shared/src/world/worldTree.ts`` (``isHashedPath`` +
233
+ # ``contentHash``); the corpus under ``conformance/worlds/`` pins both.
234
+
235
+ @property
236
+ def broad_root(self) -> bool:
237
+ return any(self.has(marker) for marker in _BROAD_ROOT_MARKERS)
238
+
239
+ def archive_files(self) -> list[Path]:
240
+ """The hashed files, absolute, in sorted relative-path order."""
241
+ ignored = _gitignore_matcher(self.root)
242
+ broad = self.broad_root
243
+ found: dict[str, Path] = {}
244
+
245
+ def keep(path: Path) -> None:
246
+ if not path.is_file() or path.is_symlink():
247
+ return
248
+ relative = path.relative_to(self.root).as_posix()
249
+ if not is_hashed_path(relative, broad):
250
+ return
251
+ if ignored is not None and ignored(str(path)):
252
+ return
253
+ found[relative] = path
254
+
255
+ def pruned(directory: Path) -> bool:
256
+ """A symlinked directory is not part of the tree: it can point anywhere."""
257
+ return (
258
+ directory.is_symlink()
259
+ or _skipped_dir(directory.name)
260
+ or (ignored is not None and ignored(str(directory)))
261
+ )
262
+
263
+ for entry in sorted(self.root.iterdir(), key=lambda p: p.name):
264
+ if entry.is_file():
265
+ keep(entry)
266
+ continue
267
+ if not entry.is_dir() or pruned(entry):
268
+ continue
269
+ for dirpath, dirnames, filenames in os.walk(entry):
270
+ here = Path(dirpath)
271
+ dirnames[:] = sorted(name for name in dirnames if not pruned(here / name))
272
+ for filename in sorted(filenames):
273
+ keep(here / filename)
274
+ return [found[relative] for relative in sorted(found)]
275
+
276
+ def hashed_paths(self) -> list[str]:
277
+ return [path.relative_to(self.root).as_posix() for path in self.archive_files()]
278
+
279
+ @property
280
+ def content_hash(self) -> str:
281
+ """Streams every file; a large tree never has to fit in memory.
282
+
283
+ An unreadable file raises: a digest that silently omitted it would
284
+ name a tree the manifest and archive do not describe.
285
+ """
286
+ digest = hashlib.sha256()
287
+ for path in self.archive_files():
288
+ relative = path.relative_to(self.root).as_posix()
289
+ digest.update(f"file:{relative}".encode("utf-8"))
290
+ try:
291
+ with open(path, "rb") as handle:
292
+ for chunk in iter(lambda: handle.read(_HASH_CHUNK), b""):
293
+ digest.update(chunk)
294
+ except OSError as error:
295
+ raise OSError(f"cannot read {relative} while hashing the tree: {error}") from error
296
+ return digest.hexdigest()
297
+
182
298
  @cached_property
183
299
  def manifest(self) -> WorldManifest:
184
300
  path = self.world / WORLD_MANIFEST
@@ -204,8 +320,8 @@ class WorldTree:
204
320
  def v1_evidence(self) -> Iterator[tuple[str, str]]:
205
321
  """``(relpath, text)`` of the files a verifiers package's API version is
206
322
  read from: the root pyproject and every ``.py`` the archive rules keep
207
- (``benchmark_hub._archive``: no hidden, build, dist, outputs, pycache or
208
- egg-info directories)."""
323
+ (``archive_files``: no hidden, build, dist, outputs, pycache or egg-info
324
+ directories)."""
