The textarea focuses on mount and when an overlay closes (remount), but focus
could drift to the transcript scrollbox on a mouse-scroll, dropping keystrokes.
Now (opencode's keep-the-prompt-focused idea, adapted):
- onMouseDown → focus the textarea (click-to-focus).
- a global keystroke net: a PRINTABLE, unmodified key while the textarea is
unfocused reclaims focus AND recovers the char (the in-flight event went to
the global handler, not the unfocused textarea, so insert it). Nav/scroll keys
(arrows/page/home/end/…) are deliberately left alone so keyboard transcript
scroll still works; kitty `release` events are skipped to avoid double-insert.
Completion accept/dismiss handler folded into the same useKeyboard with early
returns.
Live-smoked: type → text lands; `/` → completions; Esc → dismiss; type again →
lands; clean quit. 60 pass.
Item 11 — "stopping the agent doesn't work". Ctrl+C used to immediately destroy
the renderer. Now a turn-aware state machine (opencode's double-press model, the
user's preferred behaviour):
- While a turn runs (store.info.running): first Ctrl+C → session.interrupt
{session_id} (STOP the agent), and arms a 3s quit window with a warn hint
"⏹ stopped — Ctrl+C again to quit".
- Idle: first Ctrl+C arms the window ("Ctrl+C again to quit"); a stray single
press never nukes the session.
- A second Ctrl+C within the window KILLS the TUI (renderer.destroy → clean
scope teardown → gateway child EOF).
- A blocking prompt still owns Ctrl+C (deny/cancel) — unchanged.
Wiring: renderer.ts gains an `onCtrlC` hook (owns Ctrl+C when not blocked);
entry builds the machine (gateway yielded before the renderer so it can read
`running` + send interrupt). store gains a transient `hint` slice; StatusLine
shows hint (warn, priority) or the busy face (dim).
Live-smoked: long turn → Ctrl+C shows "stopped" + idle dot; second press exits
cleanly with no orphaned gateway child (the user's installed-venv sessions
untouched). 60 pass.
Item 14: a persistent bottom-chrome status bar, ported from Ink's appChrome
StatusRule. Sourced from the session.info event (model / reasoning_effort /
fast / cwd / branch / running / usage.context_*) which was decoded but dropped
until now; also folded session.create/resume result.info and message.complete
usage into a new store `info` slice.
- store: SessionInfo slice + applyInfo(); session.info handler; message.start/
complete flip `running` (the flag the Ctrl-C interrupt will read); refresh
usage on complete.
- schema: MessageComplete.payload gains loose `usage` so it survives decode.
- view/statusBar.tsx: width-aware (Ink progressive disclosure) — context bar
drops on narrow terminals, cwd compacts to last two segments + left-truncates
so the row never wraps. Turn/connection dot ◐/●/○.
- App: status bar sits ABOVE the composer; a top-edge rule (border:['top'])
visually separates the status bar + textbox input region from the transcript.
- tests: store info slice (3) + headless status-bar render (1); bumped the
approval-prompt capture height for the taller input region. 59 pass.
Live-smoked: bar shows model·effort·context%·dir; context updates 0→4% across a
turn; running dot flips; separator divides input region from transcript.
Live-usage issue 3/5: the kaomoji faces ("(¬_¬) processing…") lingered in the
transcript. Traced (instrumented capture): they arrive via `thinking.delta` —
Hermes's transient kaomoji busy *indicator* (_INDICATOR_DEFAULT=kaomoji), which I
was rendering as a persistent reasoning part.
- store: new transient `status` field. thinking.delta / status.update → `status`
(not a part); message.start + message.complete clear it. Only the real
`reasoning.delta` still becomes a (dim) transcript part.
- view/statusLine.tsx: a dim busy line above the composer shown while `status` is
set (Ink's FaceTicker analog), rendering nothing when idle; wired into App
between the transcript and the input zone.
Verified: bun run check green (55 tests / 7 files) — store tests assert
thinking.delta → status (no transcript part) + cleared on complete; status.update
→ status. Live tmux: a turn showed "٩(๑❛ᴗ❛๑)۶ cogitating…" on the transient status
line (cleared on completion) with NO face left in the transcript.
