A Z.ai desktop user reported thinking reverting to medium after one turn,
burning ~200% of a week's credits in 4 days despite reasoning_effort: false
in config.yaml. Four compounding bugs:
- _session_info reported reasoning_effort "" for disabled reasoning,
indistinguishable from unset — the desktop adopted it after the first
turn, wiping its sticky "thinking off" pick so every later chat
reverted to the default effort.
- config.set key=reasoning always wrote agent.reasoning_effort to global
config.yaml, so every desktop model-menu selection (preset.effort ??
'medium') clobbered the user's configured value. Now session-scoped
like the messaging gateway's /reasoning, landing on
create_reasoning_override so lazily-built sessions keep it too.
- YAML `reasoning_effort: false`/`off`/`no` (boolean False) was coerced
to "" by every loader's `str(x or "")`, silently re-enabling thinking.
parse_reasoning_effort now treats False/"false"/"disabled" as
{"enabled": False}; loaders (tui gateway, gateway, cli, cron,
delegate) pass the raw value through. The desktop config reader also
crashed on the boolean (false.trim()), aborting voice/STT settings.
- The zai provider profile never sent thinking on the wire, and GLM-4.5+
defaults to thinking ON server-side — so disabling reasoning was a
silent no-op on direct Z.ai, the actual token burner. The profile now
emits extra_body.thinking {"type": "enabled"|"disabled"} for
thinking-capable GLM models, mirroring the DeepSeek profile.
Also: /new (session reset) now carries reasoning_config across the
rebuild like model_override; config.get reasoning prefers the session's
live value and maps a config False to "none"; Settings shows "Off"
instead of a blank select for hand-written false.
Salvaged from PR #3243 by @Mibayy, reimplemented against current main
(the original diff targeted a removed gateway/run.py handler).
- /compact is now a first-class alias of /compress (CLI, gateway,
Telegram/Slack/Discord command lists, autocomplete) — also fixes the
dangling '/compact' references in gateway error messages
(gateway/run.py context-exhausted banners).
- --preview / --dry-run: report what WOULD be compressed (message
counts, token estimate, 'here [N]' boundary) without touching the
transcript. Flags coexist with the existing 'here [N]' / focus-topic
args on both the CLI and gateway surfaces via shared pure helpers in
hermes_cli/partial_compress.py.
- --aggressive (LLM-free hard truncation) is intentionally NOT
implemented: it would need its own transcript-persistence branch
outside the guarded _compress_context rotation machinery (#44794
data-loss class). The flag is recognized and returns an explanatory
message pointing at '/compress here [N]' and /undo instead of being
mis-parsed as a focus topic.
- locales: gateway.compress.aggressive_unsupported added to all 16
catalogs (parity test enforced).
- release.py: AUTHOR_MAP entry for contributor credit.
Three CLI reliability fixes:
1. Interrupt reliability: chat() only re-queued the user's interrupt
message when the turn result carried interrupted=True. When the agent
thread raced past its last interrupt check (or finished) before the
interrupt landed, the message was silently dropped — and the stale
_interrupt_requested flag left on the agent instantly aborted the
NEXT turn. Un-acknowledged interrupt messages are now re-queued as
the next turn and the stale flag is cleared (only when the agent
thread actually exited). The clarify-race path also parks the message
in _pending_input instead of dropping it.
2. Slow exit (5+ min): stdlib ThreadPoolExecutor workers are non-daemon
and joined unconditionally by concurrent.futures' atexit hook — even
after shutdown(wait=False). One wedged tool worker (abandoned after
interrupt/timeout) held the process open forever. Promoted
async_delegation's daemon executor to a shared tools/daemon_pool
module and adopted it in tool_executor (concurrent tool batches),
memory_manager (background sync), delegate_tool (child timeout wrapper
+ batch fan-out), and skills_hub (source fan-out). Added a 30s exit
watchdog (HERMES_EXIT_WATCHDOG_S) armed at _run_cleanup start as a
backstop for wedged cleanup steps.
3. Exit jank: after prompt_toolkit tears down the input/status bars the
terminal sat silent for the whole cleanup window, looking hung. Print
'Shutting down… (finalizing session)' immediately at exit start.
E2E: live PTY interrupt of a foreground 'sleep 120' terminal tool now
aborts in ~1s and the typed message runs as the next turn; wedged-worker
+ wedged-cleanup subprocess exits in 5.8s (watchdog) instead of hanging.
In the interactive CLI, /journey dispatched straight to `args.func(args)`,
letting Rich write ANSI to stdout — which patch_stdout's StdoutProxy passes
through as literal `?[38;2;…m` garbage. Route the read-only views (default +
`list`) through a captured, force-color Console and re-emit via `_cprint`
(prompt_toolkit's ANSI parser), matching the `ChatConsole` idiom.
`delete`/`edit` stay on real stdio since they prompt / open `$EDITOR`.
hermes -w locks each worktree (reason 'hermes pid=<pid>'). git worktree
remove --force (single -f) refuses a locked tree, so a crashed session's
lock was never released and its worktree accumulated forever — a real
contributor to .worktrees/ bloat.
_prune_stale_worktrees now classifies each lock via _worktree_lock_is_live:
a live-owner pid is skipped at any age; a dead-owner (or foreign) lock is
unlocked first so the aggressive age-based cleanup can actually reap it.
The >72h reap tier is kept (that cleanup is intentional) but now guarded so
dirty/unpushed work is preserved, and branch deletion is gated on
git worktree remove succeeding. New fail-safe helpers _worktree_is_dirty
and _worktree_lock_is_live (pid liveness via gateway.status._pid_exists,
Windows-safe).
