Adds a generic plugin-declared selectable-context-window surface mirroring thinkingLevels: ModelCatalogEntry.contextWindows + contextWindowDefault through catalog normalization and the gateway protocol, session validation on sessions.create/patch, and a 200K/1M switch inside the Control UI model picker for Claude CLI 5-series models. The Anthropic plugin owns the option mapping: explicit 1m → `[1m]` argv suffix, 200k → bare id + CLAUDE_CODE_DISABLE_1M_CONTEXT=1, omitted → bare id (shipped default argv). Run budgets follow the selection on both CLI and native paths, so a 200K session gets a matching auto-compact window instead of a silent 1M budget.
Review fixes landed in this PR: run-owner prepared-fact plumbing so ordinary replies honor the selection; atomic catalog overlay merge and normalization for the options/default tuple; one-owner tuple reads in the picker; sessions.create key-presence patch semantics; native-run budget capping.
Feature direction and in-picker switch shape by @obviyus (maintainer review).
The cloud machine picker described each class in prose ("Cheap smoke checks
and small repos"), which was the widest thing in the row, got ellipsized, and
did not say what the operator actually picks on. Machine options now carry the
class shape and the picker renders "32 vCPU · 64 GB" instead.
Crabbox reports per-provider class shapes from `providers --json`; the plugin
reads that catalog once per lifecycle and attaches cpu/memoryGb to each option.
A missing binary, failed command, unparseable output, absent classes, or an
unmatched provider all degrade to label-only rows, so an older Crabbox keeps
working.
`description` was never released, so it is removed rather than deprecated, and
listMachineOptions becomes async for the catalog read. Deletes the duplicate
machine-option projector in the environments server method.
* fix(cloud-workers): close lifecycle ownership gaps
Own bootstrap cleanup at the operation boundary and make fallback workspace sync converge across retries. Re-establish tunnel readiness per connection, retire placements before destructive session mutation, and keep operator diagnostics lightweight and redacted. Cover destructive lifecycle paths in their original execution order.
* fix(cloud-workers): drain local claims before retirement
delete/reset drain admitted local work, re-read exact identity, retire before destructive cleanup; active-claim/race tests.
* fix(cloud-workers): bind retry cleanup to workspace owner
Attest canonical HOME and the exact managed path.
Revalidate ownership before recursive fallback cleanup.
Cover malicious paths and ownership drift with tests.
* fix(cloud-workers): fence fallback workspace receivers
* fix(cloud-workers): honor Crabbox SSH fallback ports
Select a currently reachable advertised SSH endpoint before returning a Crabbox worker lease, while preserving fail-closed teardown. Allow deletion of failed cloud sessions only after the environment record proves the provider lease is gone, and align the Cloud Workers docs with the shipped configuration and Control UI flow.
* fix(cloud-workers): retry advertised SSH endpoints
Persist provider-advertised SSH fallback ports and retry them through the shared pinned transport used by bootstrap, tunnels, reconnects, and workspace transfers. Keep candidate selection provider-neutral, preserve identity and host-key fencing, and add the same-version SQLite column through the existing additive migration path.
* fix(crabbox): recheck SSH readiness after setup
Treat provider setup as a mutating lifecycle phase: refresh authoritative inspect state and wait for SSH readiness again before returning the lease. This prevents bootstrap from racing setup-triggered SSH restarts on fresh and replayed workers while preserving security attestation and fail-closed teardown.
* fix(cloud-workers): retry bootstrap transfers across ports
Use the shared advertised-port candidate runner for bundle transfer, install, and best-effort upload cleanup as well as preflight. This keeps fresh bootstrap connections on pinned identity and host-key semantics when a previously selected SSH port disappears between phases.
Proof passed: 517 focused tests (1 skipped), final autoreview/TruffleHog clean, targeted lint/format. Full check-changed reached the unrelated max-lines baseline drift on current origin/main: this branch's merge base contains ui/src/pages/chat/components/chat-model-controls.ts while current origin/main removed that baseline entry; do not touch the unrelated baseline.
