Repository navigation
Conversation
alt-glitch
added
type/feature
matrix.process_edits was documented and tested as working via config.yaml, but _YAML_LOWER_KEYS (the apply_yaml_config_fn hook that turns config.yaml's matrix: block into env vars) never listed it, so setting matrix.process_edits: true in config.yaml silently did nothing — only the MATRIX_PROCESS_EDITS env var actually worked. Add the missing tuple entry (mirroring process_notices) and a test that drives the real _apply_yaml_config bridge instead of only asserting against a hand-constructed PlatformConfig(extra=...).
The initial edit option used a global boolean and trusted the relation inside m.new_content. An edit could therefore change another author's input, route a correction into a different thread, or enter the active turn through the gateway's interrupt or steering policy. Configure corrections with exact room IDs and boolean values, defaulting to no enabled rooms. Validate the original event's room and author with the receiving adapter's client, then use its original thread relation. Apply the current sender, room, notice and mention policies. Typed edits must supply genuine replacement content because mautrix serialization can manufacture m.new_content. Encrypted validation uses the existing client crypto and replacement reader. Queue corrections through the existing pending slot and FIFO, replacing pending edits to the same original in their existing position. Recheck both events and current authorisation in shared turn preparation, so cold and recursive queued turns reject withdrawn corrections even after cache eviction. Edited commands remain external conversational text. Append the correction as a new user turn with its original event ID. Earlier model messages and the system prompt remain unchanged. Keep the existing opening-event reply anchor for final delivery in a queued chain. Two parameterised invariants reproduce the original policy, routing and active-turn failures. Six additional rows reproduce a withdrawn pending replacement after cache eviction. All 50 rows pass with the repair, and the adjacent gateway gate passes 766 tests across 39 files.
With busy_input_mode: queue, the base adapter debounces ordinary text that arrives during a turn and merges the buffer into the pending slot when the slot holds an event from the same sender. A Matrix correction waiting in that slot is a text event from the same sender, so the user's next message was appended to it. At turn start the adapter compares the queued correction with the server's copy of the edit. The merged text no longer matched, so the gateway dropped both the correction and the new message without a reply. Debounced text is now never merged with an event that waits for a turn boundary. The text stays in the debounce buffer while the correction occupies the slot, and the existing flush at the end of each turn moves it into the slot as its own turn. When a correction arrives while text is already buffered, the adapter flushes the buffer first, so that text runs before the correction, in the order the user sent them.
Edit handling and turn-start validation resolved a correction's context with the original event ID. _resolve_message_context records that ID as the thread's latest event, which bot messages without an explicit reply target use for their reply fallback, and it sends a read receipt for it. Each correction therefore moved the thread's fallback target back to the original and sent a receipt for an older event. The edit handler did this before it checked the sender's authorisation. _resolve_message_context now accepts record=False, which leaves the thread trackers and the read receipt to the caller. The correction handler resolves the context this way and checks authorisation first. Only an accepted correction marks its thread and sends a read receipt, and the receipt is for the edit event. Turn-start validation records nothing.
A correction event's message ID is the ID of the m.replace event. Clients display a replacement as the original message and do not show the edit event itself, so the processing, success and failure reactions that the adapter placed on the edit were invisible. Corrections now use the original event as the reaction target.
The bridge test set MATRIX_PROCESS_EDITS before applying the YAML configuration and then asserted that the value was unchanged. apply_yaml_bridge never overwrites a variable that is already set, so the test would also pass if a process_edits bridge row came back. The test now starts with the variable unset and asserts that it stays unset. With a bridge row restored, it fails.
The edit follow-up section did not say which message a correction's reply refers to. A correction's reply is a Matrix reply to the original message when the correction starts its own turn or runs before the end of a queued chain. When the correction is the last queued message to run after a busy turn, the gateway sends the reply as a Matrix reply to the message that started that turn, as it does for every queued message. Document this, and document that corrections are never merged with other messages, that the model receives the corrected text as typed, and that processing reactions appear on the original message.
The effective-state parent added requires_mention as the sixth field of the tuple from _resolve_message_context. The correction handler still unpacked six fields, so every accepted edit raised ValueError before an event was built, and no correction reached the gateway. Unpack the new field as well. The handler does not use requires_mention itself: _build_inbound_event reads it from the same context and records matrix_requires_mention on the correction as it does for other messages.
The effective-state parent removed fetch_mention_context and take_turn_channel_context from the adapter, and the gateway now calls use_turn_context on the snapshot from fetch_inbound_context. The correction tests stubbed all three methods, so monkeypatch failed on the missing attributes, and a None snapshot would fail in the core. Drop the stubs so turn preparation uses the adapter's real snapshot and prepare_turn_context. The hook looks up the session through the runner's session store, so the cold runners get a SessionStore in tmp_path, and the room identity fake has room_state instead of the removed members_digest.
The effective-state parent moved the Matrix client helpers into client_events.py, which provides Method with its own stand-in for environments without mautrix. The correction module still had a private try/except that fell back to read_context's re-export. Import Method from client_events, as the other Matrix context modules do, so the plugin has one fallback definition.
A correction must wait for a turn boundary and take its own turn. The
busy-session handler already queues a turn-boundary event through the
FIFO. The runner has a second route to a running session:
_hm_handle_running_session_message, reached from _handle_message when
the runner has a running turn for the session but the adapter has no
session guard. That route ignored _queue_at_turn_boundary.
The state is reachable in production. The CLI handoff runs its
synthetic turn through _handle_message directly, so the adapter has no
guard for that session for the whole turn. Auto-resume after a restart
claims the pending-agent sentinel before its task reaches
handle_message, and expects inbound messages in that window to reach
this route. A correction arriving then was merged into the pending text
while the agent was starting, and in interrupt or steer mode it was
delivered into the running turn.
