Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
73 commits
Select commit Hold shift + click to select a range
058a57f
feat(agent): integrate tool plugins and effects
fffonion Oct 18, 2026
fb610b1
fix(plugins): backtrack resolver choices within fixed work limits
fffonion Oct 18, 2026
d83dbaf
test(auth): fence TLS fixture admission on native pump teardown
fffonion Oct 18, 2026
5f3f72f
fix(gateway): synchronize persistence gate predicates
fffonion Oct 18, 2026
ad9e470
feat(effects): complete managed synchronous publication and recovery
fffonion Oct 18, 2026
e699c73
fix(effects): honor directory durability and invocation/startup budgets
fffonion Oct 18, 2026
3f66a5b
test: make managed workspace Native fixtures explicitly private
fffonion Oct 19, 2026
9b9812d
docs: document managed workspace configuration
fffonion Oct 19, 2026
ac947e3
test: derive managed gateway scratch env from assigned root
fffonion Oct 19, 2026
fce6111
Enforce captured managed authority at Native boundaries
fffonion Oct 19, 2026
d702239
test: pin the published SQLite SDK and cover legal triggers
fffonion Oct 19, 2026
007da8b
fix: validate managed documents without an identity map_err
fffonion Oct 19, 2026
ec81c25
test: compile lease_existing_for_test only with its caller
fffonion Oct 19, 2026
2f2a14e
test: expect the reviewed SDK pin b5066dd92fcd7c07df57e3ec2c29dbcf8c1…
fffonion Oct 19, 2026
49679c4
docs: align deployment guide with the reviewed SDK pin
fffonion Oct 19, 2026
3938fe5
fix: refuse provider calls cancelled after the request commit
fffonion Oct 19, 2026
0a8db1c
feat(effects): publish reviewed graph and managed workspace slices
fffonion Nov 2, 2026
c926734
feat(plugins): add reviewed versioned state lifecycle
fffonion Nov 2, 2026
3aabb9b
feat(plugins): link pure tool libraries with fresh effect authorization
fffonion Nov 3, 2026
8d3f30e
docs(config): document host plugin control environment
fffonion Nov 3, 2026
41f0ba6
fix(gateway): retain sealed replay proofs across storage outages
fffonion Nov 4, 2026
03ef010
fix(agent): settle state owners before replies and preserve restore d…
fffonion Nov 5, 2026
0e2e6bf
test: prove maintenance lease progress and failure-safe cleanup
fffonion Nov 5, 2026
5f7e1de
feat(plugins): bound one stateless caller-callee effect
fffonion Nov 5, 2026
f5da89f
feat(plugins): admit one tool linked to two pure leaves
fffonion Nov 5, 2026
49f2021
feat(plugins): restore a pure-leaf tool graph from its captured gener…
fffonion Nov 6, 2026
0a0fbee
feat(plugins): admit one stateful callee on a stateless caller
fffonion Nov 6, 2026
c71f0aa
fix(plugins): drop the unused child import re-export
fffonion Nov 6, 2026
3b192fa
test(plugins): restore a pure-leaf checkpoint over the existing socket
fffonion Nov 6, 2026
28e9004
test(plugins): restore a two-leaf checkpoint from its captured genera…
fffonion Nov 6, 2026
fab47e4
test(plugins): refuse a pure-leaf restore for another workspace path
fffonion Nov 6, 2026
724daa8
test(plugins): cut a pure-leaf read before publication
fffonion Nov 6, 2026
2619535
test(plugins): commit one stateful child write on the admitted graph
fffonion Nov 6, 2026
6d57618
feat(plugins): restore one stateful child root from its checkpoint
fffonion Nov 6, 2026
de5a543
test(plugins): refuse a stateful restore for another workspace
fffonion Nov 6, 2026
bc4a6fb
test(plugins): cut a restored stateful read before publication
fffonion Nov 6, 2026
eddb0d8
fix(plugins): resolve a restored branch from its existing receipt
fffonion Nov 6, 2026
