fix(desktop): treat interrupted completions as metadata - #63292
Open
yingliang-zhang wants to merge 101 commits into
Open
fix(desktop): treat interrupted completions as metadata#63292yingliang-zhang wants to merge 101 commits into
yingliang-zhang wants to merge 101 commits into
Conversation
alt-glitch
added
type/bug
yingliang-zhang
force-pushed
the
fix/desktop-interrupt-completion-metadata
branch
from
July 15, 2026 00:40
26b1862 to
c1d79bf
Compare
Contributor
Author
|
Rebased onto current main lineage and resolved the interruption-path conflicts. Follow-up |
teknium1
reviewed
Jul 15, 2026
teknium1
left a comment
teknium1
left a comment
Contributor
There was a problem hiding this comment.
Thanks for tracing the cancellation path across the core loop and Desktop. The premise is confirmed on current main: agent/conversation_loop.py:2352 returns interruption prose when no text streamed, and Desktop currently treats every message.complete as successful at apps/desktop/src/app/session/hooks/use-message-stream/gateway-event.ts:384-403.
Problems
agent/agent_runtime_helpers.py:2670insertsOperation interrupted.for everytool → useradjacency, with no interruption provenance. Current main explicitly treats anassistant(tool_calls) → tool → usersequence as a valid normal redirect (agent/agent_runtime_helpers.py:394-399;tests/run_agent/test_message_sequence_repair.py:108-129). This would inject false cancellation context into the next model request for that valid path.
Suggested changes
- Scope API-only closure to an explicitly cancellation-originated tail, and add a negative sanitizer test for the normal redirect sequence.
Automated hermes-sweeper review.
teknium1
added
the
sweeper:blast-moderate
yingliang-zhang
force-pushed
the
fix/desktop-interrupt-completion-metadata
branch
7 times, most recently
from
July 24, 2026 01:50
2e1003e to
ae5dd4f
Compare
yingliang-zhang
force-pushed
the
fix/desktop-interrupt-completion-metadata
branch
4 times, most recently
from
July 30, 2026 07:10
24dbf94 to
f4fd162
Compare
yingliang-zhang
force-pushed
the
fix/desktop-interrupt-completion-metadata
branch
6 times, most recently
from
August 10, 2026 04:40
9ef5f2d to
27463c4
Compare
yingliang-zhang
force-pushed
the
fix/desktop-interrupt-completion-metadata
branch
from
August 12, 2026 03:58
27463c4 to
2078a62
Compare
yingliang-zhang
force-pushed
the
fix/desktop-interrupt-completion-metadata
branch
from
August 16, 2026 01:55
2078a62 to
b1ba8ba
Compare
…side import block)
…ith hybrid-routing-v1 callers)
The union merge dropped the single call-site wiring line for the already-configured option (patch 360's only material gap); compress options read the value but never passed it.
…hropic wire MiniMax-M3 rides MiniMax's server-side automatic prefix cache; explicit cache_control is documented for the M2 family only, so markers on M3 are dead weight (cache_creation always 0, never billable). Restore the exclusion (provider-id or host match) checked BEFORE the native-Anthropic return, so provider="anthropic" pointed at a MiniMax /anthropic proxy is also caught. The later M2 opt-in reuses the hoisted route predicate. Fixes tests/run_agent/test_anthropic_prompt_cache_policy.py ::TestMiniMaxAnthropicWire (4 candidate-only failures).
switch_model snapshots _reasoning_echo_flag into _primary_runtime and a fallback entry overwrites the live flag, but restore_primary_runtime never reverted it, so the fallback's opt-in leaked into the restored primary. Restore the flag from the snapshot (default False for snapshots taken before the field existed). Also tolerate a missing context_compressor when rebinding context-engine state; gateway/CLI paths may construct it lazily. Fixes tests/agent/test_message_sanitization_policy.py ::TestPerProviderReasoningEcho::test_restore_primary_reverts_flag.
With the inline dispatch branch amputated, setup_mcp escaped to handle_function_call while the post-hook ownership predicate also dropped its membership, breaking the emit-once contract on both executor paths. Restore the invoke_tool branch and the AGENT_RUNTIME_POST_HOOK_TOOL_NAMES entry. Fixes tests/run_agent/test_run_agent.py::TestAgentRuntimePostHookOwnershipSync (2 candidate-only failures).