The streaming think-tag suppressors in cli.py (_stream_delta) and
gateway/stream_consumer.py (_filter_and_accumulate) matched tag names
with case-sensitive str.find(), so only the exact-case literals in the
tag tuples were caught. Mixed-case variants a model may emit — <Think>,
<ThInK>, <REASONING>, <Thought> — slipped through and leaked raw
reasoning into the user-visible stream.
Match against a lowercased view of the buffer with lowercased tag names
at all three sites (open-tag boundary search, partial-tag hold-back,
close-tag search) in both paths. Only KNOWN tag names are matched — no
substring matching — and the block-boundary gating that protects prose
mentions of <think> is preserved.
- 6 parametrized case-insensitive regression tests in each of
tests/gateway/test_stream_consumer.py and
tests/cli/test_stream_delta_think_tag.py.
Salvaged from PR #27289 by @YLChen-007.
Bare print() output is swallowed by patch_stdout while an interactive
prompt_toolkit Application owns the terminal, so /sessions and /history
rendered nothing. Route those emissions through _cprint (prompt_toolkit's
native renderer) when an app is running, and fall back to print otherwise.
Fixes#36815
Whole-bug-class follow-up to the tui_gateway fix: the same -1
last_prompt_tokens sentinel (parked by conversation_compression after a
compression) leaked into other status readers, producing a raw -1 or a
NEGATIVE usage_percent on the transitional turn:
- agent/context_engine.py get_status() (the ABC default every external
context engine inherits) — highest blast radius
- gateway/slash_commands.py /usage context line
- cli.py session usage printout
All clamped to >=0, mirroring cli.py _get_status_bar_snapshot and the
tui_gateway fix. Adds an ABC get_status sentinel-clamp regression test.
HermesCLI.process_command() and tui_gateway command.dispatch both handle
type: exec quick commands via subprocess.run(shell=True) with no env=
parameter, so the child inherits the full process environment — all API
keys and bot tokens stored in os.environ are visible to the script.
Any output is returned raw to the terminal or web-UI client without
redaction.
Fix: mirror the approach applied to gateway/run.py in #23584.
Apply _sanitize_subprocess_env() before spawning the subprocess and
redact_sensitive_text() on the collected output before display.
Symmetric across all three exec quick-command paths.
Parity with gateway/run.py fix in #23584.
The CLI routes user input typed while the agent is running into
``_interrupt_queue`` (separate from ``_pending_input``) so the explicit
interrupt path can opt to deliver them as a single combined message.
That path only drains the queue when ``busy_input_mode == "interrupt"``
AND a ``pending_message`` was acknowledged.
If the agent's turn finishes naturally (no interrupt fires), any
messages typed during the turn stay stuck in ``_interrupt_queue``
forever. Subsequent ``Enter`` presses route input to the same blocked
queue and the CLI appears to hang. Original report: lunarnexus in
The fix restores the post-turn drain that was originally part of
drain off as "worth its own review" and never re-landed it; the user-
visible regression is that any non-interrupt-mode user typing during
a turn is silently dropped.
Implementation: extract the drain to a small helper
``_drain_interrupt_queue_to_pending_input`` matching the existing
``_maybe_continue_goal_after_turn`` style. ``process_loop``'s
``finally`` block calls it once per turn after the status-line refresh
and before goal continuation (so re-queued user input preempts an
auto-continuation prompt). The helper swallows ``Exception`` so it
can never break the main loop.
Addresses #20271.
The --nous flag was only wired into the argparse `hermes debug share`
subcommand. The /debug slash command (classic CLI + TUI, both via
process_command -> _handle_debug_command) built a hardcoded args
namespace with no `nous` attribute, so it always took the default
paste.rs path.
Pass cmd_original through to _handle_debug_command and parse an optional
destination word:
/debug -> public paste (default, unchanged)
/debug nous -> Nous-internal S3
/debug local -> stdout, no upload
local wins over nous (never touches the network); unknown words fall
back to the default. Add args_hint="[nous|local]" so help/autocomplete
surface it. New TestDebugSlashCommand covers the parsing + dispatch.
Two narrow fixes that contribute to the TUI input black hole reported in
issue #16803, where the CLI keeps rendering but stops consuming user input.
1. MCP reload no longer blocks process_loop. _check_config_mcp_changes()
runs in process_loop's idle branch; the prior _reload_thread.join(timeout=30)
froze input consumption for up to 30s (longer if an MCP server hung). The
reload daemon already reports its own status via print(), so the join is
removed and the reload runs purely in the background.
2. Voice recording flag can no longer leak. _voice_recording was set True
before create_audio_recorder(), which runs outside any try/except. A
recorder-creation failure (no input device, PortAudio init error) left the
flag stuck True, so every future voice start was silently skipped by the
double-start guard. Recorder creation is now wrapped to reset the flag and
re-raise on failure, matching the existing start() handler.
Closes#16803
Interrupting the agent while an approval/clarify/sudo/secret prompt is up
left the overlay state dict set with no thread servicing it. The prompt's
worker thread is torn down on interrupt, but read_only (gated on
_command_running) plus the keypress filter kept the CLI input locked until
the prompt's own timeout expired — the terminal appeared frozen.
Drain and clear all four input-blocking overlays on interrupt via a single
helper (_clear_active_overlays_for_interrupt): approval -> deny,
clarify/sudo/secret -> cancel, each guarded so a dead queue can't block the
others; sudo restores the pre-modal draft. Wired into all three interrupt
paths — new-message interrupt, Ctrl+C, and Ctrl+Q. Blocking overlays now
clear AND fall through so one keypress both clears a stale overlay and
interrupts a still-running agent; the /model picker and slash-confirm
foreground prompts keep their cancel-and-return behavior.
