hermes-agent

Files

rodricksz4h5 7299015092 fix(agent): track image rejections per model across a fallback chain

The first head stored a single rejecting (provider, model) and kept the
turn-global `_vision_supported` as the recovery guard. In a fallback
chain that fails: model A rejects images and retries text-only, a later
error activates model B, the restart rebuilds api_messages from history
so B receives the images, and when B rejects them too the branch is
skipped because `_vision_supported` is already False — the request
falls through to generic error handling. Recording B also overwrote A,
so A was no longer treated as text-only on later turns.

`_image_rejecting_models` is now a set of every rejecting model, and it
is also the guard: each model's first rejection runs the recovery and a
repeat rejection from the same model still falls through, so the retry
cannot loop. image_model_key() names the key in one place.

Adds a test for the two-model sequence (fails on the previous head) and
one pinning that a repeat rejection from the same model does not retry.

Thanks to @ehz0ah for the review.

(cherry picked from commit 225fd76ccf90cd3ec509a9327f725ac02d996f24)

2026-09-21 21:43:51 +05:30

lsp

fix(lsp): one stalled request no longer silences a workspace for good — retry window, cold-root warm-up budget, per-root exclusion

2026-09-20 18:54:24 -07:00

monitoring

…

pet

fix(security): route remote-party-supplied URL fetches through the SSRF guard

2026-09-18 10:22:14 -07:00

proxy_sources

…

secret_sources

fix(env_loader): restore only values a source is authoritative for

2026-09-19 16:20:29 +05:30

transports

feat(providers): external-process provider support for standalone model-provider plugins (from #105863 )

2026-09-20 14:29:39 -07:00

vault_backends

…

verify

…

__init__.py

…

account_usage.py

fix(agent): release the portal account fetch caller at the wall-clock bound

2026-09-20 00:35:04 -07:00

acp_openai_bridge.py

…

activity_tracking.py

…

agent_init.py

fix(agent): track image rejections per model across a fallback chain

2026-09-21 21:43:51 +05:30

agent_runtime_helpers.py

fix(context): re-derive model.context_length on model/provider change

2026-09-20 16:31:10 -07:00

AGENTS.md

docs(agent): list turn_recovery_autorecover in the turn-phase sibling map

2026-09-19 12:38:55 -07:00

anthropic_adapter.py

fix(anthropic): key_cmd Claude Code OAuth identity survives on custom api.anthropic.com routes (main + aux)

2026-09-20 12:49:40 -07:00

anthropic_credentials.py

fix(anthropic): key_cmd Claude Code OAuth identity survives on custom api.anthropic.com routes (main + aux)

2026-09-20 12:49:40 -07:00

anthropic_endpoints.py

…

anthropic_message_convert.py

fix(vision): accept the image/jpg alias and give the Anthropic converter the same unsupported-subtype downgrade

2026-09-19 09:56:12 -07:00

api_error_summary.py

…

api_request_hooks.py

…

async_utils.py

…

aux_accounting.py

…

auxiliary_async_rebuild.py

fix(aux): keep key_cmd bearer and extra_headers on async aux calls to named custom providers

2026-09-17 08:53:15 -07:00

auxiliary_client.py

fix(auth): prevent repeated copilot raw token exchange warnings (#114740 )

2026-09-20 13:22:30 -07:00

auxiliary_fallback_recovery.py

fix(aux): fallback ladder route carries the client base_url; test builds the route by name

2026-09-17 08:45:46 -07:00

auxiliary_health.py

fix(aux): quarantine a transient fallback-candidate failure for seconds, not the payment hold

2026-09-17 09:07:51 -07:00

auxiliary_reasoning_floor.py

fix: step reasoning up to the floor when a route refuses to disable it

2026-09-19 02:37:36 -07:00

auxiliary_structured_output.py

fix(aux): scope the response_format rejection memo to the model and to capability errors

2026-09-17 09:11:38 -07:00

auxiliary_unavailable.py

fix: exhausted OAuth credential pool reports the cooldown and reset time, not "no credentials"

2026-09-19 10:25:30 -07:00

auxiliary_wire.py

fix: drop replayed reasoning_details on every chat-completions route that does not read it

