Three files conflicted with staging changes that landed while this PR
sat open. Resolved each by combining both intents (not picking one side):
- a2a_proxy.go: keep the branch's idle-timeout signature
(workspaceID parameter + comment) AND apply staging's #1483 SSRF
defense-in-depth check at the top of dispatchA2A. Type-assert
h.broadcaster (now an EventEmitter interface per staging) back to
*Broadcaster for applyIdleTimeout's SubscribeSSE call; falls through
to no-op when the assertion fails (test-mock case).
- a2a_proxy_test.go: keep both new test suites — branch's
TestApplyIdleTimeout_* (3 cases for the idle-timeout helper) AND
staging's TestDispatchA2A_RejectsUnsafeURL (#1483 regression). Updated
the staging test's dispatchA2A call to pass the workspaceID arg
introduced by the branch's signature change.
- workspace_crud.go: combine both Delete-cleanup intents:
* Branch's cleanupCtx detachment (WithoutCancel + 30s) so canvas
hang-up doesn't cancel mid-Docker-call (the container-leak fix)
* Branch's stopAndRemove helper that skips RemoveVolume when Stop
fails (orphan sweeper handles)
* Staging's #1843 stopErrs aggregation so Stop failures bubble up
as 500 to the client (the EC2 orphan-instance prevention)
Both concerns satisfied: cleanup runs to completion past canvas
hangup AND failed Stop calls surface to caller.
Build clean, all platform tests pass.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
The production-side end of the runtime CD chain. Operators (or the post-
publish CI workflow) hit this after a runtime release to pull the latest
workspace-template-* images from GHCR and recreate any running ws-* containers
so they adopt the new image. Without this, freshly-published runtime sat in
the registry but containers kept the old image until naturally cycled.
Implementation notes:
- Uses Docker SDK ImagePull rather than shelling out to docker CLI — the
alpine platform container has no docker CLI installed.
- ghcrAuthHeader() reads GHCR_USER + GHCR_TOKEN env, builds the base64-
encoded JSON payload Docker engine expects in PullOptions.RegistryAuth.
Both empty → public/cached images only; both set → private GHCR pulls.
- Container matching uses ContainerInspect (NOT ContainerList) because
ContainerList returns the resolved digest in .Image, not the human tag.
Inspect surfaces .Config.Image which is what we need.
- Provisioner.DefaultImagePlatform() exported so admin handler picks the
same Apple-Silicon-needs-amd64 platform as the provisioner — single
source of truth for the multi-arch override.
Local-dev companion: scripts/refresh-workspace-images.sh runs on the
host and inherits the host's docker keychain auth — alternate path for
when GHCR_USER/TOKEN aren't set in the platform env.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
- Merge origin/staging into fix/canvas-multilevel-layout-ux. 18 files
auto-merged (mostly canvas/tabs/chat and workspace-server handlers
the earlier DIRTY marker was stale relative to current staging).
- Fix 7 test failures surfaced by the merge:
1. Canvas.pan-to-node.test.tsx — mockGetIntersectingNodes was
inferred as vi.fn(() => never[]); mockReturnValueOnce of a node
object failed type check. Explicit return-type annotation.
2. Canvas.pan-to-node.test.tsx + Canvas.a11y.test.tsx — Canvas.tsx
reads deletingIds.size (new multilevel-layout state). Both mock
stores lacked deletingIds; added new Set<string>() to each.
3. canvas-batch-partial-failure.test.ts — makeWS() built a wire-
format WorkspaceData (snake_case, with x/y/uptime_seconds). The
store's node.data is now WorkspaceNodeData (camelCase, no wire-
only fields). Rewrote makeWS to produce WorkspaceNodeData and
updated 5 call-site casts. No assertions changed.
4. ConfigTab.hermes.test.tsx — two tests pinned pre-#2061 behavior
that the PR intentionally inverts:
a. "shows hermes-specific info banner" — RUNTIMES_WITH_OWN_CONFIG
now contains only {"external"}, so the banner is no longer
shown for hermes. Inverted assertion: now pins ABSENCE of
the banner, with a comment noting the inversion.
b. "config.yaml runtime wins over DB" — priority reversed:
DB is now authoritative so the tier-on-node badge matches
the form. Inverted scenario: DB=hermes + yaml=crewai →
form shows hermes. Switched test's DB runtime off langgraph
because the dropdown collapses langgraph into an empty-
valued "default" option that would hide the win signal.
