OFFSEC-003 follow-up: a malicious peer could inject fake [A2A_ERROR],
[SYSTEM], [A2A_QUEUED] or [IGNORE] blocks via response_preview/summary
in delegation results. The agent's prompt sees these unescaped markers and
may interpret them as system commands.
Add _sanitize_a2a.py — a dedicated sanitizer that inserts ZERO-WIDTH SPACE
(U+200B) INSIDE the opening bracket of each known marker (e.g.
"[A2A_ERROR]" → "[A2A_ERROR]"). The raw marker string no longer
matches naive substring/regex checks while the text remains human-readable.
Apply sanitize_a2a_result() to both summary and response_preview fields
in read_delegation_results() before formatting them for prompt injection.
Tests: 18 new cases in test_sanitize_a2a.py covering marker escaping,
idempotency, injection scenarios, and edge cases. 1 integration case in
test_executor_helpers.py.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Root cause: the sop-tier-check.sh script uses jq extensively for all
JSON API parsing (whoami, labels, team IDs, reviews). Gitea Actions
runners (ubuntu-latest label) do not bundle jq — script exits at
line 67 with "jq: command not found", producing "Failing after 1-3s"
status on every staging PR.
Fix: add apt-get install -y jq step before the script run.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cherry-pick of d79a4bd2 from PR #318 onto fresh main base (PR #318 closed).
Issue #310: platform a2a-proxy logs ~300/hr
`timeout awaiting response headers` because ResponseHeaderTimeout was hardcoded
to 60s. Opus agent turns (big context + internal delegate_task round-trips)
routinely exceed 60s, so the proxy gave up before headers arrived even when
the workspace agent was healthy.
Changes:
- a2a_proxy.go: ResponseHeaderTimeout: 60s hardcoded →
envx.Duration("A2A_PROXY_RESPONSE_HEADER_TIMEOUT", 180s).
180s gives Opus turns comfortable headroom. The X-Timeout caller header
still bounds the absolute request ceiling independently.
- a2a_proxy_test.go: TestA2AClientResponseHeaderTimeout verifies the 180s
default and env-override parsing logic.
Env var: A2A_PROXY_RESPONSE_HEADER_TIMEOUT (e.g. 5m, 300s).
Closes#310.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Plugin adapters in molecule-skill-* repos do:
from plugins_registry.builtins import AgentskillsAdaptor as Adaptor
But _load_module_from_path() used exec_module() with a fresh module
namespace that did NOT have plugins_registry or its submodules in sys.modules,
causing:
ModuleNotFoundError: No module named 'plugins_registry'
Fix: before exec_module(), import and register plugins_registry + all three
submodules (builtins, protocol, raw_drop) in sys.modules so adapter imports
resolve correctly. Follows the Option 1 recommendation from issue #296.
Also adds test_resolve_plugin.py verifying the fix for both the
AgentskillsAdaptor import and the full InstallContext/resolve/protocol import.
Closes#296.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ports the bounded retry+backoff around each `git clone` in
scripts/clone-manifest.sh onto main, mirroring PR #298 which landed the
same change on staging. CI-infra carve-out: publish-workspace-server-image.yml
fires on `push: branches:[main]`, so the retry mitigation must be on main for
the workflow to be resilient to the OOM-killed-git-mid-clone flake
(`error: git-remote-https died of signal 9`, run 4622) when triggered by a
main push. Same one-file change as #298 (+45/-5), POSIX-sh, sh -n clean.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The Docker daemon health-check fix should not change which branches trigger
the build. Revert accidental addition of 'staging' to branch filters.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cover the canvas image publish workflow with the same `docker info`
guard added to publish-workspace-server-image.yml (commit 5216e781).
publish-canvas-image.yml was the only docker-build workflow still
missing the step.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The publish-workspace-server-image / build-and-push job clones the full
manifest (~36 repos) serially in the "Pre-clone manifest deps" step on a
memory-constrained Gitea Actions runner. Under host memory pressure the
OOM killer SIGKILLs git-remote-https mid-clone:
cloning .../molecule-ai-plugin-molecule-skill-code-review.git ...
error: git-remote-https died of signal 9
fatal: the remote end hung up unexpectedly
❌ Failure - Main Pre-clone manifest deps
exitcode '128': failure
Observed in run 4622 (2026-05-10, staging HEAD b5d2ab88) — died on the
14th of 36 clones, which red-lights CI and wedges staging→main.
Wrap each `git clone` in clone-manifest.sh with bounded retry + backoff
(3 attempts, 3s/6s), wiping any partial checkout between tries. A single
transient SIGKILL / network blip no longer fails the whole tenant image
rebuild. Benefits every caller of the script (publish-workspace-server-image,
harness-replays, Dockerfile builds, local quickstart).
