Third-party Anthropic-compat providers (MiniMax, GLM, Kimi, DeepSeek)
all reuse the Anthropic SDK's wire format, which means the claude CLI
and claude-code-sdk read the bearer token from ANTHROPIC_AUTH_TOKEN no
matter which vendor is being talked to. Pre-#244:
* Canvas surfaced the vendor-specific name (MINIMAX_API_KEY, etc.)
to the user — so a user who saved only MINIMAX_API_KEY hit a
silent 401 on first call.
* The boot audit said `MINIMAX_API_KEY=set`, making it look like an
SDK bug rather than a routing gap.
* A user with multiple vendor keys could only run one workspace at a
time because they all fought over the shared ANTHROPIC_AUTH_TOKEN
slot.
Diagnostic-only audit logging shipped earlier (#32) but the actual
routing was never written — task #244 was mismarked complete.
Changes:
* config.yaml: third-party model `required_env` now references the
per-vendor name (MINIMAX_API_KEY, GLM_API_KEY, KIMI_API_KEY,
DEEPSEEK_API_KEY) so canvas asks the user for the right key.
First-party Anthropic models still use ANTHROPIC_AUTH_TOKEN /
CLAUDE_CODE_OAUTH_TOKEN.
* config.yaml: each third-party provider's `auth_env` lists the
vendor name FIRST (priority order) so projection picks the
vendor key over a stale ANTHROPIC_AUTH_TOKEN.
* adapter.py: new `_project_vendor_auth(provider)` helper, called
from `setup()` right after `_resolve_provider`. Idempotent — only
projects when ANTHROPIC_AUTH_TOKEN is unset (operator override
always wins). Logs the projection by NAME, never by VALUE
(mirrors `_audit_auth_env_presence`).
* tests/test_provider_routing.py: 6 new tests pin the contract —
vendor-key-set projects, AUTH_TOKEN-already-set is never
clobbered, first-party providers skip projection, secret value
never leaks into a log record, empty-string vendor env doesn't
trigger projection, and the same routing fires for GLM / Kimi /
DeepSeek.
Mirrors the parallel hermes-side fix from task #249 / hermes PR #38;
keeps the two runtimes' multi-vendor UX in lockstep.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds two operator-visible boot diagnostics that close the diagnosis gap
exposed by the 2026-05-02 MiniMax E2E crash-loop. The universal
canvas-picked-model fix (Bug B) and per-model required_env (Bug D) live
in molecule-core PR #2538 — this PR adds the per-template visibility
that complements them so operators can answer "is the key missing or is
routing wrong?" from `docker logs` alone.
Changes
-------
adapter.py:
- _AUTH_ENV_AUDIT tuple of 8 vendor env names (CLAUDE_CODE_OAUTH_TOKEN,
ANTHROPIC_API_KEY/AUTH_TOKEN/BASE_URL, MINIMAX/GLM/KIMI/DEEPSEEK_API_KEY).
- _audit_auth_env_presence() helper — single INFO line of NAME=set/unset
pairs. NEVER logs values; the test pins this with a "fake-secret-MUST-
NOT-LEAK" sentinel that must never appear in the log message.
- One call site at the end of setup()'s boot banner so every workspace
start emits both "which provider got picked" and "which envs are present"
in adjacent log lines.
entrypoint.sh:
- log_boot_context() function fired once before the gosu drop (as root)
and once after (as agent) so an operator can spot env values lost
across the privilege drop. Emits uid/gid/user/hostname/workspace_id/
platform_url/configs_dir/workspace_dir + the same 8 env names as
NAME=set/unset. Mirror of _AUTH_ENV_AUDIT — list pinned in sync by a
new AST-style test (test_audit_env_list_matches_entrypoint_sh) that
parses entrypoint.sh and asserts set-equality with adapter.py's tuple.
tests/test_adapter_logging.py (new):
- 4 tests covering the audit contract: every name appears, all-unset
scenario, empty-string treated as unset (matches routing semantics),
and the cross-file sync gate against entrypoint.sh's for-loop.
- Stubs molecule_runtime + a2a so the helpers can be imported without
the real wheel installed in CI (mirrors test_adapter_prevalidate.py's
scaffolding pattern).
Why this complements molecule-core PR #2538
-------------------------------------------
- PR #2538 makes Bug B (canvas-picked model silently dropped) impossible
by resolving model centrally in workspace/config.py:load_config —
every adapter (claude-code, hermes, codex, future ones) gets the
passthrough for free.