d72063c
test(plugins): roll a stateful graph back without dropping its root
fffonion Nov 6, 2026
f56ba90
test(plugins): refuse a stateful rollback when the live schema is newer
fffonion Nov 6, 2026
a126902
test(plugins): keep a referenced state root through orphan collection
fffonion Nov 6, 2026
3739f63
test(plugins): exercise the stateful graph HTTP lifecycle
fffonion Nov 6, 2026
f4ebc6e
test(plugins): read the captured root after a restored-read cut
fffonion Nov 6, 2026
e104c0d
feat(tools): dispatch captured RSS tools through callable routes
fffonion Nov 6, 2026
d1d6c1a
test(effects): reject real foreign branch and publication requests
fffonion Nov 6, 2026
fa4ef8c
test: cover plugin state migration across process crash
fffonion Nov 6, 2026
8c2258b
test: cover concurrent plugin activation and pinned rollback
fffonion Nov 6, 2026
0c61a97
test: exercise real HTTP approvals after cancellation and expiry
fffonion Nov 6, 2026
06d6292
test(plugins): enforce the original clock after child admission
fffonion Nov 6, 2026
1e9b4a9
feat(effects): manage shared process views and fenced finalization
fffonion Nov 6, 2026
17794fb
test(plugins): cover cold graph restore and selection boundaries
fffonion Nov 6, 2026
12456f5
fix(admission): reject unsupported prepared graphs before durable commit
fffonion Nov 6, 2026
17f76b2
fix(effects): propagate native intent cuts without recoverable tool r…
fffonion Nov 6, 2026
3879ec4
fix(plugins): revoke supported solo tool graphs
fffonion Nov 7, 2026
ed0c49c
test(effects): separate durable revoke and immediate cancellation fix…
fffonion Nov 7, 2026
be6be62
test(process): qualify bounded terminal ENOSPC
fffonion Nov 7, 2026
8665e0d
test(schema): prove canonical snapshot mutation causality
fffonion Nov 7, 2026
e553c46
test(effects): align process control authority assertions
fffonion Nov 7, 2026
95a3483
feat(effects): complete managed graph lifecycle and restore
fffonion Nov 7, 2026
b3ccc31
fix(effects): reconcile restore and checkpoint writer lifetimes
fffonion Nov 8, 2026
aed3e0c
fix(plugins): preserve live graph revocation refusals
fffonion Nov 8, 2026
d3c982e
fix(service): reject started unmanaged preflight failures
fffonion Nov 8, 2026
35b7871
fix(effects): map missing branch heads to typed not-found
fffonion Nov 8, 2026
2161dec
docs(config): document captured managed completion policy
fffonion Nov 8, 2026
f9e72b5
test(effects): exercise the reachable two-registration CAS race
fffonion Nov 8, 2026
d45d37d
test(effects): batch native lifecycle fixture operations
fffonion Nov 8, 2026
c1a8083
test(process): compare ENOSPC diagnostics with host dd
fffonion Nov 8, 2026
5a7051a
test(lifecycle): observe second provider deadline boundary
fffonion Nov 8, 2026
47d0aaf
test(lifecycle): report post-return harness error ordinal
fffonion Nov 8, 2026
ff37c1e
test(lifecycle): record caller-owned harness pin failure site
fffonion Nov 8, 2026
94c2ea8
fix(gateway): register SIGINT before publishing readiness
fffonion Nov 8, 2026
f241c33
refactor(agent): remove obsolete code and redundant tests
fffonion Nov 8, 2026
a1ffcb0
docs(agent): consolidate current design and usage guides
fffonion Nov 8, 2026
212fc3b
test(lifecycle): allow deadline child to reach native startup
fffonion Nov 8, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 4 additions & 6 deletions .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -70,12 +70,10 @@ jobs:
- name: Clippy (warnings denied)
run: cargo clippy --locked --all-features --all-targets -- -D warnings