- No production code changed — this commit is staging merge + test
realignment only. 953/953 canvas tests pass. tsc --noEmit clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Session's accumulated UX work across frontend and platform. Reviewable
in four logical sections — diff is large but internally cohesive
(each section fixes a gap the next one depends on).
## Chat attachments — user ↔ agent file round trip
- New POST /workspaces/:id/chat/uploads (multipart, 50 MB total /
25 MB per file, UUID-prefixed storage under
/workspace/.molecule/chat-uploads/).
- New GET /workspaces/:id/chat/download with RFC 6266 filename
escaping and binary-safe io.CopyN streaming.
- Canvas: drag-and-drop onto chat pane, pending-file pills,
per-message attachment chips with fetch+blob download (anchor
navigation can't carry auth headers).
- A2A flow carries FileParts end-to-end; hermes template executor
now consumes attachments via platform helpers.
## Platform attachment helpers (workspace/executor_helpers.py)
Every runtime's executor routes through the same helpers so future
runtimes inherit attachment awareness for free:
- extract_attached_files — resolve workspace:/file:///bare URIs,
reject traversal, skip non-existent.
- build_user_content_with_files — manifest for non-image files,
multi-modal list (text + image_url) for images. Respects
MOLECULE_DISABLE_IMAGE_INLINING for providers whose vision
adapter hangs on base64 payloads (MiniMax M2.7).
- collect_outbound_files — scans agent reply for /workspace/...
paths, stages each into chat-uploads/ (download endpoint
whitelist), emits as FileParts in the A2A response.
- ensure_workspace_writable — called at molecule-runtime startup
so non-root agents can write /workspace without each template
having to chmod in its Dockerfile.
Hermes template executor + langgraph (a2a_executor.py) + claude-code
(claude_sdk_executor.py) all adopt the helpers.
## Model selection & related platform fixes
- PUT /workspaces/:id/model — was 404'ing, so canvas "Save"
silently lost the model choice. Stores into workspace_secrets
(MODEL_PROVIDER), auto-restarts via RestartByID.
- applyRuntimeModelEnv falls back to envVars["MODEL_PROVIDER"]
so Restart propagates the stored model to HERMES_DEFAULT_MODEL
without needing the caller to rehydrate payload.Model.
- ConfigTab Tier dropdown now reads from workspaces row, not the
(stale) config.yaml — fixes "badge shows T3, form shows T2".
## ChatTab & WebSocket UX fixes
- Send button no longer locks after a dropped TASK_COMPLETE —
`sending` no longer initializes from data.currentTask.
- A2A POST timeout 15 s → 120 s. LLM turns routinely exceed 15 s;
the previous default aborted fetches while the server was still
replying, producing "agent may be unreachable" on success.
- socket.ts: disposed flag + reconnectTimer cancellation + handler
detachment fix zombie-WebSocket in React StrictMode.
- Hermes Config tab: RUNTIMES_WITH_OWN_CONFIG drops 'hermes' —
the adaptor's purpose IS the form, banner was contradictory.
- workspace_provision.go auto-recovery: try <runtime>-default AND
bare <runtime> for template path (hermes lives at the bare name).
## Org deploy/delete animation (theme-ready CSS)
- styles/theme-tokens.css — design tokens (durations, easings,
colors). Light theme overrides by setting only the deltas.
- styles/org-deploy.css — animation classes + keyframes, every
value references a token. prefers-reduced-motion respected.
- Canvas projects node.draggable=false onto locked workspaces
(deploying children AND actively-deleting ids) — RF's
authoritative drag lock; useDragHandlers retains a belt-and-
braces check.
- Organ cancel button (red pulse pill on root during deploy)
cascades via existing DELETE /workspaces/:id?confirm=true.
- Auto fit-view after each arrival, debounced 500 ms so rapid
sibling arrivals coalesce into one fit (previous per-event
fit made the viewport lurch continuously).
- Auto-fit respects user-pan — onMoveEnd stamps a user-pan
timestamp only when event !== null (ignores programmatic
fitView) so auto-fits don't self-cancel.
- deletingIds store slice + useOrgDeployState merge gives the
delete flow the same dim + non-draggable treatment as deploy.