From live-usage feedback (driving the real TUI):
- Mouse ON by default (opencode parity; HERMES_TUI_MOUSE=0 opts out). Was hardcoded
off, which is why transcript wheel-scroll, scrollbar drag, and click-to-expand
tools didn't work and the terminal's native region-select polluted copy. With
useMouse the scrollbox handles the wheel + scrollbar and tools are click-expandable;
selection becomes OpenTUI's text-aware select. (Mouse can't be driven via tmux
send-keys — verify wheel/drag/click interactively.)
- Streaming markdown: match opencode's v2 text path —
<code filetype="markdown" streaming drawUnstyledText={false}>. The previous
drawUnstyledText:true drew raw text then overlaid styling each delta (a flash);
false avoids that and re-tokenizes incrementally for smoother streaming. (The
native renderable's tree-sitter doesn't settle in the headless test renderer with
drawUnstyledText:false, so the two markdown frame tests now assert the assistant
text via the store — paint is verified in the live smoke; render.ts also settles
to waitForVisualIdle.)
Verified: bun run check green (53 tests / 7 files). Live mouse + streaming
smoothness for glitch to confirm. Part of the live-feedback polish goal.
Repoint hermes_cli/main.py `_make_opentui_argv` from the superseded React entry
to the v4 Solid + Effect-at-boundary entry: it now prefers
`ui-tui-opentui-v2/src/entry/main.tsx` (cwd ui-tui-opentui-v2) and falls back to
`ui-tui-opentui/src/entry.real.tsx` only if the v2 package is absent (graceful
during coexistence). The engine gate (_resolve_tui_engine: HERMES_TUI_ENGINE /
display.tui_engine → opentui; Windows/Termux → Ink fallback) and the dual-engine
dispatch in _make_tui_argv are unchanged; Ink (ui-tui/) is untouched. The spawned
tui_gateway's source-root default lands on PROJECT_ROOT (package at
<root>/ui-tui-opentui-v2), so it loads Python from the same checkout, no extra env.
So `HERMES_TUI_ENGINE=opentui hermes --tui` now launches the v4 engine — the exact
`bun …/v2/src/entry/main.tsx` invocation live-smoked across P1–P5e, making every
first-class surface reachable from the real CLI.
Also: a consolidated 3-way acceptance summary (Ink ↔ opencode ↔ build) at the top
of opentui-feature-map.md covering all 7 first-class surfaces + the foundation +
the launcher, each ✅ + tested + smoked.
Verified: py_compile main.py OK (dev-skill rule for the 4k-line file); imported
the worktree CLI with HERMES_TUI_ENGINE=opentui → _resolve_tui_engine()='opentui',
_make_opentui_argv() → [bun, …/ui-tui-opentui-v2/src/entry/main.tsx] (cwd
ui-tui-opentui-v2, --watch in dev). v2 `bun run check` green (53 tests / 7 files).
Smoke P8 + matrix updated. Remaining: header chrome detail (5b), agent-feature
trail (5d), distribution (§10) — polish, not first-class blockers.
The agents dashboard (spec §2b; Ink agentsOverlay) — the last first-class
interactive surface. Subagent delegations are tracked from the `subagent.*`
event stream and shown in a full-height overlay.
- store: subagents[] built from subagent.{spawn_requested,start,thinking,tool,
progress,complete} by subagent_id (status·goal·model·depth·lastTool·summary);
clearTranscript clears them. dashboard flag + openDashboard/closeDashboard.
- view/overlays/agentsDashboard.tsx: full-height overlay (replaces transcript+
composer), depth-indented subagent rows colored by status, scroll via
scrollBy/scrollTo, Esc/q close. Empty state prompts to delegate.
- view/App.tsx: content zone is now a <Switch> — pager / agents dashboard /
(transcript + input zone).
- logic/slash.ts: /agents, /tasks → openDashboard (SlashContext.openDashboard).
Verified: bun run check green (53 tests / 7 files) — subagent reducer + a
dashboard frame test (seeded tree renders, transcript replaced) + /agents
dispatch. LIVE tmux: /agents opened empty; then a REAL delegation spawned a
subagent → /agents showed "⛓ Agents · 1 subagent · ● completed <goal>
(model) ⚡terminal". ALL 7 first-class surfaces are now ✅+tested+smoked
(blocking prompts, pager, session switcher, model picker, skills hub,
completions, agents dashboard). Smoke P5e + matrix updated. Remaining: chrome
(5b), agent-feature polish (5d), launcher (8).
A live slash-completion dropdown renders above the composer as you type `/…`
(spec §1 autocomplete) — the 6th and final first-class overlay surface.
