Request shaping is selected by version predicates in llm-core, so a Claude id
from a generation newer than anything the plugin encodes fell through to pre-4.6
shaping: manual budget_tokens thinking plus caller sampling parameters, both of
which current models reject.
Resolve such an id through the plugin's existing forward-compat path and stamp
params.canonicalModelId, the same seam Bedrock and Mantle use to map a
provider-native id onto a canonical Claude contract. Shared contracts are
untouched, so Claude served through third-party providers keeps its own
shaping.
* fix(models): authenticate Anthropic discovery with OAuth tokens
Live model discovery sent every credential as x-api-key. Claude subscription
credentials are OAuth access tokens, which Anthropic rejects in that header, so
discovery 401d and silently fell back to the shipped catalog.
Select the auth header from the detected credential shape: API keys keep
x-api-key, OAuth tokens use Authorization: Bearer. The two are mutually
exclusive because Anthropic rejects requests carrying both.
* fix(models): keep shipped Anthropic models when discovery succeeds
Live discovery replaces the seed catalog with the /v1/models response. Anthropic
does not publish every model it serves, so a shipped entry with no live row was
dropped from the provider listing once discovery started working.
Re-add manifest models the live response omitted; discovered rows still win on
shared ids.
* feat(plugins): support additional upstream session kinds
* feat(acpx): monitor upstream Pi sessions
* feat(opencode): monitor upstream OpenCode sessions
* fix(opencode): ignore hidden user text activity
* refactor(plugins): share upstream echo filtering
* fix(opencode): preserve marker tuple narrowing
* refactor(opencode): detect upstream turns via event_sequence cursor
Replace timestamp/grace-window completeness heuristics with OpenCode's own
per-session event_sequence.seq cursor, which advances inside the same
transaction as the message/part projections. Detects change instead of
inferring completion, so staged part writes can no longer drop a turn.
Handles cursor regression from OpenCode migrations that clear event_sequence,
suppresses compaction replay and summary re-publish, and restricts human-turn
classification to visible non-synthetic text.
* docs(plugins): record why the adoption coordinator requires complete
* fix(opencode): align cursor proof with strict types
* fix(acpx): keep Pi JSON parser private
* feat(anthropic): gate live model discovery on contract coverage
Live catalog discovery cloned a template for any newly discovered Claude id, so
a future model generation would be selectable while request shaping treated it
as pre-4.6 and 400d. Add an opt-in acceptUnknownModel gate to the shared live
catalog seam and have the Anthropic plugin accept a discovered model only when
Anthropic's advertised capabilities agree with the contracts we would apply.
Manifest-published ids bypass the gate and keep their metadata.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs(plugins): document the live-discovery admission hook
Record acceptUnknownModel in the provider-plugin SDK reference: when to use
it, that manifest-published ids bypass it, and the fail-closed guidance.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Main already resolves bare opus to Claude Opus 5 but documents nothing about
the rolling behavior. State that bare family aliases follow the current
generation and that pinning is the opt-out, and add regression coverage
separating bare aliases, pinned aliases, and retired-ref upgrades.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(anthropic): detect Claude CLI routes pinned to non-default models
Route detection probed one hardcoded model id, so moving the packaged CLI
default stopped advertising session creation for configs routing an older
Claude model, and adoption stamped the packaged default onto sessions the
operator never routed. Probe the seeded CLI model ids instead and reuse the
routed model for both paths.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(anthropic): keep the canonical CLI model id module-local
The route probe replaced its only external consumer, so the export is dead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* feat(models): add Claude Opus 5 support
* test(models): align Opus 5 setup fixtures
* test(models): update ambient Opus 5 detection
* fix(models): reconcile Opus 5 support with main
* feat(anthropic): add Claude Opus 5 model support
Wire claude-opus-5 through the Claude 5 contract seams: adaptive-by-default
thinking with the full low..max effort range, default-sampling and prefill
stripping, streaming refusal contract, model-bound thinking replay, and 1M/128k
catalog metadata across anthropic, claude-cli, Vertex, Bedrock, and Mantle.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(anthropic-vertex): normalize explicit Vertex Opus 5 model rows
Extend normalizeAnthropicVertexResolvedModel so user-configured Vertex Opus 5
rows regain reasoning, image input, 1M/128k limits, and the native thinking
map; update the live-model priority expectation and regenerate docs_map.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(anthropic): accept "cli" entrypoint in Claude session catalog discovery
Claude Code v2.x writes entrypoint: "cli" in its JSONL session files.
Previously only "sdk-cli" was recognized, causing all v2.x sessions to
be silently skipped.
Extract CLI_ENTRYPOINTS set and isCliEntrypoint() helper, then apply
the same fix to both TypeScript catalog and macOS paired-node native
catalog. Add matching regression coverage on both platforms.
Closes#105164
* fix(anthropic): refresh Claude CLI session discovery
Apply the reviewed cli/sdk-cli allowlist, mirrored negative coverage, and fallback-scope documentation to current main.