# The default suite runs all non-ignored tests. The only remaining
# #[ignore]d suites are the openai_responses / anthropic_messages
# placeholder references (not_implemented adapter stubs — pure agent
# work, unrelated to the core contract gate); the core repro driver
# and all OpenAI Chat suites run by default — see
# plans/2026-08-13_a3-provider-core-blocker.md.
# The default suite runs all non-ignored tests; child-only re-execution
# helpers, manual live-network cases and not-implemented adapter
# placeholder references stay explicitly ignored. The core repro driver
# and all OpenAI Chat suites run by default.
- name: Test (workspace, all features, all targets)
run: cargo test --locked --all-features --all-targets

Expand Down
13 changes: 10 additions & 3 deletions Cargo.lock

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

7 changes: 5 additions & 2 deletions Cargo.toml
Original file line number Diff line number Diff line change
Expand Up @@ -25,11 +25,12 @@ hyper = { version = "1", features = ["client", "http1"] }
hyper-util = { version = "0.1", features = ["client-legacy", "http1", "tokio"] }
parking_lot = "0.12"
rustls = { version = "0.23", default-features = false, features = ["ring", "std", "tls12"] }
rustscript-vm = { package = "pd-vm", git = "https://github.com/rustscript-lang/rustscript.git", rev = "d9387f98df78dd88d2d1c04eebac4415f844efaf", default-features = false, features = ["runtime", "http-client", "sqlite"] }
pd-host-function = { git = "https://github.com/rustscript-lang/rustscript.git", rev = "d9387f98df78dd88d2d1c04eebac4415f844efaf" }
rustscript-vm = { package = "pd-vm", git = "https://github.com/rustscript-lang/rustscript.git", rev = "b5066dd92fcd7c07df57e3ec2c29dbcf8c19c390", default-features = false, features = ["runtime", "http-client", "sqlite"] }
pd-host-function = { git = "https://github.com/rustscript-lang/rustscript.git", rev = "b5066dd92fcd7c07df57e3ec2c29dbcf8c19c390" }
serde = { version = "1", features = ["derive"] }
serde_json = "1"
serde_yaml = "0.9"
semver = "1.0.28"
yaml-rust2 = { version = "0.12", default-features = false }
# Meta-schema validation only; resolver features stay disabled.
jsonschema = { version = "0.52.1", default-features = false }
Expand All @@ -51,6 +52,8 @@ tracing-core = "0.1"
[features]
default = []
config-fixture = []
# Test-only one-shot simulated I/O faults. Unarmed production builds do not include it.
workspace-io-fault = []

# Optimize the embedded compiler while retaining development checks.
[profile.dev.package.pd-vm]
Expand Down
15 changes: 9 additions & 6 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -39,10 +39,13 @@ RUSTSCRIPT_AGENT_STATE_DB=/var/lib/rustscript-agent/state.db \
cargo run --release --bin rustscript-agent-gateway
```

When `RUSTSCRIPT_AGENT_SOURCE` is set it is the authoritative agent source
(`bundled:coding`, or `file:` plus an absolute path); omit it to run the
bundled coding entry. The deprecated `RUSTSCRIPT_AGENT_SCRIPT` is read only as
a fallback. Full environment-variable reference: [docs/configuration.md](docs/configuration.md).
A present `config.yaml` `agent.source` is the authoritative agent source; a
present-but-invalid config source fails closed with no environment fallback.
Only when the config is missing does the gateway fall back to
`RUSTSCRIPT_AGENT_SOURCE`, then the deprecated `RUSTSCRIPT_AGENT_SCRIPT`, then
the bundled coding entry. A valid source is `bundled:coding`, or `file:` plus
an absolute path. Full environment-variable reference:
[docs/configuration.md](docs/configuration.md).
Deployment, systemd/container examples, shutdown and backup: [docs/deployment.md](docs/deployment.md).