209
325
  for dirpath, dirnames, filenames in os.walk(self.root):
210
326
  dirnames[:] = sorted(name for name in dirnames if not _skipped_dir(name))
211
327
  for filename in sorted(filenames):
@@ -70,9 +70,9 @@ Expected output (JSON): `"kind": "harbor"`, `"tasks": ["hello"]`, the files, and
70
70
  code-bench/
71
71
  benchmark.json # kind, models, rollouts: the defaults `benchmarks run` reads
72
72
  pyproject.toml # [project] name + version: the slug and the version `push` records; not a Python package
73
- tasks/hello/task.toml # version, [metadata], [agent] timeout_sec, [verifier] timeout_sec, [environment] docker_image
73
+ tasks/hello/task.toml # version, [metadata], [agent] timeout_sec, [verifier] timeout_sec; docker_image only as a comment
74
74
  tasks/hello/instruction.md # what the agent is told
75
- tasks/hello/environment/Dockerfile # the task's container (FROM docker_image; WORKDIR /app)
75
+ tasks/hello/environment/Dockerfile # the task's container, built by the runner (FROM --image; WORKDIR /app): the ONE image source
76
76
  tasks/hello/tests/test.sh # the grader: runs after the agent, writes /logs/verifier/reward.json
77
77
  tasks/hello/solution/solve.sh # the reference solution; must score 1.0 (harbor's oracle agent runs it)
78
78
  ```
@@ -82,7 +82,40 @@ The reward contract is the only thing the platform reads: `tests/test.sh` writes
82
82
  other key a sub-reward — or a bare number to `/logs/verifier/reward.txt`. Any language, any method.
83
83
 
84
84
  Add a task: copy `tasks/hello` to `tasks/<name>`, edit `instruction.md`, `tests/test.sh`,
85
- `solution/solve.sh`, and the image (`docker_image` or the Dockerfile). Two tasks minimum.
85
+ `solution/solve.sh`, and the image. Two tasks minimum.
86
+
87
+ **One image source.** Harbor's rule (`should_use_prebuilt_docker_image`): when `task.toml` names
88
+ `[environment] docker_image`, the runner pulls that prebuilt image and runs it as it is —
89
+ `environment/Dockerfile` is **never built**, so anything it installs is absent on the runner (a tool the
90
+ agent needs, a pip package) while a Dockerfile-first local run passes. Without `docker_image` the
91
+ runner builds the Dockerfile. The scaffold therefore writes only the Dockerfile; put what the agent
92
+ needs in it. To run a prebuilt image instead, delete the Dockerfile and set `docker_image`.
93
+ `validate` fails a task that has both when the Dockerfile adds anything beyond `FROM`/`WORKDIR`
94
+ (a note when it does not), and `dry-run` follows the runner's choice (`image … (prebuilt: docker_image)`,
95
+ `pull`, no `build`).
96
+
97
+ **Secrets (an API key the task needs).** Declare it in the task's `environment/docker-compose.yaml`
98
+ — an overlay harbor applies on top of the image, prebuilt or built — and set it on the hosted
99
+ environment:
100
+
101
+ ```yaml
102
+ # tasks/q1/environment/docker-compose.yaml
103
+ services:
104
+ main:
105
+ environment:
106
+ EXA_API_KEY: ${EXA_API_KEY:?set with gateway bench secrets set}
107
+ ```
108
+
109
+ ```bash
110
+ gateway bench secrets set --env code-bench --key EXA_API_KEY # paste the value, then Ctrl-D; never on the command line
111
+ gateway bench secrets list --env code-bench
112
+ ```
113
+
114
+ `${NAME:?…}` (and a bare `${NAME}`) is a requirement; `${NAME:-default}` is not. `validate` lists a
115
+ task's required names as `secrets …`; `run` refuses before dispatch when one is not set on the
116
+ environment, naming the exact `bench secrets set` command. `gateway secret …` is a different store
117
+ (the platform's secret vault for agents and hooks) and does **not** feed hosted harbor runs.
118
+ `run_parameters.json` does not either. Locally, `dry-run` exports what your shell holds — see A3.
86
119
 
87
120
  MCP: no twin for a local scaffold; on an existing environment, `env_write_file({ containerId, path,
88
121
  newContent })` for each file, then `env_commit`.
@@ -97,15 +130,18 @@ Expected output:
97
130
 
98
131
  ```
99
132
  ./code-bench: harbor benchmark — runner reads it as harbor — ok
100
- ok hello: image python:3.11-slim; instruction yes; tests/test.sh yes; solution yes
133
+ ok hello: image build tasks/hello/environment/Dockerfile; instruction yes; tests/test.sh yes; solution yes
101
134
  benchmark ok (harbor): 1 task(s): hello; next: gateway benchmarks dry-run ./code-bench
102
135
  ```
103
136
 
104
- Per task: `task.toml` present, an image (`[environment] docker_image` or `environment/Dockerfile`),
105
- `instruction.md` non-empty, `tests/test.sh` present and writing the reward file, `solution/solve.sh`
106
- noted when absent. Tree-level: `pyproject.toml` (a note when absent) and that the runner's own
107
- dispatch rule (`detect_env_kind`) reads the tree as harbor — a `gateway-env.toml` at the root would
108
- make it a world to the runner. Exit 2 on any problem; notes never fail it.