Skip the pending-agent merge for turn-boundary events and send them
through the queue-mode branch, which uses the same FIFO as the busy
handler. The drain check still comes first, as in the busy handler.
The new test builds a correction with the real Matrix adapter and passes
it to the real runner while the session has a running turn and no
adapter guard. Before the change it failed in three of four modes: the
pending-agent case produced "earlier text\nlatest correction", interrupt
mode called interrupt("latest correction"), and steer mode steered the
correction into the turn. Queue mode already queued it.
The effective-state parent now contains NousResearch#125648, which quotes the triggering event once per threaded response. A send anchored on an event that Hermes has already answered in the thread is a continuation: it sets is_falling_back and points at the latest event in the thread. A correction that waits behind a paused turn runs as the last message of the queued chain. Its answer is anchored on the message that started the paused turn, as for every queued message, and the first response already answered that message. The native test still expected both responses to be genuine replies to that message, so the plain and encrypted enabled cases failed with is_falling_back true and m.in_reply_to pointing at the first response. The client now expects the first response to quote the blocked message and each later response to fall back to the response before it, in the expected order. The Matrix guide describes the correction's reply in that case as a thread message after the running turn's answer.
…b it In busy_text_mode queue, text that arrives during a turn waits in a per-session debounce buffer. _flush_text_debounce_now moves the buffer into the adapter's pending slot, but only when the slot is empty or its event may absorb the text. Two senders' texts may not merge, and since the Matrix edit follow-ups a queued correction absorbs no text at all. When the slot could not absorb the buffer, the flush returned without moving it and cancelled its timer. The buffer then waited for the next flush. _queue_text_debounce made that loss visible. A message from a different sender first flushes the buffer. When the flush failed and the slot could not absorb the new message either, the function returned without storing it, and nothing logged the drop. Threads are shared between their participants by default, so in queue mode a correction followed by texts from two people lost the second text. On main the same happened to a third sender's text while the slot contained text from someone else. The unflushed buffer also broke arrival order: text sent between two corrections ran after the second one, because the second correction went to the runner's FIFO while the text stayed in the buffer. When the slot cannot absorb the buffer, hand the buffered event to the runner's _queue_or_replace_pending_event, which appends it to the FIFO behind the slot with the same cap and album rules as other busy follow-ups. The adapter already reaches the runner through gateway_runner for delivery scope and live-adapter lookups. The buffer is then empty, so the next sender starts a fresh buffer instead of reaching the drop. Each sender's burst still runs as one turn, in arrival order. An adapter without a runner keeps the previous behaviour, because it has no FIFO. The platform-neutral test sends text from two senders while the slot contains a third sender's text, and while it is empty. Before the change, "three" was missing in both rows and a buffer stayed stranded with no timer. The Matrix test covers a pending correction followed by text from Alice and Bob, and from Bob and Carol, with and without the timer firing between them. Before the change the second message was lost in all four rows.
Text that the user sends after one correction and before another must run between them. Before the previous commit, the text stayed in the debounce buffer because the first correction in the pending slot could not absorb it, while the second correction went to the runner's FIFO. The observed order was $edit1, $edit2, $text. The debounce flush now hands the text to the FIFO behind the first correction, and the flush before the second correction's busy handling runs first, so the order is $edit1, $text, $edit2. This test asserts that order. It failed on the parent of the previous commit with [$edit1, $edit2, $text].
A correction's m.new_content has no reply fallback, so a leading quote in it is the user's own text. After the merge of the effective-state parent, three conditions in the adapter strip a quote only when reply_fallback is true and _has_reply_fallback recognises a legacy fallback. The existing quote-on-reply row starts with "> quoted line", which _has_reply_fallback rejects because it has no sender pill, so that row passed even with reply_fallback removed from the conditions. Add rows whose body quotes Bob with a pill and a blank line, which is the shape of a legacy fallback, for an original that is a reply and for one in a thread. With "reply_fallback and" removed from the three conditions, both rows fail with "my answer" in place of the full quoted body. The two older rows pass under that mutation.
A started input can be deferred into the FIFO overflow and rescued by a later turn. After the rescued turn finishes, the runner clears its durable active-turn marker before reporting processing completion. Cancellation during that storage await skips the completion hook, so the input remains marked as processing even though its turn has finished. Run rescued-input completion in the marker cleanup finally block. Keep both awaits after the generation-scoped session and lease releases, and retain the adapter-owned marker handoff and marker token comparison. Extend the marker cancellation contract with the actual adapter start and defer lifecycle followed by FIFO rescue. Both normal and rescued turns release ownership and admit a successor; the successful rescued input also reports exactly one successful completion after cleanup cancellation. Historical attribution for original source 51c0ecf6197ce1489fb31bdcbddfb64c3d2ba125 (audit record 755). Public PR NousResearch#126283 already includes this cancellation correction in 48902ee. This empty commit records the missing original contribution credit. The source tree is unchanged.
Matrix message dispatch and context admission remain in the oversized adapter facade. Opt-in correction handling needs both entry paths. Move their existing methods into MatrixInboundEventMixin before adding that behaviour, preserving every dispatch, mention, thread, timestamp and receipt decision. Resolve facade-bound helpers lazily so existing runtime SDK bindings and patch targets remain effective. Declare the actual MatrixAdapter host for the constructed and admitted events. Update the module guide to include its expanded topic.
Speech reference tests used empty history, so they did not exercise channel-state acknowledgement while references appeared in stored transcript rows and adapter-generated context. Extend the existing A/B/A contract with empty and stored history for all five speech variants. Check the complete prepared message, unchanged transcript, lookup and adapter calls, returned channel state and single raw authored preprocessing input. Stored references and generated channel notes remain literal while caption and successful speech references expand under the owning profile. The complete file passes 21 cases with strict asynchronous warnings and zero retries. This adds coverage to the existing preparation behaviour; it does not repair a reproduced production defect.