This is a mitigation; the durable fix is more runner RAM/swap on the
operator host — tracked separately with Infra-SRE.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The A2A proxy can return three error shapes:
{"error": "plain string"}
{"error": {"message": "...", "code": ...}}
{"error": {"message": {"nested": "object"}}} ← value at .message is a string
builtin_tools/a2a_tools.py:72 called data["error"].get("message")
without guarding against error being a string, which raised:
AttributeError: 'str' object has no attribute 'get'
This broke every delegation attempt through the legacy a2a_tools path
(the LangChain-wrapped version used by adapter templates). The
SSOT parser a2a_response.py already handled string errors; the
legacy inline sniffer in a2a_tools.py did not.
Fix: branch on isinstance(err, dict/str/other) before calling .get().
Also update both publish-workflow files to remove the dead
`staging` branch trigger — trunk-based migration (PR #109,
2026-05-08) removed the staging branch.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The Molecule-AI GitHub org was suspended 2026-05-06; canonical SCM is
now git.moleculesai.app. external_connection.go was still emitting
github.com URLs in operator-facing copy-paste blocks, breaking
external-agent onboarding silently.
Per-site decisions (8 emit sites in 1 file):
- L124 (channel template doc comment): swap source-of-truth comment to
Gitea host.
- L137 /plugin marketplace add Molecule-AI/...: swap to explicit Gitea
HTTPS URL form. End-to-end-verified path per internal#37 § 1.A.
- L138 /plugin install molecule@molecule-mcp-claude-channel: marketplace
name is molecule-channel (per remote .claude-plugin/marketplace.json),
not the repo name. Fix to molecule@molecule-channel.
- L157 --channels plugin:molecule@molecule-mcp-claude-channel: same
marketplace-name fix.
- L179 user-facing GitHub URL: swap to Gitea.
- L261 pip install git+https://github.com/Molecule-AI/molecule-sdk-python:
not on PyPI; swap to git+https://git.moleculesai.app/molecule-ai/...
- L310 hermes-channel doc comment: swap source-of-truth comment.
- L339 pip install git+https://github.com/Molecule-AI/hermes-channel-molecule:
not on PyPI; swap to Gitea.
- L369 issue-tracker URL: swap to Gitea.
Verification:
- molecule-ai-workspace-runtime, codex-channel-molecule are on PyPI (200);
no swap needed for those pip lines (they were already package-name form).
- molecule-mcp-claude-channel, molecule-sdk-python, hermes-channel-molecule
are NOT on PyPI; swapped to git+https://git.moleculesai.app/molecule-ai/
form. All three repos are public on Gitea (default branch main) and
serve git-upload-pack unauthenticated (verified curl 200 against
/info/refs?service=git-upload-pack).
- Third-party github URLs (gin import, openai/codex, NousResearch/
hermes-agent upstream issue trackers, npm @openai/codex) intentionally
preserved.
Adds TestExternalTemplates_NoBrokenMoleculeAIGitHubURLs regression guard
to prevent the same broken URLs from re-emerging on future template
edits.
go vet / go build / existing TestExternal* — all clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two surfaces in workspace-server hardcoded `ghcr.io` and silently bypassed
the `MOLECULE_IMAGE_REGISTRY` env override that flips every other image
operation to the configured private mirror (e.g. AWS ECR in production):
1. internal/imagewatch/watch.go — image-auto-refresh polled
`https://ghcr.io/v2/...` and `https://ghcr.io/token` directly. Post-
suspension, with the platform pointed at ECR, the watcher silently
stopped seeing digest changes (every poll either 404'd or hung on a
registry it has no business talking to).
2. internal/handlers/admin_workspace_images.go — Docker Engine auth
payload pinned `serveraddress: "ghcr.io"`, so when the operator sets
`MOLECULE_IMAGE_REGISTRY=…ecr…/molecule-ai` the engine matched the
wrong credential entry on every authenticated pull.
Fix: extract `provisioner.RegistryHost()` returning the host portion of
`RegistryPrefix()` (e.g. `ghcr.io` ← `ghcr.io/molecule-ai`, or
`004947743811.dkr.ecr.us-east-2.amazonaws.com` ← the ECR mirror prefix),
and route both surfaces through it. Default behavior is unchanged for
OSS users on GHCR.
Tests
- New `TestRegistryHost_SplitsHostFromOrgPath` and
`TestRegistryHost_NeverEmpty` pin the helper across GHCR / ECR /
self-hosted Gitea / bare-host edge cases.
- New `TestGHCRAuthHeader_RespectsRegistryEnv` asserts the Docker auth
payload's `serveraddress` follows MOLECULE_IMAGE_REGISTRY (and never
leaks the org-path suffix).
- New `TestRemoteDigest_RegistryHostFollowsEnv` stands up an httptest
server, points MOLECULE_IMAGE_REGISTRY at it, and confirms both the
token endpoint and the manifest HEAD land there — i.e. the full image-
watch loop respects the env override end-to-end.
