Coordinate bounded native publication progress during frame advance while preserving watchdogs, guest budgets and resource ownership. Add controlled and packaged reopen regressions, clarify the runtime specs, and retain qualification evidence.
24 KiB
Headless Automation Protocol v1
Domain: runtime host tooling. Function: normative.
This chapter defines the v1 automation wire contract owned by the headless host. It does not change guest ABI, firmware, VM, frame/input semantics or the desktop DevTools protocol.
The observation model is explained in Deterministic Observation Must Follow the Active Producer.
Process and transport
prometeu headless --stdio starts the distributed prometeu-headless tool from
the same installation. One process owns one HeadlessHost. No display, native
audio, network service or realtime pacing is required. A client must use its
configured installation; missing binaries/capabilities are explicit failures.
Transport is UTF-8 JSON objects separated by LF (CRLF is accepted). stdout is
protocol-only; diagnostics go to stderr. Fragmented reads must be accumulated;
allocation is bounded by the command limit. An oversized line is rejected and
drained through LF before reading the next request. EOF with an incomplete line
reports invalid_request, then closes the session. EOF always requests shutdown.
All sizes exclude the LF delimiter and include the complete JSON envelope and
base64 payload. Responses exceeding their bound become response_too_large,
never truncated data. Clients must continuously consume stdout, including events.
Backpressure may block output; the supervising client owns wall-clock deadlines
and may terminate the process tree. There is no transport replay or durable queue.
Handshake and compatibility
The server first emits a hello object with sessionId, protocolVersion
(major: 1, minor: 2), real runtimeVersion, capabilities and limits.
Session identity is opaque and process-local, excluded from deterministic oracles.
Before work, the client sends initialize with protocolVersion and an array
requiredCapabilities. An incompatible major returns incompatible_version;
missing capabilities return unsupported_capability. The session remains
uninitialized and cannot execute work. Compatible minor additions are permitted;
the initialize result returns the lesser supported minor. Initialize is accepted
once; repeated initialization returns already_initialized. Shutdown is permitted
even before initialization. All other work requires successful initialization.
Envelopes
{"requestId":"a1","command":"advance","params":{"frames":60}}
requestId is a nonempty string of at most 128 UTF-8 bytes, selected by the
client; concurrent requests must have distinct identities. command and params
are required. Unknown fields and invalid parameter types are rejected. The
request schema is distributed in devtools/headless-protocol/request.schema.json;
response.schema.json defines hello, progress and every response/result shape.
Valid and invalid request fixtures are shipped beside the schemas. Process tests
validate actual server output against the response schema (including image results).
{"type":"response","sessionId":"opaque","requestId":"a1","result":{"requested":60,"completed":60,"stop":"completed"}}
{"type":"response","sessionId":"opaque","requestId":"a1","error":{"code":"busy","message":"...","details":{}}}
A response contains exactly one of result/error. Errors have stable code,
human-readable message and structured details. Unparseable/unusable request
identity is represented by null; valid identity is preserved on parameter errors.
Malformed JSON, invalid schema and unknown commands must never disappear silently.
Wire requestId is correlation, not deduplication: the client must not retry an
advance blindly after a lost response. Studio task-level clientRequestId and
operationId are a separate boundary, not part of this wire protocol.
Commands
| Command | Parameters | Result/meaning |
|---|---|---|
| initialize | protocolVersion, requiredCapabilities | negotiated version and capabilities |
| boot | path (nonempty directory-cartridge path) | current state; existing boot preconditions apply |
| reset | empty object | reset state, clears input/fault/frame correlation |
| set_input | complete snapshot fields listed below; omitted fields reset to defaults | input replaced, not patched |
| clear_input | empty object | input restored to defaults |
| advance | frames: unsigned integer, including zero | requested, completed, stop, optional fault/error details |
| capture | empty object | PNG/base64 plus canonical frame metadata |
| get_state | empty object | pc, stackTop, logicalFrameIndex, tickIndex, appId, status |
| get_fault | empty object | fault object or null |
| get_logs | required afterSequence: null/uint64, limit: 1..1024 | native log page; requires negotiated 1.2 log_read |
| cancel | targetRequestId | acknowledgement that cancellation was requested for that active advance |
| shutdown | empty object | acknowledgement after the active operation stops and host is dropped |
Input snapshot fields are up, down, left, right, a, b, x, y, l,
r, start, select, touch (booleans), and touchX, touchY (signed i32).