* fix(cloud-workers): persist fallback ports compatibly
Create the fallback-port companion table lazily as an additive surface, preserving downgrade tolerance for binaries that still expect the older shared-state schema.
* fix(cloud-workers): preserve bundle artifact narrowing
Bind the immutable discriminated artifact before the SSH retry callbacks so bundle-only fields remain narrowed across fresh candidate invocations. Use the shared temp-directory owner in the tunnel fallback test for deterministic cleanup.
* fix(cloud-workers): retire deleted placements
Terminal placement rows retire by exact CAS after session deletion, and managed AWS docs now pin the Gateway SSH CIDR.
* fix(crabbox): await pending AWS attestation
missing authoritative metadata is pending only while the lease is non-ready; ready leases still require explicit false.
* fix(cloud-workers): fence remote command retries
require explicit idempotent/never classification; remote worker launch never retries after ambiguous SSH loss.
* fix(cloud-workers): bound SSH fallback operations
* feat(plugins): support the Agent Plugins bundle format
* docs(plugins): document the Agent Plugins bundle format
* test(agents): preserve agent bundle runtime discovery
* fix(plugins): isolate Agent Plugins data-dir failures and align MCP support reporting
* docs(plugins): list Agent Plugins in the canonical plugin-format guides
* fix(plugins): gate Agent Plugins detection on schema, pure inspection, root-relative cwd
* fix(plugins): record Agent Plugins data-dir ownership explicitly
* docs(plugins): cover Agent Plugins in the CLI install detection guide
* fix(plugins): carry Agent Plugins data-dir and transport contracts through external MCP projections
* fix(plugins): preserve multi-entry pack identity and safe install compensation
Multi-entry plugin packs registered every entry under the single manifest id,
so only one entry survived registry normalization; discovery now records an
entry-scoped effectivePluginId that the manifest registry honors. Managed
install compensation no longer silences runtime cache invalidation on the
success path, and local linked installs never delete operator-owned source
directories on persistence failure.
* fix(plugins): reject colliding derived entry ids in multi-entry packs
ClawSweeper P1: ./a/index.ts and ./b/index.ts both derived pack/index and one
entry silently lost same-root dedupe. Colliding entries now emit an error
diagnostic naming the sources and register no candidate; docs note the
entry-scoped plugins.entries key for multi-entry packs.
Aggregators republish first-party models under a namespaced id and varying
case, so the shared-model check could not group novita/moonshotai/kimi-k3 with
moonshot/kimi-k3. Matching now ignores one leading namespace segment and case,
which catches that whole class automatically instead of needing an
upstreamModel marker per row.
Declares the tier on the 18 rows this surfaces across baseten, deepinfra, gmi,
novita, nvidia, and together. All are "capable", matching what the runtime
already applied for an absent flag, so behavior is unchanged.
* feat(plugins): catch code-mode drift between catalogs sharing one model
Adds a contract test that groups bundled catalog rows by shared upstream
model and requires every row in a group to declare compat.codeMode once any
sibling does. Rows sharing a model id group automatically; rows under
different ids opt in with the new manifest-only `upstreamModel` marker.
Moves the kimi catalog into its manifest so the scan can see it, and records
the tier reseller catalogs were silently missing as explicit "capable".