- PR #2538 makes Bug D (preflight rejects valid auth for non-default
models) impossible by REPLACE-not-union per-entry required_env.
- This template PR is the per-template observability layer: when one
of those universal fixes regresses (or when an operator misconfigs a
vendor key), the boot logs say exactly which env was present at each
tier. Validated end-to-end on workspace
be27badd-00a7-4cef-91e8-af428175c76f (clean boot, MINIMAX_API_KEY=set
audited, no crash-loop).
Closes part of molecule-monorepo task #248. Sibling of #2538 for
molecule-core.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two review nits:
1. Narrow the import-arm catch in _mark_sdk_wedged and
_clear_sdk_wedge_on_success to (ImportError, ModuleNotFoundError).
The bare `except Exception:` would have masked an AttributeError /
TypeError from a runtime_wedge API rename — silently degrading the
mirror to "no-op" and making heartbeat + the smoke gate (#131)
blind to claude-code wedges. The structural snapshot test in
molecule-core (task #169) catches the rename at PR-time. Older
runtimes that don't ship runtime_wedge at all still hit ImportError
and silently no-op — the local sticky flag still gates is_wedged()
inside this module so internal callers keep working.
2. Add mirror-CALL-failure injection tests. The recorder used by the
original tests never raised, so the inner try around
_mark_runtime_wedged(reason) (and the symmetric clear) wasn't
pinned. New tests inject a recorder whose mark/clear raise on call,
then assert: (a) the call attempt was recorded, (b) the local
sticky flag stayed correct, (c) the failure was logged at ERROR.
Pins both the contract ("mirror is best-effort, local is source of
truth") AND the operator-visible signal (an ERROR log line is the
only way to see a silent mirror regression).
Regression-injection-checked: removing the call-side try arm makes
both new tests fail with clear messages. Tests: 7 in
test_runtime_wedge_mirror.py, 45 across the whole tests/ tree.
The local _sdk_wedged_reason flag was only observed inside this module
— heartbeat reads runtime_wedge.is_wedged() (universal cross-cutting
holder) and so does the new boot-smoke gate from molecule-core PR
#2473 / task #131. Without the mirror, a wedged claude-code workspace
stayed green-dot on the canvas while every chat hung, AND the
publish-image gate could not catch PR-25-class init wedges before
the broken image shipped to GHCR.
_mark_sdk_wedged now mirrors into runtime_wedge.mark_wedged, and
_clear_sdk_wedge_on_success mirrors into runtime_wedge.clear_wedge.
Both are best-effort — older runtimes that don't ship runtime_wedge
silently no-op the mirror, so a template pinned to an older runtime
still boots. Mirror exceptions are logged but don't suppress the
local sticky flag, so internal callers (retry loop, cancel handler)
see consistent state regardless of the universal-side outcome.
Tests cover: mark mirrors with reason, first-call-wins propagates,
clear mirrors, no-op when not wedged, ImportError-resilience.
Regression-injection-checked: silencing the mirror branch fails the
mark+first-wins tests at unit-test time with a clear message naming
the missing runtime_wedge call.
Claude Code 2.1.x changed the flag's signature to take an *allowlist* of
tagged entries — `server:<name>` for manually-configured MCP servers,
`plugin:<name>@<marketplace>` for plugin channels. PR #25's
`{flag: None}` rendered as a bare `--<flag>` with no value, the CLI
rejected with `argument missing`, and the SDK timed out at `initialize`,
surfacing upstream as `Control request timeout: initialize` (caught
live on workspace dd40faf8 on 2026-05-01 — 100% of A2A turns wedged).
Pass `server:molecule` so the SDK forwards
`--dangerously-load-development-channels server:molecule`. Live-verified
end-to-end: A2A returns coherent replies AND the host claude session
renders inbound canvas messages as `<channel source="molecule" ...>`
tags inline (push UX without inbox poll).
Tests: replace the unconditional `None` pin with a tagged-form pin
that asserts the exact `server:molecule` value, plus a defense-in-depth
test that pins the invariants (non-None, non-empty, contains tag
colon) so any regression to the bare-switch shape fails at unit-test
time instead of surfacing as a live SDK initialize wedge. 38/38 pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Live-test revealed a regression in PR #24's setup() strip: the wheel-
default `anthropic:claude-opus-4-7` paired with an OAuth workspace
(CLAUDE_CODE_OAUTH_TOKEN set, no ANTHROPIC_API_KEY) is the realistic
production shape. Stripping in setup() routes those users into the
`anthropic-api` provider entry, after which the CLI hangs at
`initialize` because no API key env is set. Caught on workspace
dd40faf8 on 2026-05-01 — banner went `provider=anthropic-api` and
A2A wedged on Control request timeout.