- view/composer.tsx: onContentChange → onType (reads ta.plainText); a dropdown
of candidates (display + meta) renders above the textarea when completions are
set. The textarea owns key input (live refine-by-typing), so Tab accepts the
top match (ta.clear()+insertText) and Esc dismisses; arrow-nav would fight the
cursor (noted polish).
- store: completions state + setCompletions/clearCompletions; CompletionItem.
- logic/slash.ts: mapCompletions(complete.slash result) → candidates.
- entry: onType queries complete.slash for `/word` (no space) and sets/clears the
store completions; cleared on submit / non-slash / space.
Verified: bun run check green (49 tests / 7 files) — mapCompletions + a
composer-dropdown frame test. LIVE tmux: typing `/comp` showed /compress,
/composio, /compact (with descriptions); Tab accepted the top + cleared the
dropdown. ALL 6 first-class overlays are now ✅+tested+smoked (blocking prompts,
pager, session switcher, model picker, skills hub, completions). Smoke P5a +
matrix updated. Remaining: chrome (5b), agent features (5d), agents dashboard (5e).
A full-height scrollable pager (the FloatBox analog) — porting it unlocks the
long-output slash commands (/status /logs /history /tools) at once (spec §2b).
- view/overlays/pager.tsx: bordered full-height overlay (title + scrollbox +
footer), scrolling driven explicitly via useKeyboard → scrollBy/scrollTo (no
reliance on scrollbox auto-focus), Esc/q/Ctrl+C close. §8 #2 scrollbox gotchas.
- store: pager state + openPager/closePager.
- view/App.tsx: content zone swaps to the Pager (replacing transcript+composer)
when store.state.pager is set; the close is deferred a tick so the closing key
can't leak into the remounting composer.
- logic/slash.ts: present() routes output to the pager when long (>180 chars or
>2 non-empty lines, Ink parity) else a system line; titled by command; /logs
always pages. New openPager on SlashContext.
Verified: bun run check green (41 tests / 7 files) — present() routing
(short→system, long→pager) + a pager frame test (renders title/content, replaces
the transcript/composer). LIVE tmux: /logs → pager (title "Logs", scroll via
PageDown, Esc closed → composer refocused, no key-leak); /version (5-line output)
→ pager titled "Version". Smoke P5a + parity matrix updated. Completions dropdown
+ pickers + chrome are the next slices.
HERMES_TUI_RESUME=<id|recent> resumes a session instead of creating one:
session.most_recent (for "recent") → session.resume {cols, session_id} →
commitSnapshot(mapResumeHistory(messages)), buffering live events across the RPC.
- logic/resume.ts: maps the session.resume history into Message[]. Resumed tool
rows arrive as {role:'tool', name, context} (NO text — gotcha §8 #5); they're
FOLDED into the preceding assistant turn's ordered parts (state:'complete',
summary=context) so a resumed transcript renders the tools INLINE like a live
one. Assistant text gets a text part (renders via native markdown). User/system
stay flat. Unknown roles / non-arrays are ignored.
- logic/store.ts: hydrate split into beginBuffer() + commitSnapshot() so the live
event buffer spans the async resume RPC (events that arrive during resume are
replayed after the snapshot, in order).
- entry/main.tsx: bootstrap branches create vs resume; the resume path is timed
(rpc_ms / hydrate_ms) for profiling.
Verified: bun run check green (40 tests / 7 files) — resume mapper (fold tool
rows, standalone holder, ignore junk) + beginBuffer/commitSnapshot replay. LIVE
tmux: Launch A created a session with a ⚡terminal tool call; Launch B
(HERMES_TUI_RESUME=recent) hydrated user + assistant + the tool row inline.
STRESS+PROFILE on a real 103-message session (~/.hermes/sessions): client hydrate
= 76ms, bun RSS = 214MB STABLE (no leak), tool rows hydrated, PageUp scroll works;
the 1.6s cost is the server-side session.resume RPC, not the TUI. Smoke P4 +
matrix updated. Note: rows instantiate for the full history (scrollbox culls
render only) → RSS ~linear in turns; list virtualization is the lever if
multi-thousand-turn sessions become a target.
The composer now routes `/command` through the Ink-parity dispatch ladder
instead of submitting it as a prompt (spec §1):
- logic/slash.ts: parseSlash + dispatchSlash — client-local command →
slash.exec {command, session_id} (output → system line) → on reject
command.dispatch {arg, name, session_id} with typed handling
(exec/plugin→system · alias→re-dispatch · skill/send→submit a turn ·
prefill→notice). 6 client commands: help/quit/exit/clear/new/logs.
