Workflow REST adapter (B86)

The existing HTTP adapter is opt-in, loopback-only and token authenticated. Every workflow route requires a full remote token: snapshots, revisions and ledgers contain user text. No command is added to READ_ONLY_COMMANDS. Select a registered project with directory / x-opencode-directory or workspace / x-mendcode-workspace. Run IDs cannot cross that boundary, including native /command requests.

Routes and payload schemas are included in /doc and /openapi.json. Malformed payloads/queries and invalid transitions return 400, missing resources 404, workspace mismatch/read-only authority 403, unsupported donor plan features (retry limits, human output schemas or inputs, unrecorded historical runs) 501. JSON bodies use the existing byte limit. Event/artifact arrays expose X-Next-Before and X-Truncated; snapshot pagination is also in pagination. Lists stop before the native frame byte budget and mark X-Truncated. Empty/truncated history never claims evidence that was not retained.

Plan compatibility

Plans accept native v2 JSON or v1 formatVersion: 1 JSON. The donor form maps phases (name, ordinal, taskIDs), task kinds agent, synthesize, verify, validate, map, dependencies, selectors, output contracts, model/provider pairs, agent profiles, tool restrictions, workspaces, budgets, completion criteria/checks and overlap policy to native execution. Verify/validate agent tasks run read-only. Donor task/phase identifiers remain visible in HTTP and are mapped to bounded internal slugs. Donor JSON is retained on the immutable revision, preserving original metadata and IDs across restarts; native callers use their native IDs and plans. Output captures also enter the existing bounded ledger. Ledger IDs/event types remain native v2 identities inside donor-shaped envelopes; this is not an exact v1 SDK ABI.

Required gates, human tasks, per-task retry and per-plan retention are accepted. Each requiredGates entry (and a human task) becomes a native human task that waits with no session; the run is needs_input until it is answered. A human task takes no inputs, and its output is none or text (the approval note). Answer with POST /workflow/:runID/approve-task or /reject-task ({taskID, reason?}, full token; gate-N or the gate id for gates), a native extension of the donor API; rejecting fails the task. The model cannot approve. retry maps to {maxAttempts, backoffMs, retryOn}: with neither maxAttempts nor retryOn it is 3 transient attempts after backoffMs (default 1 s), retryOn alone means no retry, and maxAttempts above 6 or backoffMs above one hour return 501. Budget failures never retry. budget.retention (maxArtifacts, maxAgeMs, maxBytes) prunes the run's events and artifacts in the ledger; receipts are kept and the run is marked truncated. Native bounds remain: 32 task nodes including map overhead, four concurrent tasks, supported JSON-schema keywords, and the native text/plan byte limits. Historical native runs without immutable revision metadata cannot produce the full donor snapshot and return 501; their native APIs remain available. The native map scheduler also invalidates dependents and later phase outputs on retry, preserving its stronger stale-output protection.

Completion uses v2 contracts: donor completion criteria become required native criteria; explicit donor completion requests a fresh independent audit. Legacy donor plans use native automatic audit rather than v1 same-run semantic closure. No provider is called by preview/save. Starting and completion use the configured engine; tests use local fake ports only. Real provider compatibility is not live-tested.

Conservative decisions

• The HTTP listener stays off by default, loopback only, with token auth; no new listener, dependency or external service.
• IDs, state ownership, storage, model catalog and ledger retention remain native. Loop IDs are durable provenance; worktree reuse uses the existing origin-session lease, not a new loop runtime.
• overlapKey is a project-scoped idempotency key; concurrent repeats return the existing run, including terminal runs. It is separate from plan overlap scheduling.
• Permission changes are durable runtime overrides; they do not rewrite the frozen plan or task settings. Session aliases map conservatively: approval → ask, smart → edits, full_access → full, global_default/null → inheritance. Effective permission never exceeds the initial grant or inherited restrictions. Captured live policies update at the next tool/provider boundary; an operation already in flight cannot be undone. Report-only changes reject workflows with host checks, which cannot safely be retroactively restricted. Native custom retains its existing approval semantics.
• The native client adds setWorkflowPermissions and retryWorkflowPhase alongside existing typed save/start/get/ledger helpers. Terminal UI uses the existing workflow factories and panel grid; no new chrome or keys.

Script workflows (workflows.md) are not exposed here: a route that reads a script run answers 501, and script runs are left out of the run list.

App-harness frames use the actual HTTP adapter, workflow host, SQLite records and renderer with local pending-task ports:

bun scripts/render-frames.ts .agents-frames/B86 --b86 100x30 180x50