Pre-fix routing (let prefixed strings fall through to providers[0] =
anthropic-oauth) is actually correct for this combo. The strip is only
needed at the CLI invocation site (create_executor) where claude's
`--model` arg must be a bare id.
Tests: replace `test_setup_strip_routes_prefixed_anthropic_to_anthropic_api`
with `test_setup_keeps_prefix_routing_oauth_for_anthropic_prefix`,
which pins the inverse — prefixed model + OAuth env stays on oauth and
emits no API-key warning. The 5 unit cases on `_strip_provider_prefix`
plus the `create_executor` strip pins remain unchanged. 36/36 pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The wheel-side push UX gates (capability + instructions, molecule-core
PR #2463) only matter if the host claude CLI is willing to register a
non-allowlisted experimental channel. During the channels research
preview the CLI requires --dangerously-load-development-channels to
bypass its allowlist; without it, every notifications/claude/channel
fired by the inbox bridge arrives at the host and is silently dropped.
claude-agent-sdk forwards arbitrary CLI flags to the spawned subprocess
via ClaudeAgentOptions.extra_args (claude_agent_sdk/_internal/transport/
subprocess_cli.py:340). Wire the flag in unconditionally — the flag is
harmless on builds that already allowlist the capability and required
on builds during the research preview, so there is no version skew to
guard. Remove the line once channels graduate to the default allowlist.
Test pins the wiring with a stubbed ClaudeAgentOptions recorder; runs
in CI without claude_agent_sdk / a2a / molecule_runtime installed via
the same _ensure_module/_ensure_attr pattern as the existing adapter
prevalidate test, but tolerates real packages being present locally.
Verified by injection: removing the extra_args line makes the test
fail with a message naming the missing flag and citing the SDK file
that consumes it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The molecule-runtime wheel's config.py defaults model to
`anthropic:claude-opus-4-7` so langchain/crewai consumers get a uniform
provider:model string out of the box. The claude CLI's --model arg
expects the bare model id and silently exits 1 (no stderr) on prefixed
strings — root cause of the 2026-05-01 "Agent error (Exception)" mid-A2A
bug. Diagnosed via strace on a live workspace: the CLI received
`--model anthropic:claude-opus-4-7` and exit_group(1)'d before any
non-fatal output.
Add `_strip_provider_prefix` and call it in both setup() (so
_resolve_provider routes anthropic:claude-X correctly to anthropic-api
instead of falling back to oauth) and create_executor() (so the bare
id reaches the CLI). Only known-Claude prefixes are stripped; unknown
ones (openai:, bedrock:) pass through so the CLI fails loudly instead
of being silently mangled.
Coverage: 8 new tests — unit tests for the helper across all branches,
end-to-end `create_executor` strip on dict + dataclass shapes, and a
caplog-based setup() test that pins provider=anthropic-api routing
after the strip (the silent-fallback failure mode this fix eliminates).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two correctness issues spotted in self-review of c6f4912:
1. String-as-prefix typo split into character tuple. ``model_prefixes:
mimo-`` (operator forgot brackets) used to iterate over characters
→ ``('m','i','m','o','-')``, silently routing every model id starting
with 'm', 'i', or '-' through the entry. Now: non-list values coerce
to empty tuple (entry survives but matches nothing — operator notices
in boot banner, not via misrouted requests).
2. Single bad provider entry nuked the whole registry. _load_providers
built the registry via a generator inside tuple(...). One AttributeError
mid-comprehension (e.g. ``[mimo-, 123]`` — int's missing .lower())
propagated out, broad except caught it, registry silently fell back
to _BUILTIN_PROVIDERS (oauth + anthropic-api only). Every third-party
model would then route to anthropic-oauth — exactly the silent-fallback
failure mode this PR was meant to eliminate. Now: per-entry try/except
drops the bad entry with a warning, rest survives.
Also: entries without a string ``name`` field are now dropped with a
warning instead of silently using the placeholder ``<unnamed>`` —
operator typos surface in boot logs.