- /help renders the live `commands.catalog` (reads the `pairs` shape).
- view/prompts/confirmPrompt.tsx + store.setConfirm: a LOCAL (non-gateway) Y/N
dialog for /clear and /new; store gains pushSystem + clearTranscript.
- entry: a Promise-returning `request` adapter + the SlashContext wiring (quit →
renderer.destroy, confirm, clearTranscript, logTail, submit).
Also fixes a keystroke-leak: the key that ANSWERED a prompt was bleeding into the
freshly-refocused composer (`/clear`→y left "y" in the input, breaking the next
`/quit`). PromptOverlay now defers the prompt-clear (composer remount) past the
current keystroke — this hardens every Phase 3 prompt too.
Verified: bun run check green (36 tests / 6 files) — slash.test covers parse + the
full ladder against a fake context. LIVE tmux: /help → full gateway catalog;
/version → slash.exec output; /clear → confirm → cleared, no key-leak (typed "hi"
not "yhi"); /quit → clean quit, child reaped. Remaining TUI-only commands,
completions, pager routing, and session resume are 4b/4c. Smoke P4 + matrix updated.
The 4 gateway *.request events now drive a blocking-prompt overlay instead of
deadlocking the agent (spec §8 #6). Native OpenTUI paradigm (per glitch's steer):
- view/prompts/approvalPrompt.tsx: native <select> (once/session/always/deny)
→ approval.respond {choice, session_id}.
- view/prompts/clarifyPrompt.tsx: native <select> over choices + an "✎ Other…"
option that swaps to a native <input> for free-text → clarify.respond
{answer, request_id}.
- view/prompts/maskedPrompt.tsx: sudo (🔐) / secret (🔑) — native <input> has no
mask, so we own a buffer via useKeyboard and render '*' per char →
sudo/secret.respond {password|value, request_id}.
- view/prompts/promptOverlay.tsx: dispatches by prompt kind, binds each
answer/cancel to the matching *.respond; Esc/Ctrl+C → deny/empty so the agent
always unblocks.
Wiring: store gains ActivePrompt state + the 4 reducer cases + clearPrompt;
App swaps Composer↔PromptOverlay on store.state.prompt (so the composer textarea
stops capturing keys while blocked); renderer.ts gates the global Ctrl+C-quit on
isBlocked() so a prompt owns Ctrl+C (→ cancel); entry adds a generic `respond`
runFork callback + passes sessionId.
Verified: bun run check green (28 tests / 5 files) — reducer set/clear for all 4,
+ a frame test (approval overlay renders the command + all options as a bordered
modal, composer hidden while blocked). LIVE tmux: a real `rm -rf` approval fired;
Approve-once → command ran → unblocked; Esc → deny → "BLOCKED by user" →
unblocked; Ctrl+C-while-blocked cancelled WITHOUT quitting; Ctrl+C-unblocked quit
clean, no orphan. Smoke P3 + parity matrix updated. confirm (local) → Phase 4.
Assistant text parts now render through the NATIVE markdown renderable instead of
plain spans — bold/headings/lists/fences render, raw `**`/backtick markup is
concealed (spec §7; never hand-roll a parser).
- view/markdown.tsx: `<code filetype="markdown" streaming conceal drawUnstyledText>`
(CodeRenderable — opencode's v2 AssistantText path; `<markdown>` +
internalBlockMode="top-level" deferred paint headlessly). SyntaxStyle.fromStyles
is derived from the theme (markup.* → theme.color.*, non-hex colors guarded) and
cached by theme-object identity so all text parts share one instance, rebuilt
only on skin change. drawUnstyledText paints raw text immediately while
Tree-sitter highlighting settles (and makes it headless-capturable).
- view/messageLine.tsx: text-part Match renders <Markdown> instead of <text>.
- test/lib/render.ts: settle async markdown via flush(); captureFrame gains an
`until` option (waitForFrame) for content that paints after the first pass.
Verified: bun run check green (23 tests / 5 files). Live tmux: a markdown reply
(heading + bold word + 2-item list) rendered with `**` concealed (grep -c '**' = 0);
Ctrl+C clean, no orphan. Phase 2 complete (2a shell + 2b-i parts/tools + 2b-ii
markdown) — smoke steps 1–4 run live. Next: Phase 3 blocking prompts.