2026-09-19 09:28:45 -07:00

azure_identity_adapter.py

fix(agent): refuse ambient credential chains for multiplex profiles

2026-09-20 12:09:47 -07:00

backend_identity.py

…

background_review.py

fix(background_review): say so when the configured review route falls back to the main model

2026-09-19 12:30:32 -07:00

battery.py

…

bedrock_adapter.py

fix(agent): refuse ambient credential chains for multiplex profiles

2026-09-20 12:09:47 -07:00

billing_links.py

…

billing_usage.py

fix(billing): the /billing portal fetch releases the caller at its wall-clock bound too

2026-09-20 00:35:04 -07:00

billing_view.py

…

bounded_response.py

…

browser_provider.py

…

browser_registry.py

…

chat_completion_helpers_relay.py

feat(providers): external-process provider support for standalone model-provider plugins (from #105863 )

2026-09-20 14:29:39 -07:00

chat_completion_helpers.py

fix(agent): pin the reset_at reach through try_activate_fallback and add opt-in fallback.min_switch_reset_seconds

2026-09-20 17:00:43 -07:00

chat_completion_nonstream.py

fix(wait-notice): derive the near-deadline flag once; retire the stale 'provider may be slow' docstrings

2026-09-19 09:53:48 -07:00

chat_completion_stream_monitor.py

fix(wait-notice): derive the near-deadline flag once; retire the stale 'provider may be slow' docstrings

2026-09-19 09:53:48 -07:00

chat_completion_wait_notice.py

fix(wait-notice): derive the near-deadline flag once; retire the stale 'provider may be slow' docstrings

2026-09-19 09:53:48 -07:00

client_lifecycle.py

feat(providers): external-process provider support for standalone model-provider plugins (from #105863 )

2026-09-20 14:29:39 -07:00

codex_headers.py

fix(agent): extract residency claims from Codex OAuth JWT for workspace auth

2026-09-19 09:28:13 -07:00

codex_responses_adapter.py

chore: merge origin/main (resolve tests/agent/transports/test_codex_transport.py)

2026-09-19 10:52:11 -07:00

codex_runtime_history_seed.py

fix(codex): a codex app-server thread started from scratch is seeded with the session's prior turns

2026-09-19 20:44:24 -07:00

codex_runtime.py

fix(codex): a codex app-server thread started from scratch is seeded with the session's prior turns

2026-09-19 20:44:24 -07:00

coding_context.py

…

command_token_source.py

…

compaction_display.py

…

compression_facade.py

fix(compression): track commit fences as a registration stack

2026-09-20 15:27:11 -07:00

context_breakdown.py

…

context_compressor_summary.py

fix(compression): fence the durable cooldown rollback per compressor; one stale-attempt helper

2026-09-20 15:50:49 -07:00

context_compressor.py

fix(compression): cap the protected tail at 20% of the context window

2026-09-21 01:04:16 -07:00

context_engine.py

…

context_file_sources.py

fix(prompt_builder): load the user's own SOUL.md on a scanner hit instead of blocking it

2026-09-16 17:17:47 -07:00

context_pin.py

feat: label a model.context_length pin and warn once when it disagrees with the provider

2026-09-19 10:28:34 -07:00

context_references.py

fix(desktop): large text pastes attach from HERMES_HOME/composer-pastes when the chat cwd is elsewhere

2026-09-20 19:45:30 -07:00

conversation_compression_manual.py

…

conversation_compression.py

fix: return a raced-successful compression from _await_worker_within_budget

2026-09-21 21:34:44 +05:30

conversation_loop.py

fix: vision_analyze does not re-embed an image already attached natively to the current turn

2026-09-19 09:52:18 -07:00

copilot_acp_client.py

refactor: resolve runtime temp paths via tempfile/TMPDIR instead of literal /tmp

2026-09-19 10:44:26 -07:00

credential_persistence.py

…

credential_pool_admin.py

fix(auth): hermes auth reset survives a live session's next pool flush

2026-09-19 09:40:56 -07:00

credential_pool_model_cooldowns.py

fix: pin entitlement rotation through recover_with_credential_pool; bench until reset