Preparation can derive a user-input receipt from the transport message before shutdown assigns a pending UID. Reconstruction then derives a second receipt from that UID, so an already persisted user turn appears unexecuted. Input without a transport ID also minted a new random receipt on each call. Cache a typed owner with the actual transport namespace and bind the saved owner to the pending UID at capture. Reconstruct that binding only after checking its profile, workspace, conversation, pending UID and derived receipt. Event aliases and selective withdrawal retain the aggregate owner even when the first contribution is removed. Existing transport-ID receipts keep their original derivation. The real A/B/A SessionStore contract reproduces the mismatch at the prepare-before-capture, withdrawal and alias boundaries. Capture before preparation remains a control. Scope and receipt changes in a saved record are refused. This records an execution identity; RAM admission still does not consume a pending recovery record.
Speech reference tests used empty history, so they did not exercise channel-state acknowledgement while references appeared in stored transcript rows and adapter-generated context. Extend the existing A/B/A contract with empty and stored history for all five speech variants. Check the complete prepared message, unchanged transcript, lookup and adapter calls, returned channel state and single raw authored preprocessing input. Stored references and generated channel notes remain literal while caption and successful speech references expand under the owning profile. The complete file passes 21 cases with strict asynchronous warnings and zero retries. This adds coverage to the existing preparation behaviour; it does not repair a reproduced production defect.
Discord event normalization and Slack's inbound guard still occupy their large adapter facades. Move those methods into their platform siblings, with deferred facade bindings for runtime SDK and helper resolution. Extract the pending withdrawal traversal behind its existing sender predicate so native deletion notifications can reuse every pending store. The traversal and existing deletion policy remain unchanged. Preserve every complete current Discord notification and Slack intake guard in its extracted owner, including current subtype admission and workspace checks. Retain all current Discord mixins and the required deferred SDK and helper bindings. Keep the optional changed-message result separate until the existing None guard, then assign the narrowed event.
Cached text can contain an opaque marker. Exercise normal and restored queue execution with the actual model request path so the test detects transcript substitution, role changes and profile owner leaks before the request reaches the model.
An accepted leftover /steer can arrive while provisional pending input has already filled the queue. The drain ignored FIFO admission failure, so it dropped the correction while keeping the provisional input. The branch that prioritised a steer over a queued event also restored that event directly without releasing its provisional dispatch reservation. Retain an unadmitted steer for the next turn. Check each pending input through the recognised-command guard before combining the remaining text, so a residual command is discarded without losing the correction. Restore deferred FIFO events through the existing dispatch helper, which releases reservations while preserving order. Three assertion failures reproduce the lost correction with ordinary provisional text, the residual /help safety case, and the restored FIFO state on the parent. The FIFO contract releases each returned dispatch before measuring queue depth, as the caller does in normal operation. All eight contracts in the existing file pass after the fix. Integrate the pending-admission correction with the current pinned channel-input constructor. The drain now uses the actual FIFO admission result, preserves unadmitted steering text and releases a deferred event reservation through the existing restore helper. The complete original regression contracts and assertions remain unchanged.
A steer accepted by the running agent can return unconsumed while earlier queued input and provisional dispatches fill the pending queue. Admission then fails when the drain moves the steer to its own turn. Dropping that accepted input loses the correction and its processing completion. Restore the tracked steer through the existing pending-dispatch helper and park its lifecycle before transferring completion ownership. Restoration preserves accepted input, while the admission threshold still rejects new messages. The earlier queued event keeps its own dispatch and runs first; the restored steer retains its message identity and completes with its own turn or with cancellation if the queue is stopped. The regression drives the real runner and adapter with a scripted model, queued input and a full set of provisional reservations. It checks model input, admission rejection, processing outcomes and retained reservations for successful delivery and /stop.
The receipt teardown contract observed the former text-only flush call. The current adapter writes a complete pending snapshot before clearing its slots, so that observer missed a successful write. Wrap the actual snapshot writer and retain the completion, ordering and empty-slot assertions. The wrapper still writes the real snapshot. Supply joined room members for the current Matrix authorization check. Combine the native feedback settings with the existing paused-queue configuration in one display section, and retain the additive auxiliary configuration fixture required by the unread tool. These fixtures use the current adapter and configuration paths without replacing their behaviour assertions. The fixture contribution is split between the receipt and unread layers. Record its complete rationale in the receipt history. The current source uses the real snapshot observer, the original membership assertions and the consumed auxiliary fixture, with one parsed YAML display document. The paused-context public fixture selects input mode `queue` and text mode `interrupt`. Busy follow-ups enter the runner FIFO because interruption requires effective input mode `interrupt`. The queued-context consumer checks retained quoted context, the preceding prompt prefix and eventual dispatch; its contract does not require adapter debounce or coalescing. This accepts direct FIFO admission as the adaptation of the former queued-text fixture. Normal-mode steer cases keep their explicit override. This records contributor credit and the public fixture adaptation without changing the source tree.
Native SimpleX chat items include the quoted message separately from incoming content. The receiver retained that raw payload but omitted the incoming item ID and quote fields when constructing MessageEvent. The shared batcher therefore treated replies to different targets as unquoted input, and the prompt could omit the quoted text. Populate the existing reply fields before batching or direct dispatch. Keep the incoming item ID separate from the quoted target. Use a shared message context identity when the local target ID is unavailable, or an identity specific to the incoming item when neither target ID exists. Derive ownership and author data from the quoted direction, including an unknown group author, and reuse the normaliser for deferred voice input. Two parameterised native-payload invariants reproduce the missing event metadata and batching boundaries on the exact prerequisite. They also check complete prompt framing and deferred file replay on the repair. Retain the captured UTC timestamp when normalising native SimpleX reply context. Keep the dataclass import, complete quote normaliser and incoming identity handling from this contribution.