Closes#13618.
A Kanban task referencing a non-existent skill (e.g. a typo'd name)
crashed the worker on startup via ValueError, which the dispatcher
retried until the task auto-blocked. Both cli.py and tui_gateway/server.py
now skip the unknown skill(s), log a warning, and continue with whatever
loaded — but still hard-fail when EVERY requested skill is missing, so a
fully-misconfigured worker fails loudly instead of running blind.
Closes#27136
Co-authored-by: Jimmy Johansson <jimmyjohansson84@users.noreply.github.com>
Terminal rendition of the desktop Star Map / Memory Graph: learned skills
and memories on a timeline, shared by `hermes journey` and the TUI
`/journey` overlay via one size-aware Python renderer
(agent/learning_graph_render.py).
- TUI overlay mirrors /agents: static chart overview + selectable slice
list → slice detail → single skill/memory body, with the shared
inverse-row selection treatment and a pinned footer.
- Reuse primitives: extract OverlayScrollbar into its own module (now
shared with agentsOverlay), scroll the item body via ScrollBox, and
unify both lists through one table-driven ListRow.
- No animation/playback in the TUI — pure data; the renderer's reveal
scrubber stays available in the CLI (`--play`, `--reveal`).
* feat(display): friendly human-phrased tool labels for built-in tools
Built-in tools now render ChatGPT-style status verbs ('Searching the web
for ...', 'Reading <file>', 'Browsing <url>') on the CLI spinner and
gateway/desktop tool-progress instead of the raw tool name.
- agent/display.py: _TOOL_VERBS map + build_tool_label() + set/get
friendly-labels flag (default on). Custom/plugin/MCP tools fall back to
the raw preview; verbose gateway mode left untouched (debug surface).
- tool_executor.py / tui_gateway / gateway: route the three spinner sites,
the TUI _tool_ctx, and the gateway all/new progress line through the label.
- config: display.friendly_tool_labels (default True, per-platform aware).
Zero new core tool / schema footprint — pure display layer.
* docs: add PR infographic for friendly tool labels
* fix(display): preserve arg preview in gateway friendly labels + update tests
The first gateway pass re-derived the label from the callback's `args`, which
is empty ({}) at the gateway tool.started callsite — the command/query lives in
the `preview` string, so terminal rendered as a bare '💻 Running' and dedup
collapsed consecutive commands. Now the gateway prefixes the verb onto the
already-computed preview via get_tool_verb/tool_verb_connector/verb_drops_preview,
preserving the command/url/query. CLI spinner path (real args) keeps build_tool_label.
Tests: update test_run_progress_topics exact-format assertions to the friendly
form ('💻 Running pwd'), add a format-agnostic preview extractor for the
truncation tests (works for both quoted-legacy and verb-prefixed output).
* test(tui): update resume-display context to friendly tool label
_tool_ctx now uses build_tool_label, so the desktop resume-view context for a
search_files turn reads 'Searching files for resume' instead of the bare
'resume' preview — consistent with live tool-progress. Update the assertion.
* test(tui): harden no-race worker test against sibling shard leakage
test_session_create_no_race_keeps_worker_alive flaked under -j 8: a daemon
build thread leaked from a prior session.create test in the same shard process
fires close/unregister against its own (foreign) session_key after this test
patches the global approval hooks, polluting the captured lists. Scope the
assertions to this session's own session_key so the regression intent
(this session's worker/notify must survive) is preserved while the test
becomes immune to shard composition. Not related to friendly-tool-labels.
The startup config/manifest reads used PyYAML's pure-Python SafeLoader,
which is ~8x slower than the libyaml-backed CSafeLoader C extension.
config.yaml is parsed several times during launch (cli config, raw
config, early interface/redaction bridge, logging config) and every
plugin manifest is parsed once — all on the slow path.
Add utils.fast_safe_load (CSafeLoader-preferring, pure-Python fallback,
true drop-in for safe_load) and route the hot startup parse sites
through it: hermes_cli/config.py (config + manifest reads),
hermes_cli/plugins.py (manifest parse), env_loader, cli.load_cli_config,
hermes_logging, and the two pre-config early YAML bridges in main.py.
Behavior is identical (same restricted safe tag set); only speed changes.
safe_load calls on the startup path drop from ~79 to ~0, cutting the
YAML parse cost from ~0.9s to ~0.15s under profiling.
Adds tests/test_fast_safe_load.py asserting equivalence with safe_load
across input shapes, empty-doc falsiness, C-loader preference, and that
python/object tags are still rejected (safe, not full loader).
Pull the #33271 post-interrupt recovery (flush_stdin + _force_full_redraw)
out of process_loop's finally block into _recover_terminal_after_interrupt(),
and replace the inline-logic-copy tests with ones that exercise the real
helper plus a source guard that process_loop still invokes it behind the
_last_turn_interrupted gate.
When the agent is interrupted during processing, prompt_toolkit's
renderer and VT100 input parser can be left in an inconsistent state.
CSI 6n cursor position report responses leak as literal text
(^[[19;1R) and the terminal stops accepting keyboard input.
Fix: in process_loop's finally block, after an interrupted turn:
- flush_stdin() to drain stray escape bytes from the OS input buffer
- _force_full_redraw() to reset prompt_toolkit's renderer cache
Closes#33271
When a MoA preset is selected, each reference model's answer now renders in the
CLI as a thinking-style block labelled with its source model, BEFORE the
aggregator responds — so the mixture-of-agents process is visible instead of a
silent pause. The aggregator's response (and its tool actions) follow as normal.