2026-09-19 09:37:23 -07:00

credential_pool_plugin.py

feat: model-provider plugins classify their own API errors via ProviderProfile.classify_api_error

2026-09-19 21:24:05 -07:00

credential_pool.py

fix(auth): skip the Copilot token exchange while copilot is only an ambient gh-CLI credential

2026-09-20 13:22:30 -07:00

credential_sources.py

feat(auth): opt out of borrowing Codex CLI / Claude Code logins (auth.adopt_external_logins)

2026-09-18 20:55:24 -07:00

credits_tracker.py

refactor(credits): gate the re-warm on the pricing cache's own failure window

2026-09-20 14:02:17 +05:30

curator_backup.py

fix(curator): tarball skips regeneratable dirs; ledger --compact garbage-collects blobs

2026-09-19 09:20:54 -07:00

curator.py

fix(background_review): say so when the configured review route falls back to the main model

2026-09-19 12:30:32 -07:00

deadline.py

…

delegation_context.py

fix: kanban guidance and stop nudge resolve worker identity through one helper

2026-09-16 17:48:48 -07:00

display.py

fix(guardrails): mark every file-tool loop refusal and exempt it on the live failure seam

2026-09-18 10:03:04 -07:00

empty_response_guard.py

…

error_classifier.py

fix: name the CommandCode upstream-outage pattern

2026-09-20 15:28:21 -07:00

error_surface.py

chore: merge origin/main (resolve apps/desktop/src/components/assistant-ui/thread/assistant-message.tsx, website/docs/user-guide/desktop.md)

2026-09-19 10:54:43 -07:00

errors.py

…

estop.py

…

fallback_cooldown.py

fix(agent): pin the reset_at reach through try_activate_fallback and add opt-in fallback.min_switch_reset_seconds

2026-09-20 17:00:43 -07:00

fast_mode.py

…

file_safety.py

fix: drop no-op suppress around named-home expanduser in _guard_homes

2026-09-18 10:08:58 -07:00

gemini_native_adapter.py

fix(gemini): stop rerouting AQ. keys off the default Studio surface

2026-09-20 10:50:27 -07:00

gemini_schema.py

…

i18n.py

…

image_eviction_policy.py

fix(agent): quantize eviction by the fit window so heavy tool results hold the frontier

2026-09-17 13:52:36 +05:30

image_gen_provider.py

…

image_gen_registry.py

…

image_routing.py

fix(image-routing): drop the provider-wide supports_vision probe; mirror registry via public types in setup test

2026-09-19 20:56:31 -07:00

image_token_cost.py

…

inline_tool_executors.py

feat(connectors): one backend-owned connection operation, with a setup card on Desktop, TUI and CLI (#111008 )

2026-09-21 10:04:32 +05:30

insights.py

…

interrupt_compat.py

…

interrupt_control.py

fix(agent): clear_interrupt drops an unconsumed steer only on hard cancel

2026-09-18 13:01:12 -04:00

interrupt_scope.py

fix: attribute gateway lifecycle stops to their system issuer, not the user

2026-09-16 17:48:17 -07:00

iteration_budget.py

…

jiter_preload.py

…

kanban_stop.py

fix(kanban): stop-guard nudge names every worker exit; sync terminal-tool copies

2026-09-18 09:21:42 -07:00

lazy_forward.py

…

learn_prompt.py

…

learning_graph_render.py

…

learning_graph.py

fix: validate the skill name before opening its lock; key lock files on a digest

2026-09-15 19:03:33 -07:00

learning_mutations.py

…

lmstudio_reasoning.py

…

manual_compression_feedback.py

…

markdown_tables.py

…

memory_manager.py

…

memory_provider.py

…

message_content.py

…

message_metadata.py

…

message_sanitization.py

fix(agent): track image rejections per model across a fallback chain

2026-09-21 21:43:51 +05:30

micro_compaction.py

fix(agent): pop _db_persisted on in-place mutations of stamped live dicts

2026-09-20 15:26:34 -07:00

moa_alternation.py

fix: MoA aggregator merges adjacent same-role messages only for destinations that reject them