## Status
Expand All @@ -57,7 +60,7 @@ items are explicit, and no placeholder route is advertised.
| Single-run runner (`rustscript-agent --script … --allow-host …`) | Implemented |
| Gateway HTTP API: sessions, runs, SSE events, stop, jobs CRUD, subagent interrupt | Implemented (see [plans/2026-07-30_rustscript-agent-gateway-api.md](plans/2026-07-30_rustscript-agent-gateway-api.md)) |
| Durable SQLite state: sessions/messages/runs/events/jobs/approvals/compactions, restart recovery | Implemented |
| Gateway entry selection | `RUSTSCRIPT_AGENT_SOURCE` is authoritative; deprecated `RUSTSCRIPT_AGENT_SCRIPT` is a fallback; with neither set, RSS selects `bundled:coding`. The gateway does not yet select its entry from `config.yaml`'s `agent.source`. |
| Gateway entry selection | A present `config.yaml` `agent.source` is authoritative; a missing config falls back to `RUSTSCRIPT_AGENT_SOURCE`, then the deprecated `RUSTSCRIPT_AGENT_SCRIPT`, then bundled `bundled:coding`. |
| Legacy chat completion path (`/api/sessions/{id}/chat`) | Legacy-constructor-only; production gateway selection via `RUSTSCRIPT_AGENT_SOURCE`, legacy `RUSTSCRIPT_AGENT_SCRIPT`, or bundled `bundled:coding` returns `501 agent_source_not_configured`. Normal coded-agent runs use `POST /v1/runs`. |
| OpenAI Chat Completions (`openai`, `openai_chat`, or default) | Current source has buffered and streaming requests through the opaque native credential transport; this is source status, not live-provider validation. |
| Codex Responses (`openai-codex` / `codex-responses`) | Access-only source path is present. Full refresh/lifecycle support is not claimed here. |
Expand All @@ -67,7 +70,7 @@ items are explicit, and no placeholder route is advertised.
| Observability (A9): bounded metrics registry, `GET /metrics`, structured terminal tracing | Implemented (see [docs/deployment.md](docs/deployment.md)) |
| RSS coding loop + durable compaction policies (A5) | The gateway's bundled coding entry runs through the agent worker; compaction policies remain implemented and tested (`rss/agent/compact.rss`). |
| RSS coding tools + workspace-confined loop | Implemented for the library worker: `read_file`, `search_files`, `write_file`, `patch`, `terminal`, `process` run serially through `rss/tools/dispatch.rss`. Local E2E: `cargo test --test coding_agent_e2e_tests` and `cargo test --test coding_agent_edge_e2e_tests`. See [docs/configuration.md](docs/configuration.md). Parallel tools remain excluded (A6). |
| Harness and approval machinery (A4) | Not implemented (excluded from the current milestone scope); approval **repository** CRUD exists, there is no approval flow driving runs |
| Approval execution chain (A4) | Implemented for tool dispatch: durable approval records, `ask` policy for write/process, HTTP and Telegram approve/reject, one typed terminal result. A general-purpose harness layer beyond tool approval remains excluded. |
| Parallel tools and subagents (A6) | Not implemented (excluded from the current milestone scope) |
| Scheduled / durable job execution | **Not implemented (explicitly excluded)**. Job CRUD, pause/resume, and latest-output routes exist, but there is no scheduler; `POST /api/jobs/{id}/run` is intentionally absent and answers `404`. |
| Telegram gateway (A8) | **Implemented**: native Bot API transport (https/rustls), deny-by-default account/chat/user allowlists, durable delivery cursors (at-least-once), bounded 429/5xx/401 retries, fail-closed first-boot drain, bounded shutdown drain. See [docs/deployment.md](docs/deployment.md). |
Expand Down
86 changes: 79 additions & 7 deletions docs/configuration.md
Original file line number Diff line number Diff line change
Expand Up @@ -73,6 +73,68 @@ the library's `state.db` path, but the gateway's SQLite selection remains
separate: `RUSTSCRIPT_AGENT_STATE_DB` chooses the gateway state file and an
unset value keeps gateway state in memory.