Omitted booleans are false and omitted coordinates are zero. The ordinary runtime
begin-frame sampling derives press/release/hold edges. No new input semantics exist.
State statuses are running, paused, terminated, faulted; running does
not mean background execution. Stack values preserve their type through tagged
kind/value objects; i64 values are decimal strings and float values carry an
exact round-trippable string (including non-finite values). Faults expose kind,
pc, summary and structural trap code/opcode/message/source span or init-error code.
No client must parse a summary to distinguish trap, panic or initialization failure.
Resource traps include a nullable trap.resource object with resource, reason,
requested and limit; resource and reason are enumerated machine values.
initError is discriminated by code: registry_load_failed includes load,
runtime_resource_limit includes resource, entrypoint_arguments_invalid
includes entry, and loader_patch_failed includes binding. Other initialization
codes carry only their code and human message. load.code uses the stable
LOAD_* tag and retains typed reason, section/offset, record/member, ID domain,
referenced ID, function/PC and quota context, with null for unavailable fields.
Binding context preserves canonical native identity, version, capabilities or
declared/expected ABI counts as applicable. A load-relative PC remains nested
inside its diagnostic; initialization failure does not invent a runtime PC.
The published response schema rejects unknown fields and invalid contexts.
Ordering, cancellation and lifecycle
Only one host operation is admitted at a time. Conflicting work, including state
inspection while the host is executing, returns busy; it is not implicitly
queued. The command reader remains available for cancel/shutdown while an advance
executes. Cancellation targets an active advance's requestId; no active matching
advance returns not_cancelable. An acknowledgement is not a completed cancellation.
Advancement is repeated advance_frame(), checking cancellation between frames.
The current frame keeps the existing deterministic host-tick watchdog. Zero frames
is a no-op. Results preserve requested/completed counts and stop values:
completed, faulted, terminated, canceled, host_error.
Guest faults are outcomes, not malformed-request errors. Host errors include a
stable code and known completed progress; they do not manufacture guest faults.
Each completed frame emits type: progress, requestId, sessionId, requested and
completed counts before the terminal response. A client can retain known progress
if the process dies. The final response is authoritative when delivered.
Shutdown/EOF requests cancellation, waits for the active work to reach its boundary, drops the host and exits; explicit shutdown receives its response after cleanup. Negotiated memcard shutdown may refuse with native state and keep the channel available, as specified in Memcard controls. Disconnect never starts/resumes execution. A failed process requires a new explicit session, not replay. Host session loss does not implicitly release Studio authority.
Capture
Capture returns mimeType: image/png, encoding: base64, data, width, height,
stridePixels, pixelFormat: RGBA8888, logicalFrameIndex and frameId.
The envelope carries sessionId; the Studio consumer adds build correlation.
Frame identities/counters on this wire are unsigned JSON integers and must be
decoded losslessly. logicalFrameIndex and frameId have no numeric equality rule.
PNG stores visible RGBA channels in row order, losslessly, without stride padding.
stridePixels describes the original canonical buffer. Before a completed frame,
including after reset, capture returns no_frame_available; it cannot reuse an
old run. No desktop screenshot, hash-only substitute, temporary-image-file transport
or continuous video is permitted.
Configurable limits and supervision
| CLI option / handshake field | Default |
|---|---|
| --max-frames / maxFrames | 600 |
| --max-command-bytes / maxCommandBytes | 65536 |
| --max-response-bytes / maxResponseBytes | 33554432 |
| --handshake-timeout-ms / handshakeTimeoutMs | 5000 |
| --call-timeout-ms / callTimeoutMs | 60000 |
| --shutdown-timeout-ms / shutdownTimeoutMs | 5000 |
Limits must be positive; message bounds must be at least 4096 bytes to admit the handshake and bounded errors. Limits apply to serialized bytes, not characters. Clients must enforce the advertised handshake/call deadlines and 5-second default cancel/shutdown grace, then terminate only the process tree they own. Deadlines are host supervision, not simulated guest time. Runtime enforces frame/message bounds without clamping. Requests over limits fail explicitly. Clients configure startup timeouts before seeing the handshake, then verify the announced values.