2026-09-18 20:56:35 -07:00

moa_loop.py

fix(moa): mark reused advisor guidance as predating the tool results

2026-09-20 15:51:19 -07:00

moa_trace.py

…

model_metadata.py

feat(providers): external-process provider support for standalone model-provider plugins (from #105863 )

2026-09-20 14:29:39 -07:00

models_dev.py

feat(providers): plugin-declared per-model capability metadata

2026-09-19 21:08:09 -07:00

moonshot_schema.py

…

native_compaction.py

fix(compression): enable Astra native compaction on official Codex OAuth

2026-09-19 00:53:26 -07:00

notification_presentation.py

refactor(notifications): one notice classification, one config read, one diagnostic-metadata helper

2026-09-18 01:43:35 +05:30

nous_rate_guard.py

fix(nous): keep anonymous errors exclusive to anonymous users

2026-09-17 12:28:25 +05:30

nous_wire.py

…

onboarding.py

…

oneshot_footprint.py

feat: one-shot runs drop the self-improvement footprint (no skill authoring, fewer process skills, delegation cap)

2026-09-19 09:21:18 -07:00

oneshot.py

…

opencode_affinity.py

chore: merge origin/main (resolve agent/opencode_affinity.py, tests/hermes_cli/test_web_server_idle_proof.py)

2026-09-19 10:56:19 -07:00

outbound_webhooks.py

…

periodic_scheduler.py

fix(agent): preserve periodic callback context

2026-09-16 11:06:06 -07:00

plan_prompt.py

…

plugin_llm.py

…

plugin_stream_hooks.py

fix(plugins): stream observer hooks and event subscribers fail-report once, not per event

2026-09-17 09:03:29 -07:00

portal_tags.py

…

process_bootstrap.py

fix(bootstrap): race IPv6/IPv4 for every sync connect from the process bootstrap, not per client

2026-09-18 09:56:28 -07:00

prompt_builder.py

fix(agent): remote backend probe no longer puts user, $HOME and cwd into the system prompt

2026-09-21 01:04:37 -07:00

prompt_cache_boundary.py

…

prompt_cache_scope.py

…

prompt_caching.py

…

provider_base.py

…

provider_media.py

fix(video-gen): trust the operator-configured origin on the first download hop

2026-09-18 10:22:14 -07:00

provider_projection.py

…

provider_registry.py

…

proxy_bypass.py

fix: bypass the proxy on loopback HTTP CDP discovery and keep NO_PROXY=* intact

2026-09-15 06:16:38 -07:00

rate_limit_credits.py

fix(credits): re-warm the Nous catalog when a header finds it cold after the TTL

2026-09-20 14:02:17 +05:30

rate_limit_tracker.py

…

reactions.py

…

reasoning_effort.py

fix: effort pickers and /reasoning status say what ultra really sends

2026-09-19 10:20:55 -07:00

reasoning_params.py

fix: send reasoning_effort=medium on custom endpoints when agent.reasoning_effort is unset

2026-09-20 16:03:38 -07:00

reasoning_summaries.py

…

reasoning_timeouts.py

fix(agent): extend the reasoning stale-timeout floor to gpt-6-astra and pin the gpt-5.6 family

2026-09-19 09:57:53 -07:00

redact.py

fix(redact): dotted sk- body cannot span "..", so display masks survive a second pass; Zhipu matcher is 32-hex.16+alnum

2026-09-18 09:58:12 -07:00

relay_llm.py

close managed stream when stream_current unwraps a completed response

2026-09-19 23:38:49 -07:00

relay_runtime.py

fix(relay): shared-metrics task close skips a scope buried under a concurrent turn

2026-09-19 23:42:20 -07:00

relay_tools.py

…

repetition_guard.py

fix: only drop an interrupted partial for the runaway repetition shape

2026-09-16 17:10:13 -07:00

replay_cleanup.py

…

retry_utils.py

fix(gateway): real 401 envelopes keep the sign-in reply; quota reply reads the shared reset grammar