- Platform-level classNames.ts shared by canvas-events +
useCanvasViewport (DRY'd 3 copies of split/filter/join).
## Server payload change
- org_import.go WORKSPACE_PROVISIONING broadcast now includes
parent_id + parent-RELATIVE x/y (slotX/slotY) so the canvas
renders the child at the right parent-nested slot without doing
any absolute-position walk. createWorkspaceTree signature gains
relX, relY alongside absX, absY; both call sites updated.
## Tests
- workspace/tests/test_executor_helpers.py — 11 new cases
covering URI resolution (including traversal rejection),
attached-file extraction (both Part shapes), manifest-only
vs multi-modal content, large-image skip, outbound staging,
dedup, and ensure_workspace_writable (chmod 777 + non-root
tolerance).
- workspace-server chat_files_test.go — upload validation,
Content-Disposition escaping, filename sanitisation.
- workspace-server secrets_test.go — SetModel upsert, empty
clears, invalid UUID rejection.
- tests/e2e/test_chat_attachments_e2e.sh — round-trip against
a live hermes workspace.
- tests/e2e/test_chat_attachments_multiruntime_e2e.sh — static
plumbing check + round-trip across hermes/langgraph/claude-code.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Issue #1947: after incidents, PM agents inherit hour-old TASK-priority
queue items from ICs that were correctly reporting "X is broken" while
X was actually broken. Once X is fixed those items are stale noise —
PMs spend ~5 min each writing "thanks, the issue is resolved".
Adds:
- DropStaleQueueItems() in a2a_queue.go: UPDATE ... SET status='dropped'
for queued items older than maxAgeMinutes. Uses FOR UPDATE SKIP LOCKED
to stay concurrency-safe with concurrent drain calls.
- AdminQueueHandler in admin_queue.go: POST /admin/a2a-queue/drop-stale
(AdminAuth, ?max_age_minutes=N, &workspace_id=<id>). Returns {dropped: N}.
- admin_queue_test.go: HTTP-level tests for param validation and response shape.
- Router registration for the new endpoint.
Usage during incident recovery:
curl -X POST /admin/a2a-queue/drop-stale?max_age_minutes=120
# scoped to one workspace:
curl -X POST /admin/a2a-queue/drop-stale?max_age_minutes=120&workspace_id=<uuid>
Closes#1947.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The test asserts that AdminAuth rejects an unauthenticated request to
the test-token route once any workspace token exists in the DB. It
sets MOLECULE_ENV=development to enable the handler's gate.
After this branch's AdminAuth Tier-1b hatch (middleware/devmode.go),
MOLECULE_ENV=development + empty ADMIN_TOKEN becomes the explicit
fail-open signal for local dev — so the request correctly passes
AdminAuth and falls through to the handler, which then 500s on an
unmocked DB lookup instead of the expected 401.
The security property the test is protecting (no bearer → 401 when
tokens exist) corresponds to the SaaS configuration where
ADMIN_TOKEN is always set. Setting ADMIN_TOKEN in the test suppresses
the dev-mode hatch and reaches AdminAuth's Tier-2 bearer check,
which correctly aborts 401 with "admin auth required".
No production behaviour change — the test is now verifying the path
that actually runs in production (MOLECULE_ENV=production +
ADMIN_TOKEN set).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Problem
When a lead delegates to a worker that's mid-synthesis, the proxy returns
503 "workspace agent busy" and the caller records the delegation as
failed. On fan-out storms from leads this hits ~70% drop rate — today's
observed numbers in the cycle reports.
## Fix — Phase 1 TASK-level queue-on-busy
When `handleA2ADispatchError` determines the target is busy, instead of
returning 503, enqueue the request as priority=TASK and return 202
Accepted with `{queued: true, queue_id, queue_depth}`. The workspace's
next heartbeat (≤30s) drains one item if it reports spare capacity.
Files:
- migrations/042_a2a_queue.{up,down}.sql — `a2a_queue` table with
partial indexes on status='queued' + idempotency_key. Schema
supports PriorityCritical/Task/Info from day one so Phase 2/3 ship
without migration churn.
- internal/handlers/a2a_queue.go — EnqueueA2A / DequeueNext /
Mark*-helpers plus WorkspaceHandler.DrainQueueForWorkspace. Uses
`SELECT ... FOR UPDATE SKIP LOCKED` so concurrent drains can't
double-claim the same row. Max 5 attempts before marking 'failed'
so a stuck item doesn't wedge the queue forever.