### Managed workspace (primary-only)

`RUSTSCRIPT_AGENT_MANAGED_WORKSPACE` is an OS filesystem path, read only by
`rustscript-agent-gateway` and passed to `ManagedWorkspaceConfig::for_workspace`.
It has **no deprecated alias** and is not a boolean flag. Unset means
`managed_workspace = None` (direct mode). When set, it must name an existing,
non-root directory using a non-empty absolute UTF-8 path with no NUL or
parent-directory components; invalid configuration fails startup.

Managed mode requires `RUSTSCRIPT_AGENT_STATE_DB` to name an **absolute,
persistent database**; in-memory state is rejected. The runtime derives its
private object root as `<database-parent>/.<database-filename>.effects`.
The source must not overlap the database, that object root, or the sibling
`.<database-filename>.tool-packages` directory in either direction.

The setting alone grants no workspace authority. Import and queries require a
current Native workspace grant admitting the configured source through
`workspaces.allowed_roots`. For the gateway binary, explicitly set
`RUSTSCRIPT_AGENT_HOME` to enable the workspace host, with loadable config/auth
files in a private, process-owned agent home (Unix mode `0700`). A managed run's
admitted canonical workspace must match the configured source. Embeddings
supply this authority through `with_workspace_home` or their trusted workspace
host factory.

`RUSTSCRIPT_AGENT_MANAGED_COMPLETION_POLICY` selects the trusted host
completion policy for managed runs: `staged` (default) or `required`. It may
only be set to `required` together with a managed workspace; any other value
fails startup. This policy affects only the run terminal step: `required`
commits the run's branch publication in the same durable transaction as
`run.completed`. It grants no workspace permission and does not replace the
explicit publish management endpoint or its bearer/token policy.

The source is imported as a seed; managed filesystem mutations use private
working trees without writing changes back to the source. This mechanism
covers managed synchronous local filesystem effects, not OS permission
isolation or rollback of external effects. Tree quotas and retained-object
limits are described with the native field below.

### Host plugin control (primary-only)

The gateway reads these four primary-only variables directly; they have no
`PD_EDGE_AGENT_*` aliases. They configure the host plugin-control binding,
not ordinary run-token authority.

- `RUSTSCRIPT_AGENT_PLUGIN_SOURCE_ROOT`: filesystem path for the controlled
plugin source root passed to the host binding. Default: unset. If it is
unset, this binding is skipped and the other three settings below are not
read by that binding.
- `RUSTSCRIPT_AGENT_PLUGIN_DEPLOYMENT_SPACE`: deployment-space string passed
to the host plugin control. There is no default; it is required when the
source root is set. A missing or non-UTF-8 value fails startup.
- `RUSTSCRIPT_AGENT_PLUGIN_PUBLISHER_DIR`: optional directory of accepted
publisher verification keys. Default: unset, so no keys are loaded from a
directory. Each entry's filename is the publisher ID and its file contents
must be a raw 32-byte verification key, not a hexadecimal or PEM encoding.
An unavailable directory, unreadable entry, invalid filename, or wrong key
length fails startup. No filename extension is stripped.
- `RUSTSCRIPT_AGENT_PLUGIN_UNSIGNED_DEV`: optional filesystem path forwarded
as `InstallAuthorization::unsigned_dev_space`. Default: unset (`None`).
This is a development-space path, not a boolean flag; setting it does not
bypass the install, trust, or effect-policy checks.