2026-09-19 09:49:39 -07:00

review_engine.py

…

review_idle_queue.py

…

runtime_cwd.py

fix(tui): project-local skills register and dispatch for the session's repo

2026-09-18 10:22:47 -07:00

sdk_transform_bypass.py

fix(agent): keep moved chat fields as slot placeholders so the wire body is byte-identical; one shared escape hatch

2026-09-15 19:23:53 -07:00

search_policy.py

…

secret_scope.py

refactor(secret_scope): key the .env memo on utils.file_signature

2026-09-17 20:56:27 +05:30

served_model.py

feat(gateway): opt-in served_model footer field shows the model that really answered

2026-09-19 10:38:30 -07:00

session_activity.py

…

session_persistence.py

fix(gateway): a peer DM retries a transient turn failure once, like the other two lanes

2026-09-20 10:16:51 -07:00

shell_hooks.py

fix(agent): run bare script-path hooks on Windows through their interpreter

2026-09-20 11:27:39 -07:00

side_question.py

…

skill_bundles.py

…

skill_commands.py

fix(skills): built-in name collisions show a note in /skills, /help skills and the palette

2026-09-18 12:51:22 -07:00

skill_preprocessing.py

fix(windows): route every bare-bash spawn through _find_bash and surface silent interpreter failures

2026-09-20 15:50:26 -07:00

skill_utils.py

fix(skills): resolve project roots from session cwd

2026-09-18 10:22:47 -07:00

ssl_guard.py

…

ssl_verify.py

…

status_output.py

feat(notifications): opt-in suppression of user-channel warning notifications

2026-09-18 01:43:35 +05:30

stream_delivery.py

…

stream_diag.py

fix: name host, attempts and request size when connect retries are exhausted

2026-09-19 12:21:39 -07:00

stream_single_writer.py

…

subagent_lifecycle.py

…

subdirectory_hints.py

fix(agent): reject subdirectory hint files that resolve outside the tree

2026-09-20 10:51:03 -07:00

subscription_view.py

…

surface_switch.py

…

system_prompt.py

feat(notifications): opt-in suppression of user-channel warning notifications

2026-09-18 01:43:35 +05:30

terminal_approval_batch.py

fix(approval): prepare batches for desktop session sources

2026-09-16 06:27:23 -05:00

terminal_env_provider.py

…

terminal_env_registry.py

…

think_scrubber.py

…

thinking_timeout_guidance.py

…

thread_scoped_output.py

…

title_generator.py

fix: send the title model call after the turn on a shared custom endpoint

2026-09-20 14:09:57 -07:00

tool_dispatch_helpers.py

…

tool_executor.py

refactor: log the dead compression worker's exception once; trim guard comments

2026-09-21 21:34:44 +05:30

tool_guardrails.py

fix(guardrails): mark every file-tool loop refusal and exempt it on the live failure seam

2026-09-18 10:03:04 -07:00

tool_result_classification.py

fix(guardrails): mark every file-tool loop refusal and exempt it on the live failure seam

2026-09-18 10:03:04 -07:00

trace_upload.py

…

trajectory.py

…

transcript_repair.py

fix(compression): preserve steer display identity

2026-09-20 18:24:07 -07:00

transcription_provider.py

…

transcription_registry.py

…

tts_provider.py

…

tts_registry.py

…

turn_api_call.py

fix(agent): _should_stream keys ACP on the provider profile, not one vendor's name

2026-09-20 12:36:34 +05:30

turn_api_error.py

fix(agent): honor provider reset for fallback cooldown

2026-09-20 17:00:43 -07:00

turn_api_request.py

fix(agent): an image rejection strips the request, never the session history

2026-09-21 21:43:51 +05:30

turn_author.py

…

turn_context_compaction.py

fix(compression): an over-window session that cannot shrink ends the turn with /new guidance and waits one idle budget, not the ceiling

2026-09-20 12:52:07 -07:00

turn_context.py

feat(connectors): one backend-owned connection operation, with a setup card on Desktop, TUI and CLI (#111008 )

2026-09-21 10:04:32 +05:30

turn_empty_response.py

feat(notifications): opt-in suppression of user-channel warning notifications

2026-09-18 01:43:35 +05:30

turn_explainers.py

fix(sessions): storage maintenance refuses while a writer holds state.db; human-first retired-WAL guard text + recovery guide