Mechanism (shared seam, all surfaces):
- MoAChatCompletions/MoAClient take an optional reference_callback and emit
'moa.reference' (index/count/label/text) per reference, then 'moa.aggregating'
(aggregator label) once. agent_init wires this to the agent's
tool_progress_callback, which every surface already consumes — so the events
reach CLI/TUI/desktop/gateway with no new plumbing.
- CLI _on_tool_progress renders 'moa.reference' as a labelled '┊ ◇ Reference
i/n — <model>' header + a thinking-style preview (reusing _emit_reasoning_
preview), and 'moa.aggregating' as a spinner transition. Display-only; never
touches message history (cache-safe).
Turn-scoped reference cache: the agent loop calls the facade once per tool-loop
iteration, but the advisory message view is identical across iterations within a
turn, so references are now run AND displayed once per user turn (keyed by the
advisory view's signature) instead of re-running/re-spamming on every iteration.
This also cuts reference API cost from O(iterations) back to O(turns).
Verified live via interactive PTY on the opus-gpt preset (gpt-5.5 + opus refs):
reference blocks render once per turn, labelled by model, before the aggregator;
fresh blocks on each new turn; aggregator tool actions still execute.
Follow-up: TUI/desktop rich rendering + gateway batched-summary already receive
the events via tool_progress_callback; their surface-specific renderers are a
separate change.
In the interactive CLI, the aggregator's tool calls under a MoA preset (or
any non-streaming model call, e.g. copilot-acp) appeared to overwrite each
other instead of building scrollable history. Each tool only updated the
transient spinner line; no committed scrollback line was printed.
Root cause: persistent tool lines in _on_tool_progress's tool.completed
branch were gated on tool_progress_mode in {all, new}, omitting 'verbose'.
Streaming models hid the bug because _on_tool_gen_start commits a 'preparing'
line per tool during streaming; non-streaming calls (MoA forces
_use_streaming=False) never emit that, so under 'verbose' there was no
committed line at all — only the self-overwriting spinner.
'verbose' is strictly more than 'all', so it now commits the same scrollback
line. Verified live via interactive PTY on the MoA opus-gpt preset: three
terminal calls in turn 1 and two in turn 2 each render as separate persistent
lines.
Selecting 'Mixture of Agents' in the `hermes model` provider picker fell
through silently — select_provider_and_model had no moa branch, so it just
reprinted the current model/provider summary and exited. And the CLI session
banner rendered the bare preset name (e.g. 'opus-gpt · Nous Research'),
which is meaningless out of context.
- Add _model_flow_moa: always lists the available presets (even one), then
prints the full reference-models + aggregator breakdown for the selection
and persists model.provider=moa / model.default=<preset> (dropping stale
base_url + endpoint creds, since moa is a virtual local provider).
- Wire the branch into select_provider_and_model.
- build_welcome_banner takes provider; when 'moa' it renders
'MoA: <preset> · agg <aggregator>' instead of a bare slug. Both CLI call
sites pass self.provider.
Tests: 2 new banner tests (moa + non-moa unchanged); E2E verified the picker
persists the preset and clears stale base_url/api_key.
prompt_toolkit's renderer sends ESC[6n cursor-position queries before
painting in non-fullscreen mode; the terminal replies ESC[<row>;<col>R.
Over SSH/cloudflared tunnels and slow PTYs these replies race past the
input parser and land in the display as raw '20;1R21;1R' text, and the
pending-CPR future can stall the renderer so the prompt freezes after the
agent's final answer.
Build the prompt_toolkit output with enable_cpr=False so CPR is marked
NOT_SUPPORTED up front and ESC[6n is never sent. This is the root-cause
counterpart to the existing input-side _strip_leaked_terminal_responses
scrubbing. Vt100_Output.from_pty() does not expose enable_cpr in
prompt_toolkit 3.x, so _build_cpr_disabled_output() reproduces its
get_size setup and calls the constructor directly; it returns None on any
failure so startup falls back to the default output.
Verified in a real PTY: baseline emits 1 ESC[6n query, the fix emits 0,
banner/UI render identically. Layout is unaffected — with CPR off the
renderer sizes the prompt to its preferred height (the same fallback
prompt_toolkit uses on any terminal that doesn't answer CPR).
Co-authored-by: Hermes Agent <noreply@nousresearch.com>
/moa no longer does a sticky model switch. It now always runs a single
prompt through the default MoA preset and restores the prior model
afterward; the whole argument is the prompt (no preset-name matching).
To switch to a MoA preset for the session, select it from the model
picker, where presets already surface under a virtual Mixture of Agents
provider on every model-selection surface.
Also fixes#53444: the TUI one-shot only set session[model_override],
which the already-built cached agent ignored, so MoA silently never ran
and the turn used the original model. The TUI now does a real in-place
agent.switch_model() via _apply_model_switch() when a live agent exists
(with a proper restore after the turn), and falls back to a model_override
for lazy/unbuilt sessions.
Removes the redundant sticky-switch branch from the CLI, gateway, and TUI
/moa handlers; updates the command description, usage string, and docs.
A top-level delegate_task dispatches in the background and re-enters as a
fresh turn when done. Print a one-line dispatch-time note — no spinner,
nothing to poll — so the idle prompt doesn't read as "nothing happened."
* feat(moa): expose MoA presets as selectable virtual models
Reconstructed onto current main (PR #46081's base had diverged with no common
ancestor, marking the PR dirty so CI never dispatched). MoA is now a virtual
provider: each named preset is a selectable model under provider 'moa', and the
preset's aggregator is the acting model that answers and calls tools.