Conformance
The runtime gate includes fragmented/invalid/oversized lines, negotiation failures,
all commands, busy/cancel/shutdown races, EOF, real CLI distribution invocation,
PNG pixel equality, independent logical/render identity, repeat/reset determinism,
structured faults, progress on partial failure and retained desktop run/debug.
Neither Studio UI nor an AI model is required. See tests/automation_process.rs
in prometeu-cli and the in-process deterministic-control tests in the headless crate.
Native log observation
Protocol minor 1.2 adds capability log_read. Clients MUST negotiate minor >= 2
and include log_read in requiredCapabilities before using get_logs.
Requiring it with an older minor returns incompatible_version; using it without
negotiation returns unsupported_capability. Production clients negotiating 1.0
or 1.1 retain their existing command/result shapes and media behavior.
{"requestId":"init","command":"initialize","params":{"protocolVersion":{"major":1,"minor":2},"requiredCapabilities":["log_read"]}}
{"requestId":"logs","command":"get_logs","params":{"afterSequence":null,"limit":128}}
Send each request after the preceding response. afterSequence is a required
nullable exclusive uint64 cursor. Null includes the first retained event, including
sequence zero. limit is required and MUST be an integer in 1..1024. Unknown fields,
missing parameters, strings/fractional/negative cursors and invalid limits return
invalid_params. A cursor greater than newest, or any non-null cursor on an empty
stream, returns invalid_cursor.
The result contains sessionId, logEpoch, events, oldestSequence,
newestSequence, nextAfterSequence, gap and overflow. Events include complete
sequence, frame, timestampMs, level, source, appId, tag and message.
Levels are trace, debug, info, warn, error; sources are pos, hub,
vm, fs, app. appId is uint32 for app and null for all other sources.
tag is uint16. Sequences, frames, timestamps and epochs are unsigned JSON uint64
integers; clients MUST decode them losslessly. Native elapsed-time timestamps
MUST NOT determine deterministic verdicts. Messages are native stored strings,
without assertion parsing, filtering, formatting or additional truncation.
Events are returned in increasing native sequence order across all sources.
Retained oldest/newest bounds are both null for an empty ring. The next cursor is
the final returned sequence, or the input cursor for an empty page. Requerying an
unchanged ring returns the same events; it never drains records. Reads MUST NOT
tick the firmware, execute guest code or change lifecycle, input, media or faults.
Existing single-operation admission applies: reading during an advance returns
busy. Complete envelope byte bounds apply: an oversized page or single event
returns response_too_large, without a partial result. Retry a smaller page if
appropriate; a single event that cannot fit remains an explicit error.
Retention, cursors and reset
The ring retains at most 4096 native records. overflow reports whether retention
has ever evicted records in this log stream. It is historical information, not
proof of loss for this reader. gap reports missing history relative to the
requested cursor: null with retained oldest > 0, or a non-null cursor whose next
sequence is less than retained oldest. overflow: true, gap: false is a valid
page for a reader that consumed the evicted records already. Readers MUST use
cursor continuity and gap to detect unread retention loss, rather than reject
historical overflow alone. For example, oldest=100 and afterSequence=150 has no
gap; afterSequence=98 or null has a gap. No gap is hidden by source/tag filtering.
logEpoch is zero for the existing process-owned log stream. VM reset retains
the ring, native sequences and epoch; frame counters may reset independently.
It MUST NOT clear or renumber logs to simulate a new stream. A different process
has a different sessionId and its own sequence zero. Clients MUST bind cursors
and deduplication to sessionId/logEpoch; a numeric cursor alone is not a portable
identity between processes. A native sequence already applied MUST NOT be
applied again when inspecting overlapping pages.