Native SimpleX reply tests compare complete events and prompts after the common gateway batches incoming text. The gateway records the additional message IDs in a merged event and JSON-escapes quoted text in the prompt. The imported tests omitted those IDs and placed literal newlines inside the quotation. Include the second message ID in coalesced expected events and use JSON encoding for the complete expected quote. Keep the split-event cases and whole-event and whole-prompt comparisons unchanged. The canonical 40-case run reproduced 24 merged-event assertion failures and 16 prompt assertion failures on the unadapted tests. The native reply additions introduced formatter changes in the reply context, extended event construction and their fixtures. Format those scopes and remove the extra blank pair at the fixture insertion boundary. Keep unrelated inherited formatting unchanged. Complete module ASTs, parameter populations and assertions remain identical to the accepted 155-case source, so its behaviour evidence remains applicable.
The complete feedback and edit source adds a second auxiliary-config argument to the shared native gateway fixture. Native collection fails before the unread controls can run. Keep the single additive argument that the controls fixture already provides. The incoming edit and lifecycle contracts also read optional mock-call records and assign dynamic fixture fields as statically declared methods. Check that the required calls occurred before reading their arguments, use the real profile-hook registry, and bind fixture replacements through monkeypatch. All original behaviour assertions remain in place. The complete changed production, unit and native harness type comparison has no new diagnostics after the defining-owner correspondence checks. The native collection syntax error is removed; collection-only execution is diagnostic evidence, not a test pass. Source and contributor reconciliation for original 726 (1655eb9d3aa2d693af7f8685185c80850408d82d). The original message above records the composed correction and its historical checks. This contribution preserves its complete credit after the source repairs were incorporated into their separate public owners. The five edit-follow-up fixtures use direct awaited-call guards and the real ProfileHookRegistries on NousResearch#126312. The lifecycle, deferred-input and feedback fixtures use the declared typed and monkeypatch adaptations on NousResearch#126283. These complete unit definitions are also present in its current composed tree. The native auxiliary configuration fixture has one argument and its consumer on NousResearch#126398. The NousResearch#126283 native harness has no auxiliary fixture or consumer, so it requires no corresponding argument. This is source and credit attribution across those owners. It adds no product or fixture patch, and records no fresh test, type, SDK or native execution. The complete original message and both literal contributor dates are preserved.
iainlane
added a commit
to iainlane/hermes-agent
that referenced
this pull request
Oct 6, 2026
The complete feedback and edit source adds a second auxiliary-config argument to the shared native gateway fixture. Native collection fails before the unread controls can run. Keep the single additive argument that the controls fixture already provides. The incoming edit and lifecycle contracts also read optional mock-call records and assign dynamic fixture fields as statically declared methods. Check that the required calls occurred before reading their arguments, use the real profile-hook registry, and bind fixture replacements through monkeypatch. All original behaviour assertions remain in place. The complete changed production, unit and native harness type comparison has no new diagnostics after the defining-owner correspondence checks. The native collection syntax error is removed; collection-only execution is diagnostic evidence, not a test pass. Source and contributor reconciliation for original 726 (1655eb9d3aa2d693af7f8685185c80850408d82d). The original message above records the composed correction and its historical checks. This contribution preserves its complete credit after the source repairs were incorporated into their separate public owners. The five edit-follow-up fixtures use direct awaited-call guards and the real ProfileHookRegistries on NousResearch#126312. The lifecycle, deferred-input and feedback fixtures use the declared typed and monkeypatch adaptations on NousResearch#126283. These complete unit definitions are also present in its current composed tree. The native auxiliary configuration fixture has one argument and its consumer on NousResearch#126398. The NousResearch#126283 native harness has no auxiliary fixture or consumer, so it requires no corresponding argument. This is source and credit attribution across those owners. It adds no product or fixture patch, and records no fresh test, type, SDK or native execution. The complete original message and both literal contributor dates are preserved.
iainlane
added a commit
to iainlane/hermes-agent
that referenced
this pull request
Oct 6, 2026
The shared receipt fixture inserts its configured read-receipt and reaction settings into the Matrix platform block. The unread test also specified those settings in raw YAML, so composition created duplicate keys. Configuration parsing failed and the gateway used an unconfigured provider before the observer could establish the test session. Select disabled automatic receipts and reactions through the shared MatrixFeedbackSettings fixture. Keep one platform configuration block and preserve the original scoped unread and explicit receipt assertions. The fresh plain and encrypted native run reproduced the duplicate-key configuration error on the composed parent. No production behaviour or native expectation changes in this correction. (cherry picked from commit 2d573c3cb434410412ecfb6c86a8085c798d6d09) Public source qualification: the complete unread feedback fixture effect is already present on NousResearch#126398. Positional feedback values match the original named fields, and the inherited configuration fixture is exact. This credit changes no source and adds no fresh native acceptance. Original-source: 2d573c3cb434410412ecfb6c86a8085c798d6d09
The adapter's processing-start hook and topic recovery can await before message dispatch. A handler installed on entry can disappear during that await, so the earlier presence check does not make either dispatch call safe. The turn-marker handoff also had no declared state on MessageEvent. Declare the marker as an event-local field that resets when dataclasses replace an event. Select and validate the current handler at each dispatch boundary. Background selection remains after the start hook and inside the pending reservation scope. When no handler is installed, background dispatch reports failure through the existing notification and cleanup paths. Inline dispatch errors preserve command restoration and clarify lifecycle handling. The scoped type comparison removes the three background findings and the sibling inline optional call, with 48 inherited findings unchanged and no introduced findings. The canonical lifecycle, marker and withdrawal run passes 137 cases. The separate command and clarify run passes 19 cases, including handler removal during real awaited topic recovery. Scoped Ruff reports no lint findings; unrelated inherited formatter changes remain.