* docs: regenerate docs map for the shared-model code-mode section
* feat(config): give shared channel settings help text
* fix(ui): space the settings tiers and stop repeating help on list items
* feat(ui): link each channel detail to its docs page
* test(ui): mirror production channel hints in the mock dev server
* docs: note the shared channel help inheritance for plugin authors
* fix(config): let a channel suppress shared help with an empty string
* chore(config): regenerate bundled channel metadata for the new WhatsApp hints
* fix(ci): check bundled channel metadata locally
* chore(release): keep changelog release-owned
* chore(release): keep changelog release-owned
* fix(ui): keep shared channel help contract-neutral
* refactor(channels): generalize channel retry runner and retry-after parsing; drop dead configured-state probes
* fix(channels): align retry and metadata contracts
* feat(agents): derive a provider-declared default utility model when unset
When agents.defaults.utilityModel is not set, utility tasks (titles, progress
narration) now use the primary provider's declared small model
(modelCatalog.providers.<id>.defaultUtilityModel: OpenAI -> gpt-5.6-luna,
Anthropic -> claude-haiku-4-5). Auth is inherent because derivation follows
the agent's primary provider. Setting utilityModel to an empty string
disables utility routing entirely; narration turns on automatically when a
default resolves and stays off otherwise.
Formatting/lint/tests verified on Testbox (oxfmt, oxlint, plugins:inventory,
docs:map, import-cycles, check:test-types, 4 test files).
* fix(ai): drop the temperature parameter for models that reject it
The GPT-5.6 family 400s on temperature via the Responses API (live-verified:
gpt-5.6-luna/-terra reject it; gpt-5.5 and gpt-5.4-mini/nano accept it).
supportsOpenAITemperature gates all three OpenAI payload builders, with a
catalog compat override (compat.supportsTemperature) declared for the 5.6
family in the openai manifest. Without this, utility tasks (titles,
narration) would fail once gpt-5.6-luna becomes the derived default.
Gates on Testbox: oxfmt, oxlint, plugins:inventory, import-cycles,
check:test-types, 6 test files. Live proof on Testbox: gpt-5.6-luna and
claude-haiku-4-5 one-shot completions succeed with temperature requested.
* fix(agents): carry the primary model's auth profile onto the derived utility default
A primary like openai/gpt-5.5@work previously reached utility tasks via the
profiled fallback; the derived default shares the provider, so its trailing
auth profile carries over instead of silently switching to default
credentials (Codex review). Gates on Testbox: oxfmt, check:test-types, tests.
* docs: realign the manifest provider-fields table after adding defaultUtilityModel
* feat(ui): add plugin catalog management
* feat(gateway): add plugins.uninstall and richer plugin catalog metadata
Adds a plugins.uninstall gateway method (operator.admin, control-plane write)
backed by a lock-guarded uninstallManagedPlugin that mirrors the CLI flow:
config cleanup, install-record removal, managed file deletion, and registry
refresh. Bundled plugins stay disable-only. Catalog entries now carry a
manifest-derived category and a removable flag; ClawHub search results expose
download counts and verification tiers.
* feat(ui): redesign plugins page with inventory, store shelves, and cover art
Rebuilds /settings/plugins around three tabs: Installed (category-grouped
inventory with overview stats, state filters, uninstall for external plugins,
and inline MCP server management through the shared config seam), Discover
(featured/official shelves plus one-click MCP connectors and curated ClawHub
searches), and ClawHub (search with download counts and verification badges).
Every catalog entry renders bundled cover art or a deterministic gradient
monogram tile - no more empty boxes. Artwork generated with Codex CLI, shipped
as 512px WebP under ui/public/plugin-art.
* chore(ui): regenerate locale bundles for plugins manager strings
* docs: describe plugins manager tabs, uninstall, and MCP connectors
* fix(plugins): human catalog labels and un-pinned hosted fallback ids
listManagedPlugins now prefers manifest names over registry package-name
backfill, falls back to channel catalog labels and blurbs, and stops pinning
expectedPluginId when a hosted feed entry only exposes its package name
(which rejected every legitimate install of that package). Found via live
gateway testing against ClawHub.
* fix(ui): send minimal RFC 7396 merge patches for MCP server edits
config.patch merges rather than replaces, so key removal needs an explicit
null; sending the full config back made MCP server removal a no-op. Found
via live gateway testing.
* fix(ui): write explicit MCP transports for URL servers
The MCP runtime defaults URL-only servers to SSE, so streamable HTTP
endpoints saved by the add form or connector templates would fail at
connect time. Connector templates now declare their transport and the
add form infers streamable-http unless the URL follows the /sse
convention. Flagged by autoreview against the transport resolver.