Producer limits and consumer responsibility
The unchanged guest logging path accepts at most 10 messages per application per
logical frame, with at most 256 UTF-8 bytes per message. Longer guest messages
are truncated before insertion; excess calls are dropped and produce one native
warning for that application/frame. Dropped calls receive no native sequence, and
truncation does not create a native gap. Consequently gap: false does not establish
that every attempted guest event was stored. Reading cannot recover those losses.
Structured-event consumers MUST bound their encoded messages and per-frame emission, accounting for other logs from the same application. They MUST validate their own producer event continuity and explicit completion (for example, ordinals and final event counts), rejecting missing, malformed or incomplete reports. The provider supplies generic native observations and does not interpret suites, assertions or success. Do not enlarge guest budgets to transport test reports.
Consumers advancing the guest MUST read all pending pages before the next frame
block, including after boot and completion/failure when available. Choose
conservative blocks within actual retention headroom; maxFrames is an admission
ceiling, not a safe polling interval. At 10 app messages/frame, 600 frames can
already exceed the 4096-record ring, independently of system messages. Bounded
polling does not guarantee lossless delivery for arbitrary production rates or
unbounded work. Response loss MUST NOT trigger automatic advance replay.
Memcard controls
Protocol minor 1.1 adds capability memcard. Initialize negotiates the lesser
minor. These commands and parameters require minor >= 1 and memcard in
requiredCapabilities; unsupported use fails before mutation. Clients negotiating
1.0 keep the previous command/result shapes. No executable-version inference,
default image, alternate transport or headless media startup flags are defined.
| Command | Parameters | Meaning |
|---|---|---|
| memcard_mount | path, action: open/create, access: read_write/read_only | prepare explicit media before boot; bounded result |
| memcard_unmount | empty object | release idle media outside a resident session |
| get_memcard_state | empty object | coherent native observation, without guest execution |
| boot | path, optional memcardChoice: persistent/memory_fresh/memory_existing | explicit player choice required for a memcard cartridge |
| memcard_reconcile | empty object | recover selected same-card identity while suspended |
| memcard_continue_in_memory | empty object | explicit safe SO detach, preserving RAM |
| reset / shutdown | optional discardUnsaved (false), pendingWrites: wait/cancel (wait) | controlled teardown; no autosave |
Mount requires no resident game or outstanding IO. Create/read_only is invalid. Paths are resolved relative to process cwd, reported absolutely; parent directories must exist. Open does not create; create never replaces, converts or repairs media. Persistent requires writable coherent media; memory_existing requires a readable record for the cartridge appId. No error changes that choice automatically. After an admitted boot attempt, retry requires explicit reset; parsing and negotiation failures do not admit boot. Reset preserves selected media/lock.
State has closed objects media, session, io, recovery, a nullable request,
dirty and dirtySlots. Media reports path/cardId/access/preparation/error/health;
session reports identity/choice/mode/boot phase/residency/action requirement.
Requests report original identity, operation, local slot, state, deadline reason,
storage error and retention, independently of IO activity. cardId is 32 lowercase
hex digits; request IDs and mount epochs are decimal strings. Unknown values are
null, never invented outcomes. Schemas distributed with the installation specify
all fields. Inspection never releases requests or exposes private save payloads.
Domain failures use memcard_error with details.operation, details.reason
and details.state. Reasons include choice_required, unavailable, corrupt,
unsupported, read_only, access_denied, already_exists, busy, no_space,
out_of_memory, invalid_state, invalid_range, counter_exhausted, identity_mismatch,
unsaved_changes, indeterminate and restart_required. Human messages are not codes.
Original worker error, deadline cause and durable result are separate observations.
Shutdown stops an active advance at its existing frame boundary, then evaluates
IO and dirty RAM. Wait/cancel only concern already admitted work. After IO ceases,
remaining dirty RAM without discard consent returns unsaved_changes and leaves
the session/channel usable. Terminal results retained by the guest alone do not
block closure. Unconfirmed IO cessation never returns closed:true or permits
resource/lock release. Successful negotiated reset adds memcard to the state
result; shutdown returns closed:true and its final memcard snapshot, after checked
cleanup. EOF/output loss attempts bounded cancellation without save/resumption;
process death does not establish publication outcome.