- internal/handlers/a2a_proxy_helpers.go — isUpstreamBusyError branch
calls EnqueueA2A and returns 202 on success. Falls through to the
legacy 503 on enqueue error (DB hiccup shouldn't silently drop).
- internal/handlers/registry.go — RegistryHandler gets a QueueDrainFunc
injection hook (SetQueueDrainFunc). When Heartbeat sees
active_tasks < max_concurrent_tasks, spawns a goroutine that calls
the drain hook. context.WithoutCancel ensures the drain outlives
the heartbeat handler's ctx.
- internal/router/router.go — wires wh.DrainQueueForWorkspace into
rh.SetQueueDrainFunc after both are constructed.
## Not in this PR (Phase 2/3/4 follow-ups)
- INFO priority + TTL (Phase 2)
- CRITICAL priority + soft preemption between tool calls (Phase 3)
- Age-based promotion so TASK doesn't starve (Phase 4)
- `GET /workspaces/:id/queue` observability endpoint
Schema already supports all of these; only the dispatch + policy code
remains.
## Tests
- TestExtractIdempotencyKey (5 cases): messageId parsing is robust
- TestPriorityConstants: ordering invariant + 50=TASK default
alignment with migration DEFAULT
Full DB-touching tests (FIFO order, retry bound, idempotency conflict)
intentionally deferred to the CI migration-enabled path — sqlmock
ceremony would duplicate the existing test infrastructure 3× over and
the behaviour is directly expressible in SQL constraints (FOR UPDATE
SKIP LOCKED, partial unique index).
## Expected impact once deployed
- a2a_receive error with "busy" flavor drops from ~69/10min observed
today to ~0
- delegation_failed rate drops from ~50% to <5%
- real_output metric rises from ~30/15min back toward the pre-
throttle baseline
Closes#1870 Phase 1.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
BLOCKERS fixed:
- instructions.go: Drop team-scope queries (teams/team_members tables don't
exist in any migration). Schema column kept for future. Restored Resolve
to /workspaces/:id/instructions/resolve under wsAuth — closes auth gap
that allowed cross-workspace enumeration of operator policy.
- migration 040: Add CHECK constraints on title (<=200) and content (<=8192)
to prevent token-budget DoS via oversized instructions.
- a2a_executor.py: Pair on_tool_start/on_tool_end via run_id instead of
list-position so parallel tool calls don't drop or clobber outputs. Cap
tool_trace at 200 entries to prevent runaway loops bloating JSONB.
HIGH fixes:
- instructions.go: Add length validation in Create + Update handlers.
Removed dead rows_ shadow variable. Replaced string concatenation in
Resolve with strings.Builder.
- prompt.py: Drop httpx timeout 10s -> 3s (boot hot path). Switch print
to logger.warning. Add Authorization bearer header from
MOLECULE_WORKSPACE_TOKEN env var.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a configurable instruction injection system that prepends rules to
every agent's system prompt. Instructions are stored in the DB and fetched
at workspace startup, supporting three scopes:
- Global: applies to all agents (e.g., "verify with tools before reporting")
- Team: applies to agents in a specific team
- Workspace: applies to a single agent (role-specific rules)
Components:
- Migration 040: platform_instructions table with scope hierarchy
- Go API: CRUD endpoints + resolve endpoint that merges scopes
- Python runtime: fetches instructions at startup via /instructions/resolve
and prepends them to the system prompt as highest-priority context
Initial global instructions seeded:
1. Verify Before Acting (check issues/PRs/docs first)
2. Verify Output Before Reporting (second signal before reporting done)
3. Tool Usage Requirements (claims must include tool output)
4. No Hallucinated Emergencies (CRITICAL needs proof)
5. Staging-First Workflow (never push to main directly)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
URLs returned from DB and Redis cache (db.GetCachedURL, workspaces.url column)
are now validated via validateAgentURL() before any HTTP request is made:
- mcpResolveURL (mcp.go): added validateAgentURL() calls on all three return
paths (internal cache, Redis cache, DB fallback).