Reference models fan out in parallel via a bounded ThreadPoolExecutor (the same
batch pattern delegate_task uses) — all references dispatched at once, collected
when every one finishes, then handed to the aggregator. Output order is
preserved, failures and the MoA-recursion guard stay isolated per reference.
- Removed the old mixture_of_agents model tool and moa toolset.
- Added moa as a virtual provider in the provider/model inventory.
- /moa is shortcut behavior over model selection (default preset / named preset
/ one-shot prompt).
- Dashboard + Desktop manage named presets; presets appear in model pickers.
- Parallel reference fan-out in agent/moa_loop.py with regression test.
* fix(moa): thread moa_config through _run_agent to _run_agent_inner
The reconstructed gateway MoA wiring declared moa_config on _run_agent (the
profile-scoping wrapper) and used it inside _run_agent_inner, but the wrapper
never forwarded it — _run_agent_inner had no such parameter, so the runtime hit
NameError: name 'moa_config' is not defined on the compression-failure session
sync path. Add moa_config to _run_agent_inner's signature and forward it from
both wrapper call sites (multiplex and non-multiplex). Caught by
tests/gateway/test_compression_failure_session_sync.py on CI shard test(4).
* fix(moa): classify moa as a virtual provider in the catalog
The moa virtual provider has no PROVIDER_REGISTRY/ProviderProfile entry, so
provider_catalog() fell through to the default auth_type="api_key" with no
env vars — tripping two catalog invariants:
- test_provider_catalog: api_key providers must expose a credential env var
- test_provider_parity: every hermes-model provider must be desktop-configurable
moa already declares auth_type="virtual" in HERMES_OVERLAYS; consult that
overlay as an auth_type fallback so the catalog reports moa as virtual (no real
credential, no network endpoint). Exempt virtual providers from the desktop
parity union check the same way 'custom' is exempt — derived from the catalog,
not a hardcoded slug, so future virtual providers are covered too.
Ship the final pet-generation UX polish (provider picker behavior, step-2 cancel flow, banner integration, and visual consistency) and make saturated-chroma background removal C-op driven so hatch processing no longer hammers the machine during long runs.
Open-ended skill learning across every surface. /learn <free text> takes a
description of any source — a directory, a URL, the workflow you just walked
the agent through, or pasted notes — and the live agent gathers it with the
tools it already has (read_file/search_files, web_extract, the conversation,
the pasted text), then authors a SKILL.md via skill_manage following the
house authoring standards (<=60-char description, the standard section order,
Hermes-tool framing, no invented commands).
No engine, no model-tool footprint, works on any terminal backend (local,
Docker, remote): /learn builds a standards-guided prompt and hands it to the
agent as a normal turn.
- agent/learn_prompt.py: shared standards-guided prompt builder
- /learn registry entry (both surfaces) + CLI handler (inject onto input
queue) + gateway handler (rewrite turn, fall through, /blueprint pattern)
- tui_gateway command.dispatch returns a send directive -> TUI + dashboard chat
- dashboard Skills page 'Learn a skill' panel (dir + URL + open-ended text)
composes a /learn request and runs it in chat
- docs (slash-commands ref + skills feature page), 11 targeted tests
Inspired by OpenAI Codex's Record & Replay and the /learn concept from #47234
(dir-distillation engine); reworked to be open-ended and engine-free per
review.
The classic prompt_toolkit status bar already shows two background
indicators: ▶ N (/background agent threads) and ⚙ N (shell processes
spawned by terminal(background=true)). Background/async subagents
(delegate_task batches and background single delegations) had no
indicator despite being long-running work the user should be able to
see at a glance.
Add a third indicator ⛓ N sourced from
tools.async_delegation.active_count() — the count of delegations still
in the 'running' state. Renders in the plain-text builder and the
styled-fragment builder across the same width tiers as the other two
(omitted on the narrow <52 tier), guarded so a raising active_count()
leaves the snapshot at 0.
* feat(goals): add /goal wait <pid> barrier to park the loop on a background process
The /goal loop re-pokes the agent every turn via the post-turn judge. When a
goal is gated on a long-running background process (CI poller, build, test
matrix, deploy) that produces nothing to judge yet, this spins the agent into
'is it done?' busy-work and burns the turn budget.
/goal wait <pid> [reason] parks the loop: while the PID is alive, the judge is
skipped, no turn is consumed, no continuation fires, and /goal status shows a
parked indicator. The barrier auto-clears the moment the process exits (the
agent's notify_on_complete watcher is the natural wake signal), then the next
turn resumes normal judging. /goal unwait clears it manually; pause/resume/clear
drop it; a dead/stale PID can never wedge the loop.
Wired across CLI, gateway, and the mid-run command guard for parity. Barrier
persists in SessionDB.state_meta (survives /resume); GoalState gains
backward-compatible waiting_on_pid/waiting_reason/waiting_since fields. 12 new
tests; docs updated.
* fix(goals): use gateway.status._pid_exists for liveness, not os.kill(pid,0)
The Windows-footguns CI guard flagged os.kill(pid, 0) in _pid_alive — on
Windows that's not a no-op, it routes to CTRL_C_EVENT and hard-kills the
target's console process group (bpo-14484). Delegate to the canonical
footgun-safe gateway.status._pid_exists (psutil + ctypes/POSIX fallback)
instead, with a direct-psutil last resort.