The fake HTTP API discarded the client_session supplied by the Matrix adapter, so disconnect closed a mock while the real aiohttp session was left open. A reconnect test also replaced disconnect with a stub and left its new connection open after checking the guard. Strict resource warnings expose both leaks. Preserve the supplied session in the fake API. Use a wrapping disconnect spy in the reconnect test and disconnect the replacement client after the existing call assertion. The fake API and reconnect fixture predate this branch, so this focused fixture maintenance is separate from the corrections folded into their campaign contributors. Production code and all original assertions are unchanged. Apply the published fake-API session and reconnect cleanup to this consumer fixture. Preserve its additional Matrix types, membership, encryption and account assertions.
Cached auxiliary-client teardown is mixed into the large routing module. The worker-loop lifetime correction needs to await only the transports owned by a terminating worker, without expanding that routing module. Move the three transport-close functions into a lifecycle sibling and import the close function at its existing callers. Their full function ASTs and the remaining routing-module statements are unchanged.
Tool dispatch keeps one event loop per executor thread so cached async clients can reuse their transports between calls. The thread-local cache stored only the loop, so ending an executor left the loop and its selector sockets open. Auxiliary clients also remained cached after their owner thread exited. Temporary workers used inside an already running event loop also closed their loops before finishing generators and cached transports. Give each worker a Runner owner with a finalizer attached to its thread-local lifetime. Cancel remaining tasks and finish generators before closing cached clients on that loop. Wait for their close tasks, then let Runner shut down the executor and close the loop. Clients that belong to other live loops remain cached and open. The regression uses actual executor threads and SDK HTTP connections to a local server. It checks reuse between calls, generator cleanup before client closure, closed worker resources, and an unaffected main-thread client in both worker modes. Timeout cancellation still runs inside the worker loop, and executor shutdown remains nonblocking for the caller. The existing fake timeout executor now also closes the test coroutine which it deliberately never submits.
Matrix turns could read message history, but the agent could not inspect the current room state, joined members, power levels or pinned messages. Those details are available from the receiving bot and can affect what the agent should say or attempt. Extend the Matrix-only read tool with bounded inspection kinds. Fetch state and members through the owning adapter, apply its existing room and requester policy, and resolve pinned messages with the same decryption path as event reads. Report missing pin content per event so the other pins remain visible. Include the event type and creator override when presenting permission thresholds. A separate Matrix client verifies that the model receives the room details and pinned message from a disposable Synapse instance. Use the effective-state parent's client-events decoder and gateway scheduling path throughout inspection and its tests. Recheck admission after identity resolution, derive permission defaults from raw state, and preserve cancellation while refreshing pinned content. The feature and these interface migrations form one change so the initial commit uses the current prerequisite APIs.
iainlane
added a commit
to iainlane/hermes-agent
that referenced
this pull request
Oct 9, 2026
The complete feedback and edit source adds a second auxiliary-config argument to the shared native gateway fixture. Native collection fails before the unread controls can run. Keep the single additive argument that the controls fixture already provides. The incoming edit and lifecycle contracts also read optional mock-call records and assign dynamic fixture fields as statically declared methods. Check that the required calls occurred before reading their arguments, use the real profile-hook registry, and bind fixture replacements through monkeypatch. All original behaviour assertions remain in place. The complete changed production, unit and native harness type comparison has no new diagnostics after the defining-owner correspondence checks. The native collection syntax error is removed; collection-only execution is diagnostic evidence, not a test pass. Source and contributor reconciliation for original 726 (1655eb9d3aa2d693af7f8685185c80850408d82d). The original message above records the composed correction and its historical checks. This contribution preserves its complete credit after the source repairs were incorporated into their separate public owners. The five edit-follow-up fixtures use direct awaited-call guards and the real ProfileHookRegistries on NousResearch#126312. The lifecycle, deferred-input and feedback fixtures use the declared typed and monkeypatch adaptations on NousResearch#126283. These complete unit definitions are also present in its current composed tree. The native auxiliary configuration fixture has one argument and its consumer on NousResearch#126398. The NousResearch#126283 native harness has no auxiliary fixture or consumer, so it requires no corresponding argument. This is source and credit attribution across those owners. It adds no product or fixture patch, and records no fresh test, type, SDK or native execution. The complete original message and both literal contributor dates are preserved.
iainlane
added a commit
to iainlane/hermes-agent
that referenced
this pull request
Oct 9, 2026
The complete feedback and edit source adds a second auxiliary-config argument to the shared native gateway fixture. Native collection fails before the unread controls can run. Keep the single additive argument that the controls fixture already provides. The incoming edit and lifecycle contracts also read optional mock-call records and assign dynamic fixture fields as statically declared methods. Check that the required calls occurred before reading their arguments, use the real profile-hook registry, and bind fixture replacements through monkeypatch. All original behaviour assertions remain in place. The complete changed production, unit and native harness type comparison has no new diagnostics after the defining-owner correspondence checks. The native collection syntax error is removed; collection-only execution is diagnostic evidence, not a test pass. Source and contributor reconciliation for original 726 (1655eb9d3aa2d693af7f8685185c80850408d82d). The original message above records the composed correction and its historical checks. This contribution preserves its complete credit after the source repairs were incorporated into their separate public owners. The five edit-follow-up fixtures use direct awaited-call guards and the real ProfileHookRegistries on NousResearch#126312. The lifecycle, deferred-input and feedback fixtures use the declared typed and monkeypatch adaptations on NousResearch#126283. These complete unit definitions are also present in its current composed tree. The native auxiliary configuration fixture has one argument and its consumer on NousResearch#126398. The NousResearch#126283 native harness has no auxiliary fixture or consumer, so it requires no corresponding argument. This is source and credit attribution across those owners. It adds no product or fixture patch, and records no fresh test, type, SDK or native execution. The complete original message and both literal contributor dates are preserved.