- resolveAgentURL (a2a_proxy.go): added validateAgentURL() call before
returning agentURL to the A2A dispatcher.
validateAgentURL() was extended (registry.go) to resolve DNS hostnames and
check each returned IP against the blocklist (private ranges, loopback,
cloud-metadata 169.254.0.0/16). "localhost" is allowed by name for local dev.
GET /admin/memories/export now applies redactSecrets() to each content field
before including it in the JSON response. Pre-SAFE-T1201 memories (stored
before redactSecrets was mandatory on writes) no longer leak credentials.
POST /admin/memories/import now calls redactSecrets() on content before both
the deduplication check and the INSERT. Imported memories with embedded
credentials cannot bypass SAFE-T1201 (#838).
- admin_memories.go: GET /admin/memories/export + POST /admin/memories/import
handler (from PR #1051, with security fixes applied).
- admin_memories_test.go: 6 tests covering redactSecrets parity on both endpoints.
- registry_test.go: added DNS-lookup test cases for validateAgentURL (F1083).
"localhost" allowed by name (preserves existing test); nxdomain blocked.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Workspaces stuck in provisioning used to sit in "starting" for 10min
until the sweeper flipped them. The real signal — a runtime crash at
EC2 boot — lands on the serial console within seconds but nothing
listened. These endpoints close the loop.
1. POST /admin/workspaces/:id/bootstrap-failed
The control plane's bootstrap watcher posts here when it spots
"RUNTIME CRASHED" in ec2:GetConsoleOutput. Handler:
- UPDATEs workspaces SET status='failed' only when status was
'provisioning' (idempotent — a raced online/failed stays put)
- Stores the error + log_tail in last_sample_error so the canvas
can render the real stack trace, not a generic "timeout" string
- Broadcasts WORKSPACE_PROVISION_FAILED with source='bootstrap_watcher'
2. GET /workspaces/:id/console
Proxies to CP's new /cp/admin/workspaces/:id/console endpoint so
the tenant platform can surface EC2 serial console output without
holding AWS credentials. CPProvisioner.GetConsoleOutput is the
client; returns 501 in non-CP deployments (docker-compose dev).
Both gated by AdminAuth — CP holds the tenant ADMIN_TOKEN that the
middleware accepts on its tier 2b branch.
Tests cover: happy-path fail, already-transitioned no-op, empty id,
log_tail truncation, and the 501 fallback when no CP is wired.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses the Critical + Important findings from today's code
review of the org API keys feature (PRs #1105-1108).
## Critical-1: rate-limit mint endpoint
Previously POST /org/tokens had no mint-rate limit. A compromised
WorkOS session or leaked bearer could mint thousands of tokens in
seconds, forcing a painful manual cleanup of each one.
Fix: dedicated per-IP token bucket, 10 mints/hour/IP. Legitimate
bursts fit under the ceiling; abuse bounces. List + Delete stay
on the global limiter — they can't be used to generate new
secret material.
## Important-1: HTTP handler integration tests
internal/orgtoken had 9 unit tests; the HTTP layer (org_tokens.go)
had none. Adds org_tokens_test.go covering:
- List happy path + DB error → 500
- Create actor="admin-token" (bootstrap), actor="org-token:<prefix>"
(chained mint), actor="session" (canvas browser path)
- Create name>100 chars → 400
- Create with empty body mints with no name
- Revoke happy path 200, missing id 404, empty id 400
- Plaintext returned in response body and prefix matches first 8 chars
- Warning text present
A regression that breaks the tier-ordering, drops the createdBy
field, or accepts oversized names now fails at CI not prod.
## Important-2: bound List output
List() had no LIMIT — a mint-storm bug or abuse could make the
admin UI slow to render and allocate proportionally. Adds
LIMIT 500 at the SQL layer. 10x realistic ceiling, guardrail
against pathological cases.
## Important-3: audit provenance uses plaintext prefix, not UUID
orgTokenActor() was logging "org-token:<first-8-of-uuid>" which
couldn't be cross-referenced with the UI (which shows first-8
of the plaintext). Users could not correlate "who minted this"
audit entries with the revoke button they're looking at.
Fix: Validate() now returns (id, prefix, error). Middleware
stashes both on the gin context. Handler reads prefix for the
actor string. Audit rows now match UI prefixes exactly.