### Gateway variables and deprecated aliases

Every gateway `RUSTSCRIPT_AGENT_*` variable in the table has a deprecated
Expand All @@ -84,7 +146,7 @@ do not rely on them.
| Variable | Deprecated alias | Type | Default | Bounds / notes |
| --- | --- | --- | --- | --- |
| `RUSTSCRIPT_AGENT_GATEWAY_ADDR` | `PD_EDGE_AGENT_GATEWAY_ADDR` | string (`SocketAddr`) | `127.0.0.1:8090` | Bind address. An unparsable value fails startup. |
| `RUSTSCRIPT_AGENT_BEARER_TOKEN` | `PD_EDGE_AGENT_BEARER_TOKEN` | string (secret) | unset | Required unless `RUSTSCRIPT_AGENT_ALLOW_ANONYMOUS=1`. A blank value is rejected. Compared in constant time against the token carried in the `Authorization: Bearer ***` header. Treat as a secret; see [Secrets](#secrets). |
| `RUSTSCRIPT_AGENT_BEARER_TOKEN` | `PD_EDGE_AGENT_BEARER_TOKEN` | string (secret) | unset | Required unless `RUSTSCRIPT_AGENT_ALLOW_ANONYMOUS=1`. A blank value is rejected. Compared in constant time against the token carried in the `Authorization: Bearer ***` header. Treat as a secret; see [Secrets](#secrets-and-provider-credentials). |
| `RUSTSCRIPT_AGENT_ALLOW_ANONYMOUS` | `PD_EDGE_AGENT_ALLOW_ANONYMOUS` | flag | unset | Only the exact value `1` enables anonymous access. Local testing only. |
| `RUSTSCRIPT_AGENT_ALLOW_HOSTS` | `PD_EDGE_AGENT_ALLOW_HOSTS` | comma-separated list | empty (deny all) | Every HTTP(S)/WS(S) destination host must be allowlisted. Empty list denies all hosts. |
| `RUSTSCRIPT_AGENT_ALLOW_SCHEMES` | `PD_EDGE_AGENT_ALLOW_SCHEMES` | comma-separated list | `https,wss` | Replaces the default scheme set when set. |
Expand All @@ -101,7 +163,7 @@ do not rely on them.
| `RUSTSCRIPT_AGENT_RATE_LIMIT_MAX_BUCKETS` | `PD_EDGE_AGENT_RATE_LIMIT_MAX_BUCKETS` | integer `usize` | `10000` | Upper bound on tracked buckets (per-IP and per-account combined); at the bound the stalest bucket is evicted, so memory is bounded. |
| `RUSTSCRIPT_AGENT_CLIENT_DISCONNECT_POLICY` | `PD_EDGE_AGENT_CLIENT_DISCONNECT_POLICY` | enum | `keep-running` | `keep-running`: the run survives every subscriber disconnect and events stay replayable by cursor. `cancel-on-disconnect`: the run is cancelled with the typed `client_disconnect` reason when the LAST subscriber disconnects while it is still active. Unknown spellings fail startup. |
| Telegram adapter (A8) |
| `RUSTSCRIPT_AGENT_TELEGRAM_BOT_TOKEN` | `PD_EDGE_AGENT_TELEGRAM_BOT_TOKEN` | string (secret) | unset (adapter disabled) | Bot API token. When set, the Telegram poller starts alongside the API server; a blank value fails startup. Never logged; see [Secrets](#secrets). |
| `RUSTSCRIPT_AGENT_TELEGRAM_BOT_TOKEN` | `PD_EDGE_AGENT_TELEGRAM_BOT_TOKEN` | string (secret) | unset (adapter disabled) | Bot API token. When set, the Telegram poller starts alongside the API server; a blank value fails startup. Never logged; see [Secrets](#secrets-and-provider-credentials). |
| `RUSTSCRIPT_AGENT_TELEGRAM_API_BASE` | `PD_EDGE_AGENT_TELEGRAM_API_BASE` | URL | `https://api.telegram.org` | Bot API base. Must be a bare origin: `https` only (an `http` base is rejected unless the host is localhost AND `RUSTSCRIPT_AGENT_TELEGRAM_ALLOW_INSECURE_LOCALHOST=1`), no credentials, query, fragment, or path. |
| `RUSTSCRIPT_AGENT_TELEGRAM_ALLOW_INSECURE_LOCALHOST` | `PD_EDGE_AGENT_TELEGRAM_ALLOW_INSECURE_LOCALHOST` | flag | unset (`false`) | Only the exact value `1` allows an `http` api_base for a localhost host (test fixtures and local development; the token must never travel in cleartext). |
| `RUSTSCRIPT_AGENT_TELEGRAM_DROP_PENDING_UPDATES` | `PD_EDGE_AGENT_TELEGRAM_DROP_PENDING_UPDATES` | flag | `1` (drop) | Safe first-boot default: updates queued while the bot was offline are drained and dropped before polling starts (fail-closed). Only the exact value `0` keeps them for processing. |
Expand Down Expand Up @@ -184,6 +246,17 @@ above; every other field is set by embedding code.