iainlane
added a commit
to iainlane/hermes-agent
that referenced
this pull request
Oct 9, 2026
The shared receipt fixture inserts its configured read-receipt and reaction settings into the Matrix platform block. The unread test also specified those settings in raw YAML, so composition created duplicate keys. Configuration parsing failed and the gateway used an unconfigured provider before the observer could establish the test session. Select disabled automatic receipts and reactions through the shared MatrixFeedbackSettings fixture. Keep one platform configuration block and preserve the original scoped unread and explicit receipt assertions. The fresh plain and encrypted native run reproduced the duplicate-key configuration error on the composed parent. No production behaviour or native expectation changes in this correction. (cherry picked from commit 2d573c3cb434410412ecfb6c86a8085c798d6d09) Public source qualification: the complete unread feedback fixture effect is already present on NousResearch#126398. Positional feedback values match the original named fields, and the inherited configuration fixture is exact. This credit changes no source and adds no fresh native acceptance. Original-source: 2d573c3cb434410412ecfb6c86a8085c798d6d09
iainlane
force-pushed
the
feat/matrix-unread
branch
2 times, most recently
from
October 10, 2026 02:42
7b3a36c to
1e072b0
Compare
iainlane
added a commit
to iainlane/hermes-agent
that referenced
this pull request
Oct 10, 2026
The complete feedback and edit source adds a second auxiliary-config argument to the shared native gateway fixture. Native collection fails before the unread controls can run. Keep the single additive argument that the controls fixture already provides. The incoming edit and lifecycle contracts also read optional mock-call records and assign dynamic fixture fields as statically declared methods. Check that the required calls occurred before reading their arguments, use the real profile-hook registry, and bind fixture replacements through monkeypatch. All original behaviour assertions remain in place. The complete changed production, unit and native harness type comparison has no new diagnostics after the defining-owner correspondence checks. The native collection syntax error is removed; collection-only execution is diagnostic evidence, not a test pass. Source and contributor reconciliation for original 726 (1655eb9d3aa2d693af7f8685185c80850408d82d). The original message above records the composed correction and its historical checks. This contribution preserves its complete credit after the source repairs were incorporated into their separate public owners. The five edit-follow-up fixtures use direct awaited-call guards and the real ProfileHookRegistries on NousResearch#126312. The lifecycle, deferred-input and feedback fixtures use the declared typed and monkeypatch adaptations on NousResearch#126283. These complete unit definitions are also present in its current composed tree. The native auxiliary configuration fixture has one argument and its consumer on NousResearch#126398. The NousResearch#126283 native harness has no auxiliary fixture or consumer, so it requires no corresponding argument. This is source and credit attribution across those owners. It adds no product or fixture patch, and records no fresh test, type, SDK or native execution. The complete original message and both literal contributor dates are preserved.
iainlane
added a commit
to iainlane/hermes-agent
that referenced
this pull request
Oct 11, 2026
The complete feedback and edit source adds a second auxiliary-config argument to the shared native gateway fixture. Native collection fails before the unread controls can run. Keep the single additive argument that the controls fixture already provides. The incoming edit and lifecycle contracts also read optional mock-call records and assign dynamic fixture fields as statically declared methods. Check that the required calls occurred before reading their arguments, use the real profile-hook registry, and bind fixture replacements through monkeypatch. All original behaviour assertions remain in place. The complete changed production, unit and native harness type comparison has no new diagnostics after the defining-owner correspondence checks. The native collection syntax error is removed; collection-only execution is diagnostic evidence, not a test pass. Source and contributor reconciliation for original 726 (1655eb9d3aa2d693af7f8685185c80850408d82d). The original message above records the composed correction and its historical checks. This contribution preserves its complete credit after the source repairs were incorporated into their separate public owners. The five edit-follow-up fixtures use direct awaited-call guards and the real ProfileHookRegistries on NousResearch#126312. The lifecycle, deferred-input and feedback fixtures use the declared typed and monkeypatch adaptations on NousResearch#126283. These complete unit definitions are also present in its current composed tree. The native auxiliary configuration fixture has one argument and its consumer on NousResearch#126398. The NousResearch#126283 native harness has no auxiliary fixture or consumer, so it requires no corresponding argument. This is source and credit attribution across those owners. It adds no product or fixture patch, and records no fresh test, type, SDK or native execution. The complete original message and both literal contributor dates are preserved.
iainlane
added a commit
to iainlane/hermes-agent
that referenced
this pull request
Oct 11, 2026
The complete feedback and edit source adds a second auxiliary-config argument to the shared native gateway fixture. Native collection fails before the unread controls can run. Keep the single additive argument that the controls fixture already provides. The incoming edit and lifecycle contracts also read optional mock-call records and assign dynamic fixture fields as statically declared methods. Check that the required calls occurred before reading their arguments, use the real profile-hook registry, and bind fixture replacements through monkeypatch. All original behaviour assertions remain in place. The complete changed production, unit and native harness type comparison has no new diagnostics after the defining-owner correspondence checks. The native collection syntax error is removed; collection-only execution is diagnostic evidence, not a test pass. Source and contributor reconciliation for original 726 (1655eb9d3aa2d693af7f8685185c80850408d82d). The original message above records the composed correction and its historical checks. This contribution preserves its complete credit after the source repairs were incorporated into their separate public owners. The five edit-follow-up fixtures use direct awaited-call guards and the real ProfileHookRegistries on NousResearch#126312. The lifecycle, deferred-input and feedback fixtures use the declared typed and monkeypatch adaptations on NousResearch#126283. These complete unit definitions are also present in its current composed tree. The native auxiliary configuration fixture has one argument and its consumer on NousResearch#126398. The NousResearch#126283 native harness has no auxiliary fixture or consumer, so it requires no corresponding argument. This is source and credit attribution across those owners. It adds no product or fixture patch, and records no fresh test, type, SDK or native execution. The complete original message and both literal contributor dates are preserved.