Co-authored-by: 黄攀0668000858 <huang.pan@xydigit.com>
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(anthropic): complete transcript reverse-scan windows across short reads
readLocalClaudeTranscriptPage filled each reverse-scan window with a single
positional read and threw "Claude transcript changed while it was being read"
whenever bytesRead !== size. A positional read may return fewer bytes than
requested inside an unchanged file, so a benign short read failed the transcript
page load with a message implying the file changed. The forward metadata scan in
the same file already loops short reads; the reverse scan did not.
Fill each window across short reads, advancing both the buffer offset and the
file position. A zero-byte read before the window is filled is premature EOF and
still throws the same error.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(anthropic): tighten transcript short-read coverage
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(extensions/anthropic): write migrated model entries as own properties
* test(extensions/anthropic): guard optional models value in Object.hasOwn assertions
---------
Co-authored-by: chengzhichao-xydt <chengzhichao-xydt@users.noreply.github.com>
* feat(anthropic): import Claude Desktop custom groups
* fix(anthropic): resolve Claude Desktop groups from LevelDB entries
Group discovery regexed whole decompressed blocks, so matches were not
attributable to a Local Storage value and byte order decided the winner.
On a real 5.4MB store that mislabelled 59 of 159 sessions, surfacing a
mojibake label spliced out of Snappy copy-record bytes.
Parse SSTable entries properly instead: prefix-delta keys bounded by the
restart array, newest internal sequence per user key (honoring deletions),
and record scanning confined to a single value. Values are normalized so
Chromium's UTF-16 form scans like Latin-1, and unflushed WAL writes keep
precedence over SSTables.
Prod LOC grows ~77; it buys structural correctness in place of ordering
luck, and folds the old index-only walk into one shared entry decoder.
Verified against the live store: 159 assignments, 0 mislabelled versus an
independent entry-level ground truth, 0 control-character labels.
* fix(ui): sort custom session groups ahead of project groups
Custom groups were pushed into the section list as encountered, so their
position depended on roster order rather than the documented behavior;
the existing test only passed because its fixture happened to be ordered
that way. Collect custom and project groups separately and concatenate,
and assert the guarantee with a reversed-input case.
* fix(anthropic): widen LevelDB fixture key type for test typecheck
The prefix-delta helper assigned a Buffer into a Buffer-typed accumulator
whose generic argument differed, which tsgo rejects in the test lane.
* chore(anthropic): drop release-owned changelog edit from the PR
CHANGELOG.md is generated at release time, so a normal PR must not carry
an entry. The release-note context lives in the feature commit message and
the PR body instead.
* fix(anthropic): add claude-fable-5 to CLI allowlist, aliases, labels, and context window
CLI path was missing claude-fable-5 across all four metadata layers while
the direct Anthropic provider path in register.runtime.ts already had full
support via isAnthropicFable5Model() / resolveAnthropicFixedContextWindow().
- CLAUDE_CLI_DEFAULT_ALLOWLIST_REFS: add claude-cli/claude-fable-5
- CLAUDE_CLI_MODEL_ALIASES: add fable/fable-5/claude-fable-5 mappings
- CLAUDE_CLI_CONTEXT_WINDOWS: add claude-fable-5 -> 1_000_000 (was falling back to 200K)
- CLAUDE_CLI_MODEL_LABELS: add Claude Fable 5 (Claude CLI)
- resolveClaudeCliImageMediaInput: add fable-5 to 2576 max-side tier
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(anthropic): route fable alias through context/model-ref canonicalization paths
* fix(agent): include claude-cli in fable-5/mythos-5 fixed context window resolution
* fix(anthropic): add claude-fable-5 to static CLI manifest catalog
* test(anthropic): cover fable-5 alias canonicalization in bare and provider-qualified forms
* style(anthropic): fix indentation in plugin manifest
* fix(anthropic): publish Fable CLI output limit
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* perf(sdk): load Claude CLI identity from a narrow plugin artifact
src/plugin-sdk/anthropic-cli.ts snapshots CLAUDE_CLI_BACKEND_ID at module
scope through the sync facade loader, which jiti-evaluates the full
anthropic api.js barrel on source checkouts: 130.86s of self time per
cold worker on CI (Testbox import profile), silently stalling every job
whose graph reaches cli-runner/prepare.ts. A narrow cli-api.js artifact
carries the two static facts; Testbox proof: the reliability+helpers
agents-core pair drops from 157.6s to 14.3s.
* perf(ci): stripe unit-fast and tooling node tests on the full plan
core-fast ran the import-bound unit-fast graph as one job (247s vitest,
181s module evaluation) and core-tooling as one serial job (241s); both
now stripe on the full plan like the compact plan, so the compact-only
expansion is deleted and the docker helper config rides with the
isolated shard on both plans. Group hints refreshed from main run
29551077288; stale per-file commands hints dropped (the consolidation
landed); cache-writer selection and the warm workflow track the stripe
names.