An assistant turn is now ONE ordered parts[] (text/reasoning/tool) instead of a
flat string, so tool calls render INLINE between text blocks rather than dumped
as separate rows below (spec §7 — the "dump-below" bug opencode's sync-v2 avoids).
- logic/store.ts: Part discriminated union + reducer rework. message.delta
appends to the open text part (or opens one); tool.start pushes a running tool
part; tool.complete matches by tool_id and updates that part IN PLACE (state,
envelope-stripped resultText, summary, error, lineCount); reasoning.delta
accumulates a reasoning part. User/system rows stay flat text; settled/resumed
assistant rows fall back to text.
- logic/toolOutput.ts: ported pure helpers — stripToolEnvelope (unwrap
{output,exit_code}, append [exit N]/[error] suffix) + collapseToolOutput +
truncate.
- view/messageLine.tsx: <For>+<Switch> dispatch by part.type with stable id keys.
- view/toolPart.tsx: two-tier render — inline one-liner (≤1 output line) or a
capped left-bar block (TOOL_MAX_LINES, "… +N more", click-to-expand) keyed off
the theme; reactive width via useTerminalDimensions.
Verified: bun run check green (23 tests / 5 files / 64 expects) — store
interleave/in-place/reasoning, a frame test asserting the tool renders inline +
envelope stripped, and toolOutput unit tests. Live tmux: a terminal-tool prompt
rendered "⚡ terminal" with its alpha/beta output inline between the assistant's
text parts; Ctrl+C clean, no orphan. Smoke P2b + parity matrix updated. Native
<markdown> for text parts is the next slice (2b-ii).
Turns the read-only Phase-1 view into an interactive shell, split into focused
view components (spec v4 §2 layout):
- view/transcript.tsx: ONE full-height <scrollbox> with a reactive <For>
(opencode's no-scrollback model). Applies the §8 #2 gotchas exactly:
minHeight:0 on the wrapper AND the scrollbox, NO flexDirection on the
scrollbox root, stickyScroll + stickyStart="bottom".
- view/composer.tsx: a native <textarea> captured by ref — flexShrink:0,
focus-on-mount, Enter->submit via keyBindings, imperative .clear() on submit,
and a `submitting` re-entrancy guard. Wired by the entry to fire prompt.submit
(Effect.runFork on the in-hand service value); it's now the PRIMARY input, with
the HERMES_TUI_PROMPT stand-in kept only for launch-with-prompt.
- view/header.tsx + view/messageLine.tsx: extracted, themed (no hardcoded
styles). MessageLine stays flat-text this slice; ordered parts (§7) land in 2b.
test/lib/render.ts now flushes 3 renderOnce passes before capture — a <scrollbox>
needs more than one pass to measure content + apply sticky, else the transcript
row paints blank.
Verified: bun run check green (12 tests / 4 files / 31 expects). Live tmux drive:
typed into the composer -> cleared -> user row -> streamed reply ("Here are three
words"); Ctrl+C quits cleanly even with the textarea focused, no orphan child.
Composer placeholder rendered the live skin's welcome string (skin->theme live).
Smoke P2a + parity matrix updated. Phase 2b (ordered parts/tool render/markdown)
is the next slice.
GatewayService/liveGateway over the real Python tui_gateway: JSON-RPC stdio
framing (Bun.spawn), 16ms event coalescing flushed inside Solid batch(), typed
GatewayError, and a decode-once GatewayEvent Schema (~35-member tagged union;
unknown/malformed events skip via Option.none, never crash the stream).
The Solid sync-v2-style store grows to: streaming text concat (prefer
payload.text), gateway.ready{skin}/skin.changed -> fromSkin reactive re-theme,
LRU id-dedup, and hydrate-while-buffering (resume scaffold). Theming is a 1:1
port of Ink's theme.ts (DARK/LIGHT, detectLightMode, ANSI-256 normalization,
fromSkin) behind a Solid ThemeProvider so existing skins work unchanged and the
view carries NO hardcoded styles. A console-safe diagnostics log (in-memory
ring + NDJSON file) is the single logging path.
Entry gains a live launch path (default; HERMES_TUI_FAKE=1 -> scripted hello)
with an initial-prompt bootstrap (session.create -> prompt.submit) as the
Phase-2-composer stand-in, plus a minimal Ctrl+C graceful quit
(renderer.destroy -> shutdown Deferred -> scope finalizers -> client.stop) so
the engine reaps its own gateway child instead of orphaning it.