alt-glitch
added
area/install-update
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
7 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack
All campaign PRs in this section target
main. This list records logical merge order. Nesting means the parent PR must merge before its child. Every prerequisite for a dependent must merge first. Only adjacent layers are shown; each linked PR description lists its other dependencies.Matrix sessions can inspect history but cannot read the bot account's observed notification counts or deliberately acknowledge a selected event and timeline. Add
matrix_unreadandmatrix_mark_readthrough the Matrix session toolset, with ahermes toolstoggle. Counts describe the bot account's push rules and remain unavailable before server observation.Explicit acknowledgements require the exact event, scope and public/private visibility. Each write checks current requester membership, room policy, requester/target-sender authorization and the original live owner. Tool interruption cancels and drains the owning task while its gateway loop is running. If the loop stops, the tool returns without claiming a completed drain and prevents a cancelled preflight from issuing a new write after recovery. Confirmed receipts and uncertain receipt/account-data writes remain separate in the result. Only room scope resets the bot account's marked-unread flag; the fully-read marker remains unchanged.
Current prerequisite composition and review
This successor retains the actual independently published receipt supplier
dc2811135e5f3a3bdc670ad69b8973c412b33806. Review the own unread layer above that exact supplier. The complete 28 current own contributions remain in their original order, with literal author and committer metadata, dates, timezones and full messages. All thirteen meaningful empty own records remain. Every complete cumulative snapshot joins automatically onto the supplier, without another production implementation or conflict resolution.The compatibility target is captured upstream
3637c512fc3211bdccb61b7140d09ac1018508ba. Local joins with that revision and the original frozen target are clean. The current receipt supplier already includes the generic worker-loop owner and finaliser. Its four complete defining scopes are AST-exact to the independently published core correction. Both complete pin and read files pass with this production, so the earlier explicit worker fixture ports are unnecessary here. The complete request-loop result mechanism also matches the separately published #130706 source, with formatting as the sole difference. Its logical prerequisite remains.Current validation and limits
All 199 cases across seven complete SDK files pass with zero resource warnings. The 31-case pin and 17-case read files passed first on the final unchanged source; the disjoint five-file unread, sync, warning, invite and runtime scope then passed 151 cases on that same source. Each run used one worker, zero retries and all four strict warning categories. Both runs captured all 8,553 tracked Python and test-runner inputs before and after execution. The canonical runner was restored exactly.
A fresh PM-managed image of this exact source passes both original plain and encrypted native unread cases. Separate Matrix clients observe notification counts, the selected room and thread receipt scopes, public/private acknowledgements and account-data updates. The complete native child is clean through process exit, with strict warnings and no retries. The image, package and lock inputs and actual tracked native runtime bindings are recorded separately. The original native assertions and parameters remain unchanged.
The complete 17-path type assessment reports 51 inherited findings, compared with 51 on the original current branch and 59 on the current supplier. Every complete lexical, statement-AST and relative statement-span occurrence is within the original/provider maximum, with zero introduced findings. Ruff passes. Thirteen formatter paths remain inherited; this successor adds none. Cumulative whitespace checks pass.
These are exact local SDK and live Matrix receipts. Earlier broad and hosted results below remain historical, and this delivery does not claim fresh hosted CI acceptance.
Previous recorded delivery
The following counts, commit references and runtime evidence describe
those earlier recorded sources, not the current composition.
Previous recorded prerequisite composition
Merge receipt lifecycle, Matrix administration and the shared runtime lifecycle repair first. This branch composes the complete published receipt parent with the corrected inspection, pins and administration contributions. The twelve prerequisite contributions retain their original authors, dates and complete messages; runtime-equivalent contributions retain explicit credited commits. The fourteen original unread contributions are preserved above a separate mechanical extraction of the existing sync topic.
This PR targets upstream
main, so the cumulative diff includes its prerequisites. The own review comparison is the unread composition above published receipt parent6bbef00ea3501a0ba7e56de7bf24ad43648f523e. The rebase uses fixed upstream9a71d5a8b59b37d1795ab7065fbbe3b647ba5249because GitHub confirmed an upstream conflict. A later change tomainalone does not trigger another rebase.Current live authorization, transport identity, profile and physical-home checks from the receipt parent remain intact. The shared stopped-loop runtime and its complete separate-loop test module are byte-identical to the corrected published #130706 contribution. The original stopped-loop source and the published unread copy have identical author, date, full message and patch. Their explicit source aliases use the unread contribution as one canonical credit; the redundant final empty copy is omitted in this clean history. All runtime and native inputs remain exact; the inherited receipt-parent unit fixture is the only tree difference.