137
+ Per task: `task.toml` present, ONE image source (`[environment] docker_image` → `image <ref> (prebuilt:
138
+ docker_image)`, else `environment/Dockerfile` → `image build …`; both is a problem when the Dockerfile
139
+ adds anything), the compose overlay and its `secrets` when the task has one, `instruction.md`
140
+ non-empty, `tests/test.sh` present and writing the reward file, `solution/solve.sh` noted when absent.
141
+ Tree-level: `pyproject.toml` (a note when absent) and that the runner's own dispatch rule
142
+ (`detect_env_kind`) reads the tree as harbor — a `gateway-env.toml` at the root would make it a world
143
+ to the runner. Exit 2 on any problem; notes never fail it. The closing line counts the tasks and names
144
+ the first five (`24 tasks: a, b, c, d, e, … (19 more)`); `--json` lists them all.
109
145
 
110
146
  ### A3. Dry-run one task
111
147
 
@@ -131,6 +167,16 @@ runner's `harbor run` line, the static check of `tests/test.sh`) followed by
131
167
  `not executed: … is not installed; execution needs Docker — the plan is what would run`, exit 0.
132
168
  `GATEWAY_DOCKER` names the docker binary.
133
169
 
170
+ The image is chosen by harbor's rule (A1): `docker_image` set → `docker pull`, the Dockerfile is not
171
+ built and the plan says so; otherwise `docker build`. A task with `environment/docker-compose.yaml`
172
+ runs through `docker compose` instead of `docker run` — harbor's base for `main` (the image, the
173
+ mounts, written to `<tmp>/docker-compose.base.yaml`) plus the task's file, then `compose down` — so
174
+ its extra services and `${NAME:?…}` secrets are exercised locally. The names must be exported in
175
+ your shell (`export EXA_API_KEY=…`); a missing one is reported as `not executed: the task's
176
+ docker-compose file requires EXA_API_KEY … export EXA_API_KEY here for the dry-run. Hosted runs read
177
+ it from … gateway bench secrets set --env code-bench --key EXA_API_KEY`, exit 0, nothing spawned.
178
+ Values never appear in the printed argv.
179
+
134
180
  ### A4. Push
135
181
 
136
182
  ```bash
@@ -146,6 +192,18 @@ tree under a world's slug (exit 2, nothing sent); the hub's versions door refuse
146
192
  `benchmark_substrate_mismatch`. Later pushes are new content-hashed versions (`gateway bench log`,
147
193
  `diff`, `branch`, `tag`).
148
194
 
195
+ **Which slug.** stderr says where it came from: `slug code-bench (from benchmark.json)`. Order:
196
+ `--slug` > the dir's `.gateway/.env-metadata.json` (written by every push and pull, so a dir keeps
197
+ pushing to the slug it became) > `benchmark.json` > `pyproject.toml [project].name` > the directory
198
+ name. A tracked slug that disagrees with what the tree declares is refused with exit 2 and nothing
199
+ sent — `./vendor tracks slug websearch-bench-e2e (.gateway/.env-metadata.json, from an earlier push)
200
+ but its benchmark.json says websearch-bench` — because that is how one `--slug X` push made every
201
+ later bare push land on X. Settle it with `--slug <the one you mean>` (the dir retracks) or delete the
202
+ metadata file. The platform's collision suffix (`code-bench-2` minted for a taken `code-bench`) is
203
+ not a disagreement. A 409 `Branch "main" has moved` (another push landed on the slug since this dir's
204
+ base) names its way out: `gateway benchmarks push ./code-bench --force` moves the branch anyway;
205
+ `gateway bench pull code-bench` into a fresh dir builds on the tip.
206
+
149
207
  MCP: `create_environment_version` (or `env_write_file` + `env_commit`).
150
208
 
151
209
  ### A5. Run: models × tasks × rollouts