…ach, log close reason Three restorations for the slash-worker lifecycle: - Keep references to the drain threads (named slash-drain-stdout / slash-drain-stderr for leaked-thread attribution) and join them with a bounded 2s timeout in close(); previously each closed session leaked two daemon threads holding worker references (NousResearch#53303). - _close_sessions_for_transport now closes the session's slash worker immediately when detaching a non-close_on_disconnect session instead of letting it linger until the orphan/TTL reaper; tolerates worker.close() failure and a missing history_lock. The worker recreates lazily on next use. - _close_session_by_id logs "session closed: <sid> (end_reason=...)" at INFO; the teardown path was silent, making reap behavior undiagnosable. Fixes tests/tui_gateway/test_slash_worker_drain.py (2) and tests/tui_gateway/test_slash_worker_detach.py (3).
switch_model refreshed only model metadata on the compressor, so a move onto a custom codex_responses route never applied the per-route Codex threshold autoraise (0.85) and moving back never restored the preserved config baseline. Resolve the new route's threshold the same way agent_init does at startup — _compression_threshold_for_model gated by _codex_gpt55_autoraise, combined with the raise-only resolver semantics — and pass it through update_model(default_threshold_percent=...). The agent-side _compression_global_threshold attribute is optional; the compressor's immutable config baseline substitutes when the agent wasn't fully constructed. Also tolerate partially-initialised ContextCompressor instances in _effective_max_tail_message_floor (fixtures built via __new__ set only the attributes under test). Fixes tests/run_agent/test_switch_model_context.py ::test_switch_model_custom_codex_threshold_uses_resolved_window, tests/agent/test_context_compressor_cross_session_guard.py (3) and tests/agent/test_compressor_image_tokens.py ::TestTokenBudgetWithImages::test_image_heavy_turns_count_toward_budget.
…d tool tails on the wire Restoration of three lost behaviors around durable prompt/transcript reuse: 1. _restore_or_build_system_prompt reused the persisted prompt verbatim but never seeded the frozen plugin-section snapshot, so the first invalidate/rebuild in the resumed process re-evaluated section callbacks and rewrote prompt bytes, breaking the byte-identical-resume contract and the cache prefix. Seed the snapshot from the persisted prompt; section callbacks evaluate only on new sessions. 2. sanitize_api_messages lost the NousResearch#48879/NousResearch#63292 closure pass: a durable _interrupted_tool_tail tool row followed by a user redirect reached the wire as a bare tool -> user alternation, which strict providers reject or hallucinate against. Close the sequence on the per-call API copy with the interim "Operation interrupted." assistant marker, and strip the internal provenance key from the wire copy. 3. Re-align the NousResearch#68454 rotation-flush control test with this stack's schema: the partial UNIQUE index idx_messages_active_dedupe(session_id, role, content, timestamp) WHERE active=1 makes a bare flush of unstamped cold-resume rows a DB-level no-op, so the control now forbids duplicates instead of asserting the upstream double-write. Production is unchanged; the boundary/no-op and tail-only sibling tests are untouched. Fixes tests/agent/test_plugin_prompt_sections.py ::test_fresh_process_resume_restores_identical_full_prompt_without_callback, tests/agent/test_session_rotation_flush_cold_resume_68454.py ::test_rotation_flush_without_history_boundary_duplicates, tests/agent/test_interrupt_tool_tail_api_sanitization.py ::test_user_after_interrupted_tool_tail_is_closed_only_in_api_copy, and tests/agent/test_turn_finalizer_interrupt_alternation.py ::test_interrupt_after_tool_keeps_transcript_clean_and_closes_api_copy.
…e-boundary submits The 63298 queue-boundary contract emits `queued: true` on re-submission and rebinds the queued entry's session identity to the recovered runtime before the first visible submit; update the two stale assertions (3 session.resume calls, recovered id on both submits).
The runtime updater pipeline's hindsight_post_sync verifier ran fully on the promoted branch state but rejected the 0.6.1 pin; bumping to the same 0.8.4 the server env already carries.