Verified: bun run check green (tsc + eslint + 12 tests / 4 files); live tmux
drive connect -> gateway.ready -> prompt -> streamed reply ("pong") -> clean
teardown with no orphan bun/python. Parity matrix + smoke P1 run log updated.
hermes --tui launches the native OpenTUI engine (Bun) when
HERMES_TUI_ENGINE=opentui (env) or display.tui_engine=opentui (config);
Ink stays the default and the shipping path is untouched.
- _resolve_tui_engine() (env > config > ink); refuses opentui on
Windows/Termux (no Bun) -> falls back to ink with a notice.
- _make_opentui_argv() -> [bun, src/entry.real.tsx] (no build step).
- _bun_bin() with HERMES_BUN override.
- Branch at top of _make_tui_argv BEFORE _ensure_tui_node (Bun-only host
must not bootstrap Node).
- Gate _launch_tui NODE_OPTIONS/--max-old-space-size on engine==ink (Bun
is JSC; the V8 flag errors/ignores).
Verified end-to-end via tmux: real hermes --tui -> Bun -> OpenTUI ->
real Python gateway streamed a real reply. No-flag default still ink.
Completes the worktree-misroute fix from #35399, which made misroutes
visible (resolved_path) but did not prevent them: its divergence warning
only fired once a terminal command had populated the live cwd registry.
A fresh worktree session (registry still empty) with a stale TERMINAL_CWD='.'
got neither a worktree anchor nor a warning, so a relative write_file/patch
silently landed in the MAIN checkout.
Two changes in tools/file_tools.py:
- Treat sentinel TERMINAL_CWD values ('', '.', './', 'auto', 'cwd') and any
relative value as UNSET rather than a literal anchor. Previously '.' was
joined onto the process cwd, silently routing edits to wherever the process
happened to be (the main repo, in a worktree session). The gateway already
sanitizes the same set at import time; the file-tool layer now matches.
- New _authoritative_workspace_root(): prefers the live terminal cwd, else a
sentinel-free absolute TERMINAL_CWD (the worktree path cli.py/main.py set
for -w). _resolve_base_dir() and _path_resolution_warning() both use it, so
a worktree session resolves into — and warns about escaping — the worktree
from the very first write, before any cd has run.
Validation: 11 new/parametrized tests (sentinel handling, empty-registry
anchoring, early divergence warning, live-cwd precedence). 32/32 pass under
scripts/run_tests.sh. Live E2E: relative write in an empty-registry worktree
session lands in the worktree, main untouched.
When --replace force-kills an unresponsive old gateway, SIGKILL can fail
to reap it (uninterruptible sleep, zombie-reaping parent, etc.). The old
code unconditionally cleared the PID file and scoped locks and started a
fresh instance anyway, leaving two live gateways fighting over the same
bot token — a duplicate-gateway failure mode of #19471.
Re-verify the process is actually gone (via the Windows-safe _pid_exists
helper) after the force-kill; if it still appears alive, clear the
takeover marker and abort the replacement instead of duplicating.
Co-authored-by: Hermes <noreply@nousresearch.com>
PR #41822 collapsed CWD-only overrides to the shared 'default' container
via _resolve_container_task_id, but three call sites kept routing the
*env/override lookup* through that collapsed id:
- the foreground exec path read _task_env_overrides[effective_task_id],
yet register_task_env_overrides writes under the raw task_id, so a
CWD-only override's cwd was silently dropped (env spun up at the wrong
root, exit 126);
- the get-or-create env lookup keyed solely on effective_task_id, so an
env cached under the raw task_id was missed and duplicated;
- register_task_env_overrides synced the new cwd onto the env under the
collapsed id, missing a live env cached under the raw task_id.
Container *identity* still collapses to 'default' (sharing preserved);
only the per-session env/override *lookup* now prefers the raw task_id and
falls back to the collapsed id. Fixes the 3 regressions in
test_terminal_task_cwd.py left red by #41822.
eslint --fix (import sort + padding-line-between-statements) on sidebar/index.tsx
after cherry-picking @dangelo352's commits; add release.py AUTHOR_MAP entry so
CI doesn't block on the unmapped author email.
gateway/run.py is the largest god file (20k LOC, GatewayRunner with 220
methods). This lifts the cohesive kanban-watcher cluster — _kanban_notifier_watcher,
_kanban_dispatcher_watcher, _kanban_advance/unsub/rewind, _deliver_kanban_artifacts
(~1,035 LOC, 6 methods) — into gateway/kanban_watchers.py as a mixin that
GatewayRunner inherits.