Both new tests were verified to FAIL on the pre-fix code path before the
helper was wired in, so a future revert can't silently re-introduce the
bug.
Out of scope (followup needed)
ECR uses `aws ecr get-authorization-token` (SigV4 + basic-auth) instead
of GHCR's `/token?service=…&scope=…` flow. This PR makes the URL host-
configurable; the bearer-token negotiation in `fetchPullToken` still
speaks the GHCR flavor. On ECR with `IMAGE_AUTO_REFRESH=true`, the
watcher will now fail loudly at the token fetch (logged per tick) rather
than silently hitting ghcr.io. Operators on ECR should keep
IMAGE_AUTO_REFRESH=false until ECR auth is wired — tracked as a separate
task. Net effect of this PR alone is strictly better than pre-fix:
fail-loud > silent-broken.
Refs: RFC #229 P2-4
tier:low
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two bugs in yaml-utils.ts toYaml():
1. tools: [] was only emitted when config.tools.length > 0,
but the test asserts it's always present. Add blank-line
separator + unconditional list("tools", ...) so MINIMAL_CONFIG
with tools: [] renders correctly.
2. Nested list values (e.g. runtime_config.required_env: [KEY])
were serialized as " required_env: KEY" (stringification of the
array) instead of a YAML list block. Fix obj() to detect
Array.isArray(sv) and emit a list block with 4-space indent.
Closes#269.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Self-delegation deadlocks: the sending turn holds `_run_lock`, the receive
handler waits for the same lock, the A2A request 30s-times-out, and the
whole cycle is wasted (the Dev Lead system prompt warns agents off this by
hand — "Never delegate_task to your own workspace ID … there is no peer who
is also you"). The platform/runtime had no guard. Now both
`tool_delegate_task` and `tool_delegate_task_async` early-return an
actionable error when `workspace_id == effective_source` (`source_workspace_id
or _peer_to_source[target] or WORKSPACE_ID`) — before `discover_peer`, so no
network round-trip is wasted either. A genuinely different target (incl.
another of a multi-workspace agent's own registered workspaces) is
unaffected.
Tests: tests/test_a2a_tools_delegation.py — new TestSelfDelegationGuard (4
cases: rejects own ID; rejects when source_workspace_id explicitly == target;
async path rejects; a different target passes the guard through to
discover_peer). `pytest tests/test_a2a_tools_delegation.py` → 12 passed.
(tests/test_a2a_tools_impl.py's TestToolDelegateTask* suite is red on this
PC2/Windows checkout — same on `main` without this change; httpx-mock infra,
not this PR — CI validates on Linux.)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
[core-lead-agent] PR #281 merged — handles string-form errors in a2a_tools.delegate_task (was raising AttributeError on every delegation through legacy path), fixes empty-parts dict regression (#279), and drops the dead staging branch trigger from both publish workflows. Replaces the abandoned PR #268 + #277. Integration Tester unblocked for mesh recovery validation.
internal#226 follow-up #1. `molecule_runtime.config` resolves the picked
model as `MOLECULE_MODEL` > `MODEL` > (legacy) `MODEL_PROVIDER` (#280) —
this side of the boundary now matches:
- applyRuntimeModelEnv reads `MOLECULE_MODEL` ahead of `MODEL` /
`MODEL_PROVIDER`, and exports BOTH `MOLECULE_MODEL` and `MODEL`
(the latter kept for back-compat with everything that already reads
`os.environ["MODEL"]`). So a workspace whose secrets carry
`MOLECULE_MODEL` (the unambiguous name) is honoured, and the
`MODEL_PROVIDER` misnomer — which got set to provider slugs
("minimax") and even runtime names ("claude-code") — is the lowest-
priority fallback, exactly as on the runtime side.
- the resolution-order comment is updated to flag MODEL_PROVIDER as the
legacy-and-misleadingly-named var.
Also drops a stray trailing `}` in delegation_test.go (committed in
97768272 "test(delegation): add isDeliveryConfirmedSuccess helper") that
made `internal/handlers` fail to parse — one of the things keeping the
package from compiling for tests.
Tests: TestApplyRuntimeModelEnv_SetsUniversalMODELForAllRuntimes extended
to assert MOLECULE_MODEL mirrors MODEL on every case, plus two new cases
(MOLECULE_MODEL env fallback; MOLECULE_MODEL beats MODEL_PROVIDER). Could
not run `go test ./internal/handlers/` locally — the package is still
blocked behind `internal/plugins` `SourceResolver` redeclaration (the
#248 plugin-router/resolver refactor, Core-BE's lane); CI validates once
that lands. The applyRuntimeModelEnv change is mechanical (same shape as
the existing `MODEL` handling) — reviewer please eyeball.
Companion: molecule-core#280 (runtime config.py side), molecule-ai-workspace-template-claude-code#14 (CLI-stream-error surfacing).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