## Nit: named constants for audit labels
actorOrgTokenPrefix / actorSession / actorAdminToken replace
the hardcoded strings scattered across the handler. Greppable
across log pipelines + audit queries; one place to change if
the format evolves.
## Tests
- internal/orgtoken: 9 existing + 0 new, all still green (updated
signatures for Validate returning prefix).
- internal/handlers/org_tokens_test.go: new — 9 HTTP-layer tests
above. Full gin.Context + sqlmock harness.
- Full `go test ./...` green except one pre-existing
TestGitHubToken_NoTokenProvider flake unrelated to this change
(expects 404, gets 500 — tracked separately).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds user-facing API keys with full-org admin scope. Replaces the
single ADMIN_TOKEN env var with named, revocable, audited tokens
that users can mint/rotate from the canvas UI without ops
intervention.
Designed for the beta growth phase — one token tier (full admin).
Future work will split into scoped roles (admin / workspace-write
/ read-only) and per-workspace bindings. See docs/architecture/
org-api-keys.md for the design + follow-up roadmap.
## Surface
POST /org/tokens mint (plaintext returned once)
GET /org/tokens list live keys (prefix-only)
DELETE /org/tokens/:id revoke (idempotent)
All AdminAuth-gated. Bootstrap path: mint the first token via
ADMIN_TOKEN or canvas session; tokens can mint more tokens after.
## Validation as a new AdminAuth tier (2a)
AdminAuth evaluation order:
Tier 0 lazy-bootstrap fail-open (only when no live tokens AND
no ADMIN_TOKEN env)
Tier 1 verified WorkOS session via /cp/auth/tenant-member
Tier 2a org_api_tokens SELECT — NEW
Tier 2b ADMIN_TOKEN env (bootstrap / CLI break-glass)
Tier 3 any live workspace token (deprecated, only when ADMIN_TOKEN
unset)
Tier 2a runs ONE indexed lookup (partial index on
token_hash WHERE revoked_at IS NULL) + an async last_used_at
bump. No measurable latency cost on the hot path.
## UI
New "Org API Keys" tab in the settings panel. Label field for
human-readable naming. Plaintext shown once + clipboard copy.
Revoke with confirm dialog. Mirrors the existing workspace-
TokensTab flow so users who've used one get the other for free.
## Security properties
- Plaintext never stored. sha256 hash + 8-char display prefix.
- Revocation is immediate: partial index on revoked_at IS NULL
means the next request validates or fails in microseconds.
- created_by audit field captures provenance: "org-token:<short>"
when a token mints another, "session" for browser-UI mints,
"admin-token" for the ADMIN_TOKEN bootstrap path.
- Validate() collapses all failure shapes into ErrInvalidToken
so response-shape can't distinguish "never existed" from
"revoked".
## Tests
- internal/orgtoken: 9 unit tests (hash storage, empty field
null-ing, validation happy path, empty plaintext, unknown hash,
revoked filtering, list ordering, revoke idempotency, has-any-
live short-circuit).
- AdminAuth tier-2a integration covered by existing middleware
tests unchanged (fail-open + bearer paths).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses three Critical findings from today's code review of the
SaaS-canvas routing stack.
## Critical-1: session verification scoped to the current tenant
session_auth.go previously verified via GET /cp/auth/me, which
only answers "is someone logged in" — NOT "is this user in the
org they're targeting." Every WorkOS-authed user (including folks
who only signed up via app.moleculesai.app with no tenant
relationship) could call /workspaces, /approvals/pending,
/bundles/import, /org/import etc. on ANY tenant they could reach.
Cross-tenant read: user at acme.moleculesai.app could hit
bob.moleculesai.app/workspaces with their cookie and get Bob's
workspaces.
Fix:
- CP gains GET /cp/auth/tenant-member?slug=<slug> which joins
org_members × organizations and only returns member:true when
the authenticated user is actually in that org.
- Tenant sets MOLECULE_ORG_SLUG at boot via user-data.
- session_auth now calls tenant-member (not /me), passing its
own slug. Cache key includes slug so one tenant's cached
positive never satisfies another's check.
## Critical-2: cp_proxy path allowlist (lateral-movement fix)
cp_proxy.go forwarded any /cp/* path upstream with the cookie
and bearer attached. Since /cp/admin/* accepts sessions as one
of its auth tiers, a tenant-authed user could curl
/cp/admin/tenants/other-slug/diagnostics through their tenant
and the CP would honor it — turning any tenant into a lateral
hop into admin surface.