* feat(goals): judge-driven auto-wait — the loop parks itself, no manual /goal wait
Makes the wait barrier automatic. Every turn the judge is shown the agent's
live background processes (pid, command, uptime, output tail from the
process_registry) alongside the goal + response, and can return a new 'wait'
verdict instead of continue:
{"verdict":"wait","wait_on_pid":N} → park until that process exits
{"verdict":"wait","wait_for_seconds":N} → park until the deadline passes
evaluate_after_turn acts on the directive (sets the barrier, parks the loop)
so the agent isn't re-poked into busy-work while CI/builds/deploys run. Adds a
time-based waiting_until barrier alongside the pid barrier; both auto-clear and
can never wedge the loop. Drivers (CLI, gateway, tui_gateway) feed the live
registry in via gather_background_processes(). Manual /goal wait stays as an
override. Judge verdict contract widened to (verdict, reason, parse_failed,
wait_directive); legacy {"done":bool} shape still accepted.
* test(goals): update kanban _fake_judge to the 4-tuple judge contract
CI test(3) caught it: test_kanban_goal_mode's _fake_judge still returned the
3-tuple (verdict, reason, parse_failed), but the kanban loop now unpacks the
4-tuple (+ wait_directive). Update the fake to return None for the directive
and accept the background_processes kwarg.
* feat(goals): trigger-based wait — park on a process's own signal, not just exit
Addresses two gaps in the judge-driven wait: (1) the judge could only express
'wait until PID exits' or 'wait N seconds', so a long-lived watcher/server that
fires a trigger MID-RUN (and may never exit) couldn't be waited on; (2) the
process's own watch_patterns/notify_on_complete trigger was invisible to the judge.
Adds a session-based barrier (waiting_on_session) that releases on the process's
OWN trigger via process_registry.is_session_waiting(): the session exits, OR (if
started with watch_patterns) its pattern matches — even while the process keeps
running. list_sessions() now surfaces session_id + watch_patterns/watch_hit/
notify_on_complete so the judge sees the trigger and is told to prefer
wait_on_session for trigger processes. Judge verdict gains a {wait_on_session}
directive (preferred over pid). Backward-compatible GoalState field; pid + time
barriers unchanged.
Tests: TestSessionTriggerBarrier (release on mid-run pattern match while alive,
release on exit, unknown-session, full park→trigger→resume, parse, validation,
backcompat load). 105 goal-surface + 85 process_registry tests green.
display.timestamps already drove the [HH:MM] suffix on live submitted and
streamed message labels, but there was no runtime command to toggle it and
/history ignored the setting entirely. Add /timestamps [on|off|status]
(alias /ts) and render [HH:MM] in /history for turns that carry a stored
unix timestamp (resumed sessions). Live unsaved turns without a stored time
are never given a fabricated one. Uses the existing sanctioned non-wire
'timestamp' message key (stripped before the API call in chat_completions),
so message-alternation and prompt-cache invariants are untouched.
Ctrl+G already opened $EDITOR with the current draft, but used
open_in_editor(validate_and_handle=False), which only loaded the saved text
back into the input area — the user still had to press Enter. The TUI's
Ctrl+G (openEditor) submits the draft on a clean exit. Since CLI submission
is driven by the custom Enter keybinding (not the buffer accept_handler),
validate_and_handle can't route through it; instead chain a done-callback on
the editor Task that calls the new _submit_editor_buffer(), which mirrors the
Enter handler's idle/queue/slash branches and drops an empty save.
terminal.docker_extra_args passes flags verbatim to `docker run` (e.g.
--gpus=all, --shm-size=16g). It was wired into DEFAULT_CONFIG,
TERMINAL_CONFIG_ENV_MAP (so `hermes config set` bridged it),
terminal_tool._get_env_config (reads TERMINAL_DOCKER_EXTRA_ARGS), and
DockerEnvironment (applies extra_args) -- but it was MISSING from cli.py's
env_mappings and gateway/run.py's _terminal_env_map.
Consequence: a user who hand-edits config.yaml (rather than running
`hermes config set`) has docker_extra_args silently dropped on the CLI and
gateway/desktop startup paths, while docker_image / docker_volumes (which
ARE in those maps) bridge correctly -- producing the reported 'Hermes
partially reads the Docker config' symptom where --gpus=all and
--shm-size=16g never reach docker run.
This is the same bridge-coverage bug class that shipped before for
docker_run_as_host_user (cli + gateway) and docker_mount_cwd_to_workspace
(gateway). Fix by adding the key to both maps, plus a dedicated regression
pin in test_terminal_config_env_sync.py mirroring the existing
test_docker_*_is_bridged_everywhere guards.
* feat(cli): /prompt — compose your next prompt in $EDITOR
Adds /prompt (alias /compose): opens $VISUAL/$EDITOR on a temp markdown
file so you can hand-edit a multi-line prompt, then sends the saved buffer
as the next agent turn. Text after the command pre-seeds the buffer; an
empty save cancels. Reuses the one-shot _pending_agent_seed the interactive
loop already consumes (same mechanism as /blueprint), so no changes to the
input event loop or message pipeline. CLI-only.
* feat(tui): /prompt slash command opens $EDITOR (parity with CLI)
The TUI already opens $EDITOR via Ctrl+G (openEditor), but had no /prompt
slash command like the classic CLI. Wire openEditor into the slash handler
context and register /prompt (alias /compose) to call it; inline text after
the command is dropped into the composer first so it carries into the editor,
matching the CLI's /prompt <text>.
* feat(cli): /reasoning full to show complete thinking, not 10-line clamp
The post-response Reasoning recap box hard-clamped long thinking to the
first 10 lines, so there was no way to see the full reasoning trace after
a turn (live streaming already shows it in full). Add display.reasoning_full
(default off) plus /reasoning full|clamp to toggle it at runtime; the clamp
truncation note now points at the command. Addresses repeated user requests
to show all thinking tokens.