Previous recorded validation
The actual Mautrix initial-sync contract reproduced missing unread filters for both a fresh sync and a saved cursor. Both rows failed before the correction and pass after the original unread feature supplies the filter to every sync branch. The complete unread unit module passes 64 cases. Existing decorators, parameter tables and whole assertions remain intact; three callback fixtures now use actual
AsyncMockinstances, as required by the preserved contributor typing correction.The complete affected type comparison checks 146 Python paths, including every original affected unit module, with the actual SDK sources. The corrected composition has 1,336 diagnostics against 1,344 on exact published receipt parent
6bbef00ea3501a0ba7e56de7bf24ad43648f523e, with zero introduced diagnostic buckets. The comparison preserves complete rules, messages, qualified scopes and multiplicity. Only the four independently proved sync method moves receive explicit scope mappings. The previously corrected extraction annotations remain unchanged.Signed candidate
4c04f0ae9aba6ef067a2d17537787636bf499e73passes all 910 tests across 16 files in 74.8 seconds, with four workers, strict warning errors and zero failures, skips or retries. The original 15-file scope passes 892 cases; the inherited receipt-parent unit fixture adds 18. Committed-delta whitespace and Ruff checks pass for all 18 selected Python files. All 407 cumulative contributions are signed and linear. All 180 sources from the previously published unread branch retain their exact raw author, date, timezone and complete message prefixes. The clean history contains 27 own contributions and omits only the redundant final empty runtime credit. Original c3/DDDE runtime aliases resolve to the single preserved unread contribution; every runtime and native byte remains unchanged.The complete admin test module is byte-identical to the published receipt parent and the protected accepted unread tree. Its complete decorated catalogue function equals the corrected published #126437 owner. The fixture explicitly disables unrelated optional toolsets and preserves the original expected
matrix_read, all parameter rows and the A/B/A profile contract. The correction already exists in the published parent, so the unread replay adds no second fixture correction.Both original native unread cases passed on protected head
a99a6567e7a8d46e566f1cd875076a7557d00489in 52.1 seconds, with one worker, no retries and the unchanged 300-second budget. Those results transfer to this candidate through independently verified identical inputs: all 6,746 image source files, 2,441 required runtime files and 39 mounted native harness files match the accepted image. The original complete native module and every production blob and mode are unchanged. The sole tree difference is a Docker-excluded unit fixture inherited from the receipt parent. Git's existing PowerShell line-ending conversion is explicit in the input proof. No new image build or native execution is claimed for this candidate. The accepted image issha256:d7c0a1107d3be7628b3751808066766d70ff95729d280ce853fbb0de23539cbb; its SDK versions are Mautrix 0.21.1, matrix-nio 0.26.0 and aiohttp 3.14.3.The earlier protected composition passed 892 unit cases after the public admin correction disabled unrelated toolsets. The separate original catalogue owner is now published at
b9fa14ac69185bab4e0bcab904c24c22381ff0a7; the receipt parent contains that exact corrected catalogue contract. These are public source-owner corrections. The first setup attempt for this clean candidate was blocked before tests started; the complete passing runner receipt above describes the frozen candidate.Historical validation
The following receipts describe the earlier published heads and remain historical. They do not establish acceptance of the current rebased composition.
The cancellation and current-membership contracts reproduce actual interruption and policy defects before repair. Session, home and profile replacement inside fake callbacks are controls for the shared runtime; those controls do not establish a separate production race.
Before the read-only follow-up, signed head
36fad5c7d960a659b52bbe7088a9c63019160fb9passed 1,394 tests across 34 files, with four workers and zero retries or skips. A fresh image of the same head passes both plain and encrypted native cases in 42.7 seconds. Independent Matrix clients observe the selected room and thread acknowledgement scope, counts and account-data changes. The same exact image passes all 13 existing receipt visibility and processing-boundary cases in 203.0 seconds, including manual plaintext approvals and the specialized model responders. Ruff, the affected production-module type checks and whitespace checks pass.The composed receipt fixtures use observed joined members for DM controls and preserve queued fake-model tool calls. Their earlier failures are recorded in the review report and pass after these fixture changes. The encrypted local history index remains a separate follow-up.
Read-only unread requests now observe agent interruption and the shared thirty-second deadline. Running-loop cancellation stops the owning task before the worker returns. Reads never signal a write and therefore return an ordinary interruption or timeout error. An event-controlled registry contract reproduces both blocked-read failures on the previous head and preserves the accepted-read control.
Signed head
31ba23a5831034576d0e7181202efac1512789b0passes the three-row read contract and all 414 tests across nine distinct adjacent files, with four workers and zero retries or skips. Whole-tree Ruff, the edited tool's type check with the actual Mautrix source and whitespace checks pass. A fresh image of this exact head passes both plain and encrypted native unread cases in 43.1 seconds. The 13 receipt native cases above are recorded evidence from the previous head, which this read-only change does not claim to rerun.The final runtime composition passes 326 tests across five existing files, with four workers and no retries or skips. Its separate real-loop contract verifies stopped admission, preflight loss, uncertain dispatched writes, confirmed results and recovery without a late write. A fresh image of this exact composed head passes both plain and encrypted native unread cases, 2/2 in 44.2 seconds, with no retries or skips. Earlier native receipts above remain historical.
The fresh final image digest is
sha256:6e9fa0c6623b7b9d1cce7e5b233209932fb74a2a4c110fa7801139214f10dead. Hosted CI remains subject to upstream approval.Original fixture credit
The complete original shared-fixture correction is recorded in commit
8305c5aebde8e34bab7af12d56943eef57d3a4f2, with both original contributor dates and the full original message. Its parent is4c04f0ae9aba6ef067a2d17537787636bf499e73, and both commits have source tree552b35f28c62dc090c7a3b859a8a90b307673e58. The branch now includes the unchanged 407 signed contributions and this signed empty credit contribution.The complete fixture effect was already present through the shared configuration and scoped feedback settings. This credit changes no source and adds no fresh runtime or native acceptance. The validation results and native transfer qualifications above remain historical evidence for the unchanged source. The reported conflict with current main remains; this credit does not establish main compatibility.