…lersistence) uv sync --locked prunes any package without a lock entry; main-venv embed package must be part of the resolving graph to survive dependency_sync before hindsight_post_sync.
…ddedClient) Restores plugins/memory/hindsight/embedded_runtime.py from the P0 hindsight-server isolation lineage (1660e5528). Required by the updater's verify_hindsight_runtime runtime-adapter contract: imports cleanly without the full hindsight namespace and exposes _ensure_started/close/__getattr__.
…ons call (rebase fix)
…ron tests to upstream tick architecture
…_cron_session_db The done-callback _close_late_session_db_result was registered before future.result() was called, so when SessionDB() completed on time (the normal case) the callback fired immediately in the calling thread and closed the connection — returning a SessionDB with _conn=None. This caused the stale-session reaper to silently return 0 (list_open_cron_sessions hit _conn.execute on None) and run_job to fail when titling sessions. Matching the upstream pattern: register the callback only inside the except TimeoutError block so it fires exclusively for a late result. Fixes 5 TestReaperFailClosed tests and test_monitor_kind failures.
The Hindsight plugin's _get_client() still used the legacy 'from hindsight import HindsightEmbedded' path requiring hindsight-all (torch/transformers/sentence-transformers ~2-3GB) in the agent venv. The split-runtime design (DedicatedEmbeddedClient in embedded_runtime.py) was committed (ec7ba1a) but the __init__.py wiring was lost in a subsequent rebase, leaving Hindsight silently offline since Aug 14. Changes: - _check_local_runtime: probe hindsight_client + hindsight_embed + embedded_runtime (lightweight, <1s) instead of hindsight + sentence_transformers (monolith, requires 134-package install) - _get_client: use DedicatedEmbeddedClient from .embedded_runtime instead of HindsightEmbedded; drop __del__ monkey-patch; add server_executable kwarg - _probe_url: read inner Hindsight._base_url (private attr) instead of .url (would trigger __getattr__ daemon start) - _local_runtime_hint: broadened trigger; new guidance points to hermes memory setup + HINDSIGHT_EMBED_API_EXECUTABLE, not hindsight-all - Version constants: restore _MIN=0.8.4, _MAX=0.10, _CLIENT_REQUIREMENT, _EMBED_REQUIREMENT (was regressed to 0.6.1) - Setup wizard: install [client, embed] not hindsight-all for local_embedded - embedded_runtime.py: add seal cache (keyed by candidate+fingerprint, 45000x speedup) + conditional LLM config (only set HINDSIGHT_API_LLM_* when truthy, preventing empty-string shadowing of hermes.env keys) - plugin.yaml: declare hindsight-embed>=0.8.4,<0.10 - Remove hindsight-all from memory_setup.py, web_server.py, update_cmd.py - Update tests to mock DedicatedEmbeddedClient; port test_hindsight_embedded_runtime.py (621 LOC) from reference; add seal-cache regression test - Update website docs (EN + zh-Hans) to split-runtime terminology Verified: hermes memory status → available; _check_local_runtime → (True, None) in 0.49s; seal cache 5.45s→0.0001s; live recall 26 facts in 3.68s; 154 passed / 3 pre-existing failures (confirmed on pristine HEAD).
Resolved conflicts: - cron/scheduler.py: kept both hashlib (ours) and errno (upstream) imports - cron/scheduler_provider.py: adopted upstream EMFILE self-heal backoff - pyproject.toml: kept our hindsight 0.8.4 + upstream mcp 2.0.0/httpx2 upgrade Auto-merged 23 overlapping files incl: agent/conversation_loop.py, gateway/run.py, hermes_state.py, run_agent.py, tools/*, tui_gateway/server.py, apps/desktop/src/app/session/hooks/*. Upstream highlights: Bot Mode builtin plugin (6935 lines), HUD surface, multi-gateway connections registry, session list density modes, transcript tail-page hydration, per-turn duration badges, find-bar positioning fixes, titlebar clusters, sidebar filter/inbox/density, running≠busy, composer settle-hold bound, session tile reclaim self-heal.