* test(gateway): de-snapshot /reasoning help assertion
The test froze the exact args-hint literal '/reasoning [level|show|hide]',
which the new full/clamp args change to '[level|show|hide|full|clamp]'.
Convert to an invariant: assert /reasoning is in help and carries its core
args, not the exact hint string.
* feat(tui): /reasoning full|clamp parity in tui_gateway
The classic-CLI reasoning_full toggle had no TUI equivalent — typing
/reasoning full in the TUI fell through to parse_reasoning_effort and
errored. The TUI renders thinking as an expand/collapse section (no fixed
10-line recap), so map full -> sections.thinking=expanded (raw, uncapped
via thinkingPreview mode='full') and clamp -> collapsed, persisting
display.reasoning_full for cross-surface config consistency.
* feat(providers): remove google-gemini-cli + google-antigravity OAuth providers
Google now actively bans accounts for third-party tools that piggyback on
Gemini CLI / Antigravity / Code Assist OAuth, and because abuse prevention
sits at a backend layer the ban can extend to the entire Google account
(Gmail/Drive), with a second violation being permanent.
Ref: https://github.com/google-gemini/gemini-cli/discussions/20632
Removes both OAuth inference providers entirely (modules, provider profiles,
auth/runtime/config/models wiring, the /gquota Code Assist quota command,
the antigravity-cli optional skill, desktop + docs surface in en + zh-Hans).
The API-key 'gemini' provider (GOOGLE_API_KEY/GEMINI_API_KEY against
generativelanguage.googleapis.com) is unaffected and stays fully supported.
* fix(skills): keep the antigravity-cli skill — only the OAuth provider is removed
The antigravity-cli optional skill orchestrates the external `agy` binary as
a coding-agent tool via the terminal tool — it does NOT wrap Hermes inference
through the banned google-antigravity OAuth provider, so it carries none of
the account-ban risk that motivated removing that provider. Restore the skill,
its docs page, the sidebar entry, and the optional-skills catalog row. The
google-antigravity / google-gemini-cli inference providers stay fully removed.
When a /model switch resolves a valid model but the in-place agent swap
fails mid-conversation (expired key, unreachable base_url), the agent
rolls itself back to the old working model+client and re-raises. The
callers caught that re-raise, logged a warning, then committed the broken
switch anyway: wrote the failed model to the session DB, set
_session_model_overrides to the broken model/provider/key, and (gateway
direct path) evicted the working cached agent. The next message then
rebuilt a dead agent from the broken override -> permanently unusable
conversation (#50163).
Fix the whole caller class so a failed swap aborts the commit entirely:
- gateway/slash_commands.py (picker + direct /model paths): on swap
failure, early-return an error message; skip DB persist, session
override, cache eviction, and config write.
- cli.py (both /model handlers): snapshot CLI-level credential/runtime
fields before mutating, restore them on swap failure, and abort the
note + success print.
- tui_gateway/server.py: wrap the previously-unguarded swap; on failure
raise a clean error and skip worker restart, runtime persist, switch
marker, session model_override, and config persist.
The no-cached-agent path (apply-on-next-session) is unaffected.
Adds a gateway regression test that fails on the pre-fix behavior.
hermes -w created the worktree branch from the standalone clone's HEAD, which
lags origin when the clone isn't freshly updated (it's only refreshed by
hermes update, not per session). Every worktree branch then rooted on a stale
base, so the PR diff GitHub computes against current main ballooned with
unrelated changes and the agent had to discover the staleness at push time and
rebase.
_resolve_worktree_base() now fetches and branches from the freshest available
ref: the current branch's upstream if it tracks one (so a deliberate
feature-branch worktree tracks its own remote), else the remote's default
branch (origin/HEAD), else local HEAD as a fail-soft fallback (offline / no
remote / detached). A bogus 'origin/(unknown)' default is guarded, and worktree
creation retries from HEAD if branching off the remote ref fails — so this is
never worse than the old behavior.
Gated by worktree_sync (default true); set worktree_sync: false to keep the
old branch-from-local-HEAD behavior. The resolved base is printed in the
session banner.
This is the follow-up to the #50319 session, where the standalone clone was
213 commits behind origin and the worktree inherited that stale base.
Follow-up to the salvaged preflight-compression warning:
- Replace silent `except Exception: pass` at all 5 guard call sites
(cli.py x2, gateway/slash_commands.py x2, tui_gateway/server.py) with
`logger.debug(...)` so signature drift in the guard helper isn't hidden.
- tui_gateway/server.py: set the confirm dict's `warning` field to the
merged message (was bare expensive-model text) so it matches
`confirm_message` for any future consumer reading `warning`.
- Add trailing newlines to the two new files.
Adds hermes_cli/context_switch_guard.py mirroring the model_cost_guard
pattern. When a user switches models mid-session (Herm TUI picker, CLI,
or /model on Telegram/Discord), the warning surfaces on the existing
ModelSwitchResult.warning_message path used by the expensive-model
guard if the new model's compression threshold is below the current
session size.
Partial fix for #23767 — addresses only the 'user-facing guardrail
when switching from a high-context provider to a substantially
lower-context provider' slice. The other proposed fixes from that
issue (hard preflight token guard, metadata cache invalidation on
switch, compression safety invariant, oversized tool-output handling)
are out of scope for this PR.