Tests: 28 passing (3 new regression tests covering both issues plus
the no-name path).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Move the model→endpoint→auth-env mapping out of hardcoded constants
in adapter.py + entrypoint.sh into a single `providers:` list at the
top of config.yaml. The adapter loads it at boot via _load_providers;
canvas Config tab will read the same YAML for its Provider dropdown so
UI and adapter never disagree on what's available. Adding a new
provider becomes a one-line YAML edit — no Python or shell changes.
Includes 5 third-party providers ready out of the box (Anthropic-compat
endpoints, Bearer-style ANTHROPIC_AUTH_TOKEN OR ANTHROPIC_API_KEY auth):
xiaomi-mimo https://api.xiaomimimo.com/anthropic
minimax https://api.minimax.io/anthropic
zai https://api.z.ai/api/anthropic (NEW)
moonshot https://api.moonshot.ai/anthropic (NEW)
deepseek https://api.deepseek.com/anthropic (NEW)
Plus 7 new model entries in runtime_config.models (mimo-v2.5, MiniMax-M2,
MiniMax-M2.7, GLM-4.6, GLM-4.5, kimi-k2.5, kimi-k2, deepseek-v4-pro,
deepseek-v4-flash) so they show up in the Canvas Config dropdown.
Operator override unchanged: ANTHROPIC_BASE_URL set as a workspace
secret still wins over the registry default — the escape hatch for
regional endpoints (Xiaomi token-plan-sgp, MiniMax api.minimaxi.com).
entrypoint.sh: drops the `mimo-*` case mapping (adapter handles routing
now). _BUILTIN_PROVIDERS retained as malformed-YAML fallback so a
bare-bones workspace still boots with oauth + anthropic-api defaults.
Tests: 25 passing. New coverage:
- YAML parses + normalizes to expected shape
- Malformed YAML falls back to builtins (warning, not raise)
- Each new provider routes its model id to the right base_url
- ANTHROPIC_AUTH_TOKEN alone satisfies third-party auth check
- Operator-set ANTHROPIC_BASE_URL overrides registry default
- Case-insensitive prefix match (MiniMax-M2 / minimax-m2.7 / GLM-4.6)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Aligns setup()'s third-party-model-without-URL handling with
create_executor()'s pre-validate (#19) — both unrecoverable
misconfigurations now raise ValueError at boot instead of one warning
and one raising.
Why: a third-party (mimo-*) model selected without ANTHROPIC_BASE_URL
sends every LLM request to api.anthropic.com with a non-Anthropic key,
401-ing every prompt. Workspace boots, looks "online" via heartbeat,
but is structurally broken on the user-facing path. The previous
warning-only path produced the same end-user symptom as the
2026-04-30 incident (workspace looks alive, every interaction fails)
just via a different misconfig shape.
Symmetry: create_executor raises when ANTHROPIC_BASE_URL is set to a
non-Anthropic host but no model is picked. setup() now raises when a
third-party model is picked but no URL is set. Together they catch
both halves of the misconfig surface at boot, before the workspace
enters "online" status.
Adds 4 setup() tests:
- raises on third-party + no URL
- passes on third-party + URL
- passes on OAuth alias (sonnet) + no URL
- passes on Anthropic API id (claude-*) + no URL
Stubs molecule_runtime.plugins.load_plugins as a no-op so the pass-path
tests run cleanly without the runtime installed. Test count: 11 (7
create_executor + 4 setup).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The 2026-04-30 staging incident traced back to workspaces booting with
ANTHROPIC_BASE_URL pointing at a non-Anthropic shim (MiniMax / OpenAI
gateway) but no explicit model configured. The adapter silently fell
back to "sonnet" — an Anthropic-native alias the upstream didn't
recognize — and the SDK --print probe hung 30s before timing out.
Platform's phantom-busy sweep then nuked the workspace at 10min,
producing "every workspace dead" with the root cause buried in a
30s subprocess hang.
Pre-validate the combo at adapter boot: when ANTHROPIC_BASE_URL host
is non-Anthropic AND no explicit model is set, raise ValueError with
an actionable message pointing to MODEL_PROVIDER / runtime_config.model.
Also log the resolved model + base_url_host every boot so future
failures explain themselves in the workspace logs without digging
into the SDK subprocess.
Tests live under tests/ with their own pytest.ini that anchors rootdir
there — keeps pytest from importing the package __init__.py (which
does the runtime-discovery relative import that requires
molecule_runtime installed). 7 tests cover: misconfig raises with the
right message, Anthropic-native passes, no-base-url passes, custom-url
+ explicit model passes, dataclass + dict shapes, unparseable URL
no-crash. CI runs them on every push/PR.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