| `sqlite` | `SqlitePolicy` | core defaults, `max_statements = 1024` | — |
| `approvals` | `ApprovalSettings` | `read = "allow"`, `write = "ask"`, `process = "ask"`; reviewer/model/fallback/required overrides unset (`None`) | Validated by `ApprovalSettings::validate`: `reviewer` must be `guardian` or `user` when set; `reviewer_model` must be non-empty, at most 128 characters, with no control characters or surrounding whitespace; `read = "ask"` is unsupported. |
| `fuel` | `Option<u64>` | `Some(10_000_000)` | VM fuel budget; `None` disables the fuel cap. |
| `managed_workspace` | `Option<ManagedWorkspaceConfig>` | `None` (direct mode) | Trusted native startup opt-in for persistent synchronous managed effects. `ManagedWorkspaceConfig::for_workspace(source_root)` supplies default quotas and validates the source; `AgentGatewayConfig::validate` revalidates it. Gateway construction requires an absolute persistent database, and the runtime rejects a root source or overlap with persistent state. Import and queries still require current Native workspace authority; see the primary-only environment setting above. |
| `managed_completion_policy` | `ManagedCompletionPolicy` | `staged` | `staged` (default) may commit `run.completed` while the branch stays staged; `required` commits the run's branch publication in the same durable transaction as `run.completed` and is rejected unless `managed_workspace` is set. Captured at admission from trusted host configuration; affects only the run terminal step and grants no workspace authority. |

`ManagedWorkspaceConfig` contains `source_root: PathBuf`, `quota: ResourceQuota`,
and `retention: ManagedRetentionLimits`. `for_workspace` defaults to at most
1 000 000 tree entries, 64 MiB per tree, and depth 32. Native embeddings may
choose smaller positive bounds through `ResourceQuota::new`; these defaults
are also its hard ceilings. Retention defaults to 128 objects and 512 MiB;
both bounds must be positive and cannot exceed those ceilings. Retained
evidence is not evicted to admit a call. These are native configuration
fields, not additional environment variables or `config.yaml` settings.

`RateLimitConfig` bounds: `enabled` (master switch), `ip_burst`, `account_burst`,
`window`, `max_buckets`. `TelegramConfig` bounds: `bot_token` (redacted in
Expand Down Expand Up @@ -406,10 +479,9 @@ The following configuration does **not** exist in this revision. It is
listed only to reserve the namespace and to make the roadmap explicit; no
binary reads these names and setting them has no effect.

- **A4 harness/approval machinery (排除)** and **A6 parallel tools /
subagents (排除)** are out of scope for this repository's current
milestones and define no configuration. The approval repository CRUD
(`approval.request`/`get`/`resolve`/`expire` storage commands) exists, but
there is no approval flow driving runs.
- **A6 parallel tools / subagents (排除)** remains out of scope and defines
no configuration. Tool approval is implemented through the `approvals`
settings above and durable approval records; a general-purpose harness layer
beyond tool approval is not part of the current product.
- A job **scheduler** is not implemented: job CRUD/pause/resume/latest-output
routes exist, but scheduled execution defines no configuration.
Loading
Loading