2026-09-20 20:13:33 -07:00

turn_facade_lease.py

fix(agent): system watchdog interrupts are attributed to their issuer, not the user

2026-09-16 17:48:17 -07:00

turn_facade.py

…

turn_failure_copy.py

chore: merge origin/main (resolve apps/desktop/src/components/assistant-ui/thread/assistant-message.tsx, website/docs/user-guide/desktop.md)

2026-09-19 10:54:43 -07:00

turn_final_response.py

fix(agent): judge "wrong script" against the user's own message; one continuation kind per stop

2026-09-19 12:12:04 +05:30

turn_finalizer.py

fix: send the title model call after the turn on a shared custom endpoint

2026-09-20 14:09:57 -07:00

turn_iteration_prep.py

fix(agent): pop _db_persisted on in-place mutations of stamped live dicts

2026-09-20 15:26:34 -07:00

turn_liveness.py

…

turn_loop_errors.py

feat(notifications): opt-in suppression of user-channel warning notifications

2026-09-18 01:43:35 +05:30

turn_overflow.py

fix: output-cap failure copy stops pointing at the removed model.max_tokens key

2026-09-19 00:59:27 -07:00

turn_preflight_gate.py

feat(notifications): opt-in suppression of user-channel warning notifications

2026-09-18 01:43:35 +05:30

turn_preflight.py

fix(compression): probe aux feasibility before the first compaction; clear stale clamp notice on un-clamp

2026-09-18 09:49:14 -07:00

turn_recovery_autorecover.py

feat(agent): bounded auto-recovery ladder after retries and fallback are spent (#85426 , #107307 )

2026-09-19 12:38:55 -07:00

turn_recovery.py

fix(agent): track image rejections per model across a fallback chain

2026-09-21 21:43:51 +05:30

turn_request_assembly.py

fix: hand Codex reasoning-only stalls to the fallback provider instead of the incomplete sentinel

2026-09-19 00:04:42 -07:00

turn_response_check.py

chore: merge origin/main (resolve website/docs/user-guide/features/credential-pools.md)

2026-09-19 10:47:48 -07:00

turn_response_intake.py

fix: hand Codex reasoning-only stalls to the fallback provider instead of the incomplete sentinel

2026-09-19 00:04:42 -07:00

turn_retry_state.py

feat(agent): bounded auto-recovery ladder after retries and fallback are spent (#85426 , #107307 )

2026-09-19 12:38:55 -07:00

turn_stop_gates.py

fix(kanban): stop-guard nudge names every worker exit; sync terminal-tool copies

2026-09-18 09:21:42 -07:00

turn_summary.py

…

turn_tool_round.py

feat(notifications): opt-in suppression of user-channel warning notifications

2026-09-18 01:43:35 +05:30

turn_tool_validation.py

feat(notifications): opt-in suppression of user-channel warning notifications

2026-09-18 01:43:35 +05:30

turn_truncation.py

chore: merge origin/main (resolve agent/error_classifier.py,tests/agent/test_error_classifier.py)

2026-09-19 10:51:40 -07:00

turn_usage.py

fix(bedrock): restore Grok context with provider-confirmed cache provenance

2026-09-17 18:36:10 -07:00

usage_anchor.py

…

usage_pricing.py

feat(providers): external-process provider support for standalone model-provider plugins (from #105863 )

2026-09-20 14:29:39 -07:00

vault_login_classifier.py

…

vault_store.py

harden vault error handling for malformed state

2026-09-20 10:21:52 -07:00

verification_evidence.py

fix: verify-on-stop recognises python.exe / py launcher interpreters (review follow-up)

2026-09-20 10:23:04 -07:00

verification_stop.py

…

verify_hooks.py

…

vertex_adapter.py

…

video_gen_provider.py

fix(video_gen): build the OpenAI-compatible video client on the env-only-proxy httpx client

2026-09-19 09:51:12 -07:00

video_gen_registry.py

…

vision_message_prep.py

…

web_search_provider.py

…

web_search_registry.py

…