Render the reactive pet pane in the classic CLI (steady redraw,
right-aligned) and wire the /pet command to list and switch pets, plus an
enable/disable toggle. Backed by hermes_cli/pets.py and the CLI commands
mixin, registered in the central command registry. Covered by the CLI pet
pane and toggle tests.
Makes the CLI memory-provider shutdown path observable: log when CLI
cleanup calls memory shutdown (with session id + message count), warn
instead of swallowing CLI memory-shutdown exceptions, warn on
on_session_end failures during agent shutdown, and raise the
MemoryManager provider-hook failure log from debug to warning with a
traceback.
Salvaged from PR #49287 (authored by Gille / @helix4u).
Second review pass (Codex + Hermes subagent). Codex reproduced a real race with
a two-thread harness; both converged on the remaining issues.
- Generation-aware publish (fixes a lost-update race): two refresh callers (the
late-refresh daemon and the between-turns prologue around turn 1) could each
compute a snapshot outside the lock; a SLOWER caller holding an OLDER registry
generation could acquire the publish lock after a newer caller and clobber it,
deleting just-landed tools. refresh_agent_mcp_tools now captures
registry._generation before computing and refuses to publish a stale set;
agent._tool_snapshot_generation tracks the published generation.
- Context-engine routing names (_context_engine_tool_names) are now staged on a
local and published atomically with the snapshot, and only claimed when this
rebuild actually appended the schema — matching agent_init's dedup so a
registry/plugin tool of the same name keeps its own dispatch. (Previously
mutated live, before the publish lock, and on no-change refreshes.)
- CLI /reload-mcp: self.enabled_toolsets is resolved once at startup, so a
server newly ENABLED in config mid-session wasn't picked up (TUI already
re-resolved). Merge now-connected MCP server names into the override (unless
the user pinned all/*), mirroring startup, and keep self.enabled_toolsets in
sync. Closes the CLI/TUI parity hole.
- ACP (acp_adapter/server.py) routed through the shared helper — it was a 5th
sibling rebuild that re-injected memory tools but NOT context-engine tools and
bypassed the atomic/name-diff path (inert today, fragile).
- mcp_startup._resolve_discovery_timeout pulls its default from DEFAULT_CONFIG
(single source of truth) instead of a stale hardcoded 5.0 literal.
- Tests: stale-generation-no-clobber, _skip_mcp_refresh honored, timeout
fallback uses DEFAULT_CONFIG.
Consolidated findings from three independent reviewers (Codex, Claude Code, a
Hermes subagent w/ the hermes-agent-dev skill):
- BLOCKING: refresh_agent_mcp_tools rebuilt only the registry subset, silently
dropping post-build-injected memory-provider (mem0/honcho/…) and context-
engine (lcm_*) tools on every refresh. Now additive-preserving: re-applies
the same injectors agent_init uses, staged on locals and published atomically.
- Re-injection now honors the #5544 enabled_toolsets gate for context-engine
tools, so a restricted-toolset platform can't get lcm_* leaked back in.
- Atomic read-diff-publish under one lock: the returned `added` set and the
(tools, valid_tool_names) pair are consistent even under concurrent callers
(no half-swap, no TOCTOU).
- background_review fork opts out (_skip_mcp_refresh) so its byte-identical
tools[] cache parity with the parent is preserved.
- CLI /reload-mcp routed through the shared helper (was a 4th divergent copy
with the same clobber bug + missing disabled_toolsets).
- Explicit reloads (TUI RPC + CLI) pass enabled_override so a server the user
just enabled in config this session is picked up; automatic paths reuse the
agent's build-time selection.
- mcp_discovery_timeout default 5.0 -> 1.5s: correctness now comes from the
between-turns refresh, so the startup wait is only a small turn-1 UX bump
rather than a heavy dead-server latency penalty.
- has_registered_mcp_tools checks registered TOOLS (not connected servers) so a
zero-tool/prompt-only server doesn't make the per-turn hook fire forever.
- Tests: rewrote the thread-safety test to actually exercise the write path
(alternating tool sets), added the #5544-gate regression, the memory/context
preservation regression, and a "callable next turn via valid_tool_names"
contract; removed a dead monkeypatch line.
The classic CLI status bar could appear twice after a horizontal terminal
resize — two bars at two widths with two different elapsed readings.
Root cause: prompt_toolkit's Application._on_resize() calls renderer.erase(),
which does cursor_up(_cursor_pos.y) + erase_down() using the _cursor_pos.y
cached from the LAST render at the OLD width (renderer.py:745). On a column
shrink the terminal reflows the already-painted full-width chrome into extra
physical rows, so the cached y undershoots: cursor_up doesn't climb past the
reflowed rows and erase_down leaves the old bar stranded ABOVE the live
origin. The next paint stacks a fresh bar below it. The existing post-resize
suppression hides the NEW bar for ~0.35s but never erases the already-reflowed
OLD one, so the ghost survives the whole window. Ctrl+L / /redraw clears it,
confirming a viewport wipe is the fix.
Fix: on a WIDTH change, _recover_after_resize now routes through the same
recovery as Ctrl+L — _clear_prompt_toolkit_screen(rebuild_scrollback=False)
(CSI 2J, visible viewport only) + _replay_output_history() — BEFORE delegating
to prompt_toolkit's resize. Banner-safe: 2J never touches scrollback history
(that's CSI 3J, which we don't send here), so the startup banner is preserved.
Rows-only resizes skip the clear (no reflow → no ghost) to avoid an extra
repaint. Tracks _last_resize_width to distinguish the two.
Tests: replace the now-obsolete 'never clears on resize' assertion with two
tests — rows-only resize delegates without clearing; width change clears the
viewport + replays and never wipes scrollback.