* fix(state): tolerate vanished sqlite sidecars in agent-db permission sweep
existsSync+chmodSync raced SQLite's own WAL/SHM cleanup: a checkpoint or
close between the two calls throws ENOENT (observed from the transcript
reconcile worker in server-startup-web-fetch-bind on CI). chmod directly
and swallow only ENOENT, which removes the TOCTOU window.
Embedded runs targeting a CLI runtime provider fall through to the openclaw
harness and call the provider API directly with the runtime's credentials
(cli_runtime_passthrough_openclaw). Anthropic routes direct anthropic-messages
calls on subscription OAuth tokens to metered extra-usage billing — this is
long-standing behavior, not a recent change. Without extra-usage balance,
every such run (e.g. active-memory recall) fails with a billing error; with
extra-usage enabled, the run silently draws paid metered usage instead of the
plan limits the CLI runtime was configured for. Only CLI-backed execution
runs on plan limits for those credentials.
Add an opt-in RunEmbeddedAgentParams.cliBackendDispatch: "subscription-auth"
that dispatches the run through runCliAgent as a one-shot turn when the
provider is claude-cli, a CLI backend is registered, and the ordered auth
profile selection for the passthrough resolves to a subscription (oauth/token)
credential or nothing rather than an API key; resolution stays on stored
credential metadata, with no credential materialization or refresh on the
per-turn path. The dispatch translates toolsAllow into the selectable-backend
surface (native: [], allowlisted loopback MCP tools; wildcard allowlists stay
MCP-only), runs with a fresh CLI process (no live-session reuse; session-
scoped bundle-MCP retirement on run end rather than the process-wide loopback
close), bridges CLI tool result events to onAgentToolResult with native-path
semantics (normalizeToolName + isToolResultError), and drops CLI session
bindings from the result.
The selectable-backend MCP list now also bounds the loopback MCP grant
server-side: the grant carries a per-run gateway tool allowlist enforced in
scoped tool resolution, so tools outside the run's allowlist can be neither
listed nor called even under CLI bypass permission modes where
--allowedTools is advisory.
active-memory recall opts in so recall works on claude-cli subscription-only
instances and stops drawing metered extra usage where it previously could.
Scoped to claude-cli; other CLI runtimes keep the passthrough until their
direct-API contract is verified.
The dispatch also mirrors the run into the run's session transcript through the session
accessor (user turn, tool call/result records as they stream, final assistant
snapshot at run end) so transcript consumers keep parity with embedded runs:
active-memory's persistTranscripts, timeout partial-text salvage, and the
live terminal-search watcher that polls the session file mid-run.
Post-review hardening: canonical anthropic/<model> refs whose configured
agentRuntime is claude-cli resolve through the runtime policy before the
dispatch gate (they previously stayed on the failing passthrough); restricted
dispatches serve an exclusive loopback-only MCP bundle so user/plugin MCP
servers stay outside the run's tool universe; and the transcript recorder
flushes the latest assistant snapshot the moment the run aborts, so timeout
salvage sees partial text even while the killed CLI child is still settling.
Recalls routed to the claude-cli runtime default to a 45s budget (measured
CLI-dispatched runs take 14-20s, over the plain 15s default); explicit
timeoutMs config always wins.
Transcript mirror keeps bare-array tool_result content (claude stream-json
echoes MCP results without a {content} wrapper); dropping it classified every
successful recall as no_relevant_memory.
CLI dispatch resolves inside session/global lane admission so dispatched
runs obey the same lifecycle, placement, and concurrency gates as native
embedded runs.
LOC-ratchet offsets move the loopback grant-context builders to
cli-runner/mcp-grant-context.ts and dedupe the run/prep stage-summary
emitters into attempt-stage-timing.ts; unused type exports dropped and the
now-used stream-message baseline entry removed.
The recall timeout default now consumes the runner's own dispatch
eligibility through a new plugin-runtime seam
(agent.resolveCliBackendDispatchEligibility): API-key and missing-backend
routes keep the passthrough and its plain 15s default.
Eligibility honors an explicitly pinned authProfileId (the credential the
run executes on) before ordered profile selection, in both directions.
Transcript mirror composes buildAssistantMessage + buildUsageWithNoCost
directly; main trimmed the zero-usage wrapper export (ab0ccc244b) before
this branch's usage landed.
Dispatch eligibility is provider-owned: the anthropic plugin's claude-cli
backend declares CliBackendPlugin.subscriptionAuthDispatch and core reads
the registered descriptor instead of a core provider allowlist.
Dispatch fails closed on tool policy (only non-empty named allowlists are
expressible on the CLI surface; deny-all, wildcards, absent allowlists, and
disableTools/modelRun keep the passthrough), threads the pinned
authProfileId into CLI runtime resolution, and emits onExecutionStarted at
the admitted dispatch boundary.
* chore(plugins): gate deprecation hygiene in CI and purge internal deprecated usage
* fix(agents): migrate auth-profiles oauth import to OAuthProviderId
* fix(whatsapp): use durability param in delivery-recovery test
* chore(scripts): declare shouldRunDeprecationHygieneChecks in check-changed types
* test(plugins): update runtime-api guard for ssrf policy rename
Peter Steinberger