An advance stopped for resident media action returns stop:action_required with known completed progress. Resident state is paused; a host dialog is not counted as a game frame. Idle IO polling leaves guest counters/input unchanged. Reconcile accepts no replacement path, validates expected identity before changes, preserves RAM/continuation/original retained results and does not rerun initialization.
When successful bootstrap remains blocked by its pending native publication, the
owner coordinates lane progress within the same advance, outside machine ticks,
with requested waits of at most 5 ms. This does not apply to an unrelated gameplay
commit. Native deadlines remain 500 ms for cancellation and 1 s for stop/escalation,
measured from the original request admission. One additional 2 s deadline from
first wait entry in an advance_frame does not reset on wake/reentry or extend
those deadlines. Existing native action-required/fault boundaries take precedence.
If the dependency remains pending at the outer deadline without such a boundary,
the advance returns stop:host_error, exact known completed progress, and
error.code:memcard_error with details.operation:advance,
details.reason:restart_required and details.state holding the native snapshot.
The guest finishes paused with the host step request cleared. This preserves
worker ownership until actual cessation, the 100,000-tick frame watchdog, transport
responsiveness and cancellation between frames; it introduces no intraframe
cancellation, guest fault, simulated progress, command replay or new protocol shape.
Runtime qualification uses neutral PBX/ABI fixtures, real distinct processes and an isolated extracted distribution. It proves SUCCESS, termination, exact readback without resave and stable appId across cartridge versions. Consumer source-language compilation and client adaptation are separate work; the provider has no frontend build dependency. No global installation or consumer configuration update is needed.
Integration sequence
Start prometeu headless --stdio from the selected installation and consume its
hello before sending work. Require the capability explicitly:
{"requestId":"init","command":"initialize","params":{"protocolVersion":{"major":1,"minor":1},"requiredCapabilities":["memcard"]}}
{"requestId":"media","command":"memcard_mount","params":{"path":"/absolute/save.pmc","action":"open","access":"read_write"}}
{"requestId":"boot","command":"boot","params":{"path":"/absolute/cartridge","memcardChoice":"persistent"}}
{"requestId":"step","command":"advance","params":{"frames":1}}
{"requestId":"save-state","command":"get_memcard_state","params":{}}
{"requestId":"close","command":"shutdown","params":{"pendingWrites":"wait"}}
Each request above follows the successful response to the preceding one. Creating
new media requires a deliberate action:create; do not retry an open failure
with create. Memory choices are memory_fresh and memory_existing. No client
may replace a failed persistent choice with either automatically.
If closure reports unsaved_changes, keep the process/session and expose that
native state to the caller. A later explicit choice may send
{"discardUnsaved":true,"pendingWrites":"cancel"}; do not infer that consent
from a UI stop, timeout or process error. A successful commit is the guest/native
request state success, independently of its reason; neither an advance response
nor closed:true proves that a save occurred.
On reopening, launch a different process, open the same image and choose the
appropriate mode explicitly. The neutral process oracle verifies exact prior
bytes without resaving and repeats with a changed cartridge version and unchanged
appId. Run cargo test -p prometeu-cli --test automation_process memcard_public_distinct_process_save_readback_and_version_identity; add
--release to qualify that profile independently. For an extracted archive,
set PROMETEU_PROTOCOL_TEST_INSTALL and run the ignored
packaged_memcard_process_persistence test. It invokes the installed dispatcher
from another cwd without modifying any consumer project or installation.
Program Registry integration
When exposing initialization failures, retain the structured load error code, typed reason and available context. Human-readable messages are not a parsing contract; a load error has no fabricated runtime PC. Failed loads must not publish executable sessions. Existing protocol request names and transport behavior are unchanged.
Heap reference inspection values use {"kind":"heap_ref","value":{"index":u32,"generation":u32}}. Both components are required to distinguish a live incarnation from a reused slot; the wire representation must not truncate handles to indices. Inspection does not grant cross-session transfer authority.