stack/docs/scratchpads/webui-fleet-bridge-plan.md

WebUI Fleet Bridge Planning Scratchpad

Mode: Task 0 docs-only decision PR authorized; Task 1 and all executable/runtime work remain blocked. Owner: Velma Opened: 2026-08-09 Scope: One enrolled Agent Host launching one native mosaic claude OAuth session and streaming it into apps/web through apps/gateway.

Objective

Turn the approved direction in jarvis-brain/docs/scratchpads/MOSAIC-WEBUI-FLEET-BRIDGE.md into a test-first implementation plan while preserving Fred's harness-home/launcher contract and Scooby's greenfield safety findings.

Source reconciliation

• Current origin/main: b0f7d26dd9c14d91eaaefc35d6c9fd6618a0bd92.
• Current origin/next: 4df478cdd150fdf8d52ea109f02ade5d85017acd.
• Branches currently diverge (main has 11 unique commits; next has 13). next contains local-tier Redis fix #689; main contains later fleet/shell fixes.
• Fred's three-root harness-home design and promotion stack are not yet fully present on either baseline.
• Therefore neither current SHA is an acceptable implementation pin. Code may begin only from a Fred-certified reconciled SHA containing the required launcher/home contract and safe Gateway startup prerequisites.
• The deployed mosaic.woltje.com v0.0.20 UI remains reference-only.

Verified seams

• Current browser chat uses an in-process Pi SDK session.
• AgentRuntimeProvider supports list/tree/stream/send/attach/detach/terminate, but not create/start.
• InteractionController enrolls an already-existing runtime session; it cannot launch one.
• Hermes is the only runtime provider registered in Gateway.
• Tmux streaming is explicitly unsupported and remains out of scope.
• mosaic claude is the authoritative launcher and accepts Claude's machine-facing stream-json flags.
• Installed discovery version: Claude Code 2.1.226. Target Distrobox version must be independently pinned and certified.

Non-negotiable dependencies

1. Fred approves the machine-facing launcher/seat-home contract before code.
2. No WebUI/Gateway direct read of lease broker state, daemon socket, or state files (F-V3).
3. No provider OAuth token leaves the Agent Host.
4. No local Gateway/Web startup around the KBN/database hold.
5. Greenfield work runs in a Debian Distrobox with an isolated home.
6. The initial plan PR targets next; Fred binds D1–D15 on its exact head before issue/PRD/tracker completion or implementation work.

Reproducible evidence

Run from a clean Stack clone:

git fetch origin main next
git rev-parse origin/main origin/next
git rev-list --left-right --count origin/main...origin/next
rg -n "interface AgentRuntimeProvider|createSession|streamEvents|terminate" \
  packages/types/src/agent packages/agent/src apps/gateway/src/agent
rg -n "AgentService\.prompt|interaction_sessions|createRuntimeTerminationApproval" \
  apps/gateway/src packages/db/src/schema.ts

Primary inspected source seams:

• packages/types/src/agent/agent-runtime-provider.ts
• packages/agent/src/{runtime-provider-registry,hermes-runtime-provider,matrix-native-runtime-provider,tmux-fleet-runtime-provider}.ts
• apps/gateway/src/agent/{runtime-provider-registry.service,interaction.controller,durable-session.repository,durable-session.service}.ts
• apps/gateway/src/chat/chat.gateway.ts
• packages/mosaic/src/commands/{launch,interaction}.ts
• packages/mosaic/src/fleet/generated-env-boundary.ts
• packages/db/src/schema.ts
• apps/web/src/app/(dashboard)/chat/page.tsx

Planning-only investigation transcripts are local and intentionally uncommitted:

• /tmp/velma-plan-stack-surface.txt
• /tmp/velma-plan-structure.txt
• /tmp/velma-plan-scooby.txt
• /tmp/velma-plan-runtime-contract.txt

Source findings that constrain the design

• interaction_sessions.id is the stable primary key; there are no create/policy/enrollment/state columns.
• interaction_outbox has a unique (session_id, idempotency_key) index and only pending | processing | delivered.
• Baseline DurableSessionRepository.create() can replace provider/runtime identity for the same owner; M1 must remove that implicit mutation.
• Baseline termination approval is Redis-backed and currently consumes separately from PostgreSQL; M1 therefore needs durable authorization acceptance before destructive token deletion/dispatch.
• Existing interaction HTTP base is /api/interaction/:agentName; the plan extends it rather than inventing a second route family.
• Baseline launch.ts/lease launcher still use ambient lookup/literal interpreters. Section 4.1 is non-binding consumer input to Fred's W-F design; W-F's final resolved-launch contract must exist in the certified base before Velma can certify it.
• Root pnpm test is not KBN-safe: it includes PGlite migration and framework-shell/lease-broker suites.