* test(ui): wait for deferred plugin requests before resolving in e2e
* feat(ui): plugins detail view, action menus, and unified ClawHub search
Reworks the plugins page from PR #103176 feedback: merges the ClawHub tab into
Discover (typing searches ClawHub inline and appends a quiet From ClawHub
section, with Browse ClawHub demoted to a header text link), makes every row and
store card open a plugin detail overlay (hero art, primary enable/install
action, metadata table), and replaces enable/disable switches with a state chip
plus an overflow menu (Enable/Disable, Remove for external plugins, View
details) matching the ChatGPT-store install+menu pattern. MCP rows use the same
menu; refresh is now icon-only.
* chore(ui): regenerate locale bundles for plugins UI iteration
* feat(ui): vetted, grouped connector catalog for the plugins store
Expands Connect your world to 28 connectors organized into use-case shelves
(Work & productivity, Coding & infrastructure, Home & media, Everyday life).
Every entry passed a three-stage subagent review: official-docs verification
plus live endpoint probes for MCP servers, ClawHub result-quality and
malware/typosquat screening for curated searches, and an adversarial pass
that dynamically registered OAuth clients to prove one-click viability.
That review removed Figma (registration allowlisted, 403) and Atlassian
(OAuth issuer-mismatch bug upstream), downgraded GitHub to PAT-based setup
(no dynamic client registration upstream), fixed Linear (/sse retired) and
Home Assistant (/api/mcp, streamable HTTP) endpoints, retargeted poisoned or
dead searches (youtube, finance, hue dropped; calendar -> google calendar;
stocks replaces finance), and added Todoist, Airtable, Canva, Stripe,
Context7, DeepWiki, Hugging Face one-click MCP servers plus Jira, PDF,
transcription, Kubernetes, Reddit, maps, translation, and notes searches.
Keyless servers get a ready-to-use success message; new cover art included.
* chore(ui): regenerate locale bundles for connector groups
* fix(plugins): suppress hosted catalog rows once their package is installed
Hosted feed entries without a declared runtime id fall back to their package
name as catalog id, which never matches the installed runtime id, so the
Discover shelf kept offering an already-installed package. Installed package
names now also suppress official rows. Flagged by autoreview.
* fix(plugins): pin declared runtime ids and surface connector errors in place
The runtime-id pin now keys off explicitly declared catalog ids (plugin,
channel, or provider) instead of string-comparing against the package name,
so declared ids that equal their package name stay enforced while entry-id
fallbacks stay unpinned. Connector add failures on Discover now render on the
triggering card instead of the Installed tab's MCP section. Both flagged by
autoreview; regression tests included.
* feat(ui): full inventory artwork, pulse header, and two-column plugin list
Every bundled plugin now ships distinctive cover art (113 new Codex CLI
illustrations; 172 total, ~2.1MB WebP), so inventory rows and detail views
never fall back to monogram tiles. The four stat cards give way to a compact
inventory pulse: a segmented enabled/disabled/issues meter whose legend and
counts live inside the filter chips. Inventory, MCP, and search rows flow
into two columns when the panel is wide enough.
* chore(ui): regenerate locale bundles for pulse header
* fix(ui): omit stdio args from the MCP server row target
Stdio MCP args routinely carry tokens, and the inventory is visible to
read-only operators; mirror the config page and show only the command.
Flagged by autoreview; regression test included.
* fix(merge): point crestodian setup at relocated plugin commit/refresh modules
* fix(merge): add bootstrapToken to plugins page test gateway harness
* fix(plugins): name catalog install-action branches so Swift emits the union
* fix(ui): satisfy strict lint on plugins page form parsing and mocks
* chore(build): regen docs map, raise plugin-sdk declaration budget for new protocol surface
* fix(ui): type the plugins page patch mock with its real call signature
Ayaan Zaidi