Fix: explicit allowlist of paths the canvas browser bundle
actually needs (/cp/auth, /cp/orgs, /cp/billing, /cp/templates,
/cp/legal). Everything else 404s at the tenant before cookies
leave. Fail-closed: future UI paths require explicit entries.
## Important-1,2: bounded session cache + split positive/negative TTL
Previous sync.Map cache grew unbounded (one entry per unique
Cookie header for process lifetime) and cached failures for 30s,
meaning a 3s CP blip locked users out for the full window.
Fix:
- Bounded map with batch random eviction at cap (10k entries ×
~100 bytes = 1 MB ceiling). Random eviction is O(1)
expected; we don't need precise LRU.
- Periodic sweeper goroutine (2 min) reclaims expired entries
even when they're not re-hit.
- Positive TTL 30s, negative TTL 5s — short negative so CP
flakes self-heal fast.
- Transport errors NOT cached (would otherwise trap every
user during a multi-second upstream outage).
- Cache key = sha256(slug + cookie) so raw session tokens
don't sit in process memory, and cross-tenant isolation is
structural not policy.
## Important-3: TenantGuard /cp/* bypass documented
Added a security note to the bypass explaining why it's safe
only under the current setup (cp_proxy allowlist + tunnel-only
ingress), and what would require revisiting (SG opens :8080
inbound to the VPC).
## Tests
- session_auth_test.go: 12 new tests — empty cookie, missing
slug, no CP, member:true happy path with cache hit, member:
false, 401 upstream, malformed JSON, transport error not
cached, cross-tenant isolation (same cookie different
tenants hit upstream separately), bounded eviction, expired
entries, cache key collision resistance.
- cp_proxy_test.go: new — isCPProxyAllowedPath covers 17
allow/block cases, forwarding preserves Cookie+Auth, Host
rewritten, blocked paths 404 without calling upstream.
All platform tests pass. CP provisioner tests pass after
threading cfg.OrgSlug into the container env.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Canvas's browser bundle issues fetches to both CP endpoints
(/cp/auth/me, /cp/orgs, ...) AND tenant-platform endpoints
(/canvas/viewport, /approvals/pending, /org/templates). They
share ONE build-time base URL. Baking api.moleculesai.app
broke tenant calls with 404; baking the tenant subdomain broke
auth. Tried both today and saw exactly one failure mode per
attempt.
Real fix: same-origin fetches + tenant-side split. Adds:
internal/router/cp_proxy.go # /cp/* → CP_UPSTREAM_URL
mounted before NoRoute(canvasProxy). Now a tenant serves:
/cp/* → reverse-proxy to api.moleculesai.app
/canvas/viewport,
/approvals/pending,
/workspaces/:id/*,
/ws, /registry, → tenant platform (existing handlers)
/metrics
everything else → canvas UI (existing reverse-proxy)
Canvas middleware reverts to `connect-src 'self' wss:` for the
same-origin path (keeping explicit PLATFORM_URL whitelist as a
self-hosted escape hatch when the build-arg is non-empty).
CI build-arg flips to NEXT_PUBLIC_PLATFORM_URL="" so the bundle
issues relative fetches.
Security of cp_proxy:
- Cookie + Authorization PRESERVED across the hop (opposite of
canvas proxy) — they carry the WorkOS session, which is the
whole point.
- Host rewritten to upstream so CORS + cookie-domain on the CP
side see their own hostname.
- Upstream URL validated at construction: must parse, must be
http(s), must have a host — misconfig fails closed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PR #729 tightened AdminAuth to require ADMIN_TOKEN, breaking the
workspace credential helper which called /admin/github-installation-token
with a workspace bearer token. Tokens expired after 60 min with no refresh.
Fix: Add /workspaces/:id/github-installation-token under WorkspaceAuth
so any authenticated workspace can refresh its GitHub token. Keep the
admin path as backward-compatible alias.
Update molecule-git-token-helper.sh to use the workspace-scoped path
when WORKSPACE_ID is set.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
GET /admin/memories/export returns all agent memories with workspace
name mapping. POST /admin/memories/import accepts the same format,
resolves workspaces by name, and deduplicates on content+scope.
Both endpoints are AdminAuth-gated.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