Independent draft review

Seven adversarial review rounds found and drove explicit fixes for:

• a candidate resolved-seat consumer descriptor and threat model, now explicitly non-binding input to Fred's W-F-owned launcher design;
• one active launch per stable conversation, exact pending/failed encodings, durable pre-dispatch reservation, CAS activation, and crash lookup without a migration;
• generation-bound enrollment, command/event revalidation, SIGHUP config epochs, and stale-epoch rejection;
• Redis/PostgreSQL exact-stop crash safety via non-destructive verify, durable authorization acceptance, atomic claim/GETDEL, and same-operation status reconciliation;
• sequence gaps/reorder limits, deterministic UUIDv5 completion, and Gateway restart fail-closed behavior;
• shared streaming redaction before host ring/transport and again before Gateway persistence/browser;
• D2/D12-selected path-free provenance and a separate safe browser DTO—never raw or hashed path strings;
• exact F-V3 boundary: bridge has no broker API, while Fred's sealed launcher may enforce broker policy internally;
• migration-free focused tests in Tasks 1–8, with live OAuth, repository transaction, Gateway/Web, and Playwright restricted to the Fred/Scooby-certified Task 9 path;
• candidate private/public commitment and artifact-binding mechanisms that W-F may accept, simplify, replace, or defer;
• JCS event-digest recomputation and equal/different duplicate handling in both accepted and future-buffer states;
• a safe browser presentation DTO for host/workspace/seat/persona labels, readiness, and connection state;
• explicit OnApplicationBootstrap create/stop recovery enumeration with no auto-launch;
• all POSIX/Windows/UNC/file/tilde path classes in the streaming redactor and definitive failed-stop response semantics;
• the full 14-column canonical task schema, fake-only Task 5 repository tests, migration-free certified Task 9 DB test, server-owned operation correlations/routes, and exact merged-SHA smoke;
• per-commit independent review, queue guards, exact-head PR review, squash merge, exact merged-next SHA/CI wait, worktree-bound smoke, issue-state readback, and reviewed tracker-closure PR with its own merged-next CI;
• a private DB URL loader captured/exported per session without recording or echoing the credential;
• an attached Task 9 implementation branch, provider-filtered exact issue-state readback, and machine-verified smoke JSON binding source/worktree/deployed SHA;
• a capped, duplicate-key-rejecting, exact-key/type smoke schema so report extensions or JSON boolean/float coercion cannot smuggle data or fake child-count evidence.

The plan remains deliberately decision-PR-ready, not implementation-ready. Fred authorized only Task 0's initial two-document PR on next. He must still return every Section 2 value, replace all [FRED-GATE] entries, certify storage/startup, and provide the W-F-dependent IMPLEMENTATION_BASE_SHA before Task 0 closes or Task 1 starts.

Current status

• Gitea principal verified as velma; helper and API wrapper resolution are fail-closed and correct.
• Fred authorized the initial Task 0 docs-only PR in comms/20260809T094952Z__from-fred__ec0e85.md and confirmed its next lane/W-F descriptor corrections in comms/20260809T095437Z__from-fred__4ee79d.md.
• Provisional decisions: D1/D8/D9/D10/D13/D14 approved; D4 tool labels exactly {Read, Grep, Glob}; D5/D12 provisional; D11 remains a single-operator seam; D2/D3/D6/D7 and IMPLEMENTATION_BASE_SHA are BLOCKED-ON-W-F.
• Awaiting exact plan-PR-head review and Fred's complete D1–D15 binding contract.
• No source code, database, deployment, or live runtime changes made.