yingliang-zhang
force-pushed
the
fix/desktop-interrupt-completion-metadata
branch
2 times, most recently
from
August 18, 2026 01:05
ef7fe8e to
f2c8b04
Compare
Resolved 4 conflicts in tui_gateway/ — preserved our submitted_at/message_id params alongside upstream's new display_kind param: - server.py: _run_prompt_submit + _compute_host_turn_frame signatures - compute_host.py: frame kwargs forwarding - methods_prompt.py: _run_prompt_submit call in run_after_agent_ready - test_tui_gateway_server.py: _run_inline mock accepts display_kind
…ration Address review feedback on NousResearch#62492 from @jrleal10: 1. _migrate_profile_config: print an actionable per-profile warning when missing required settings prevent silent auto-migration (previously silently skipped, leaving stale profiles undiagnosed). 2. cmd_update all-profiles loop: replace logger.debug (invisible to users) with a visible stderr warning when per-profile migration fails, including the profile name and remediation command. 3. Add behavioral test coverage for both entry points: - TestMigrateProfileConfig: helper unit tests (4 tests) - TestUpdateAllProfilesMigration: update-loop coverage (2 tests) - TestDashboardStartupMigration: serve/dashboard startup coverage (3 tests) Tests cover: version-bump-only silent path, missing-settings warning, exception surfacing, up-to-date skip, multi-profile iteration, and failure visibility.
Tri-model review (3/3 ACCEPT) flagged the profiles_env fixture as dead code — no test uses it (all 12 tests use mock.Mock() profiles). Removed fixture plus unused os/pathlib/yaml imports.
…ount publication The background prefetch worker in queue_prefetch inlined its own recall/reflect calls instead of honoring the shared _do_recall contract, so it bypassed two behaviors our tests encode: (1) the query was never clipped to recall_max_input_chars before reaching arecall/areflect, and (2) the discrete memory count was never stored in _prefetch_count, leaving prefetch() to publish count=0 in the deterministic recall indicator (recall_status) regardless of what was injected, and breaking the stale-count lifecycle. Fix: extract a shared _truncate_query() helper used by both the sync and background recall paths, and have the worker capture len(results) (0 for reflect synthesis) and store it alongside _prefetch_result under the same lock; the session-rotation reset now clears the count with the buffer so the pair stays atomic.
# Conflicts: # contributors/emails/agent@agents-Mac-mini.local
yingliang-zhang
force-pushed
the
fix/desktop-interrupt-completion-metadata
branch
from
August 19, 2026 03:16
47862b2 to
441dccd
Compare
# Conflicts: # apps/desktop/src/app/session/hooks/use-prompt-actions/index.test.tsx
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
5 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.
Summary
interruptedstatus with an empty final response instead of assistant prosetool -> usertails only in the per-request API copy so strict providers still receive valid orderingRoot cause
The core loop returned local cancellation status strings such as
Operation interrupted: waiting for model response (...)asfinal_response, and the turn finalizer could persist a syntheticOperation interrupted.assistant row after a tool result. Desktop then handled everymessage.completeas a successful completion, so cancellation metadata appeared as a reply and triggered completion UX.The fix makes structured
interruptedstate authoritative. Empty cancellations stay metadata-only; actual streamed partial text is still returned and persisted. Provider role repair now happens on the API copy rather than rewriting the durable transcript.Verification
pytest -q tests/run_agent/test_message_sequence_repair.py tests/agent/test_turn_finalizer*.py tests/agent/test_interrupt_tool_tail_api_sanitization.py tests/run_agent/test_run_agent.py -k 'interrupt or tool_tail or sanitize'— 37 passednpx vitest run src/app/session/hooks/use-message-stream --environment jsdom— 12 passednpm run typecheckinapps/desktop— passeduv run --frozen --extra dev --extra acp pytest -q tests/acp/test_server.py -k interrupt— 1 passednpx vitest run src/__tests__/createGatewayEventHandler.test.ts -t 'interrupt'inui-tui— 2 passedgit diff --check— passedRelated work
This is complementary to #55317: that PR surfaces non-interrupted tool-result-only endings, while this PR keeps interrupted cancellation paths metadata-only.