Mixin (not free functions) because the methods use only self state: inheriting
keeps every self._kanban_* call site working unchanged via the MRO, making this
a behavior-neutral move. The methods' lazy imports (_kb, _decomp, _load_config,
Platform) travel with them; the mixin needs only stdlib + a matching
logging.getLogger('gateway.run').
run.py 20187 -> 19157 LOC; GatewayRunner direct methods 220 -> 214.
Behavior-neutral: gateway test suite 6582 passed / 0 failed; start() still wires
both watchers via self._kanban_*; MRO resolves all 6 to the mixin. One test
(corrupt-board quarantine retry) keyed its time-travel mock on the caller's
filename being gateway/run.py — updated to also accept gateway/kanban_watchers.py.
Establishes the mixin-extraction pattern for further GatewayRunner decomposition
(the 2406-LOC _run_agent and 1164-LOC _handle_message remain, but their callback
closures need a context-object redesign — deferred).
When register_task_env_overrides is called with only a 'cwd' key
(ACP adapter workspace tracking), the task_id should collapse to
'default' so all interactive surfaces (TUI, gateway, dashboard)
share one long-lived container.
Previously, any override registration — even CWD-only — caused
_resolve_container_task_id to return the session key unchanged,
spinning up a separate container per session. This made it
impossible to authenticate into external services once and have
that auth available across all surfaces.
Now only overrides containing isolation keys (docker_image,
modal_image, singularity_image, daytona_image, env_type) trigger
per-task container isolation.
Fixes#37361
Subcommands whose handler was a closure defined inside main() — memory, acp,
tools, insights, skills, pairing, plugins, mcp, claw — have their handler
promoted to a top-level function and their parser block extracted into
hermes_cli/subcommands/<name>.py (build_<name>_parser, injected handler).
These 9 had zero closure-over-main-locals, so promotion is a pure relocation.
acp/mcp parser blocks use the shared add_accept_hooks_flag helper.
main() 1798 -> 954 LOC (71% below the 3297 Phase-2 starting point);
add_parser calls in main.py 89 -> 28.
Deferred: sessions, computer-use, secrets handlers reference <name>_parser
(for a no-subcommand print_help fallback) — left in place to avoid the
_self_parser indirection; minority, low value.
Behavior-neutral: all 9 subcommands' --help (incl nested subactions) byte-
identical to pre-extraction (diff-verified). tests/hermes_cli/ 6519 passed /
0 failed; new test_subcommands_followup.py covers the 9 builders.
run_conversation's inner retry loop tracked recovery state in ~15 scattered
bare booleans (per-provider OAuth refresh guards, format-recovery guards,
restart signals). They are now fields on a single TurnRetryState dataclass the
loop mutates in place (_retry.<flag>), giving the recovery bookkeeping a named,
testable home.
Loop-control vars (retry_count, max_retries, max_compression_attempts) stay as
plain locals — they're while-mechanics, not recovery bookkeeping.
Behavior-neutral: pure local→attribute rewrite of 42 references; kwarg NAMES
preserved (e.g. has_retried_429=_retry.has_retried_429). Live simple + tool
turns OK.
Validation: tests/run_agent/ 1615 passed / 0 failed under per-file process
isolation; new test_turn_retry_state.py pins the field contract.
When context compaction rotates agent.session_id, it updates the gateway/tools
session context (set_current_session_id -> HERMES_SESSION_ID env + ContextVar)
but never updates the separate logging session context. The [session_id] tag on
log lines comes from hermes_logging._session_context (set once per turn in
conversation_loop.py), so post-compaction log lines in the same turn carry the
STALE old id while the message/DB/gateway state carry the new one — breaking log
correlation exactly at the compaction boundary.
Call hermes_logging.set_session_context(agent.session_id) alongside the existing
set_current_session_id, guarded so a logging failure can't regress the routing
update. Logs-only; no runtime or caching impact.
Refs #34089
The curator's idle-archival path (apply_automatic_transitions under
prune_builtins) could archive the bundled `plan` skill, killing the
/plan slash command silently — typing /plan then returned 'Unknown
command' with no signal that a skill had vanished. The archived skill's
hash stays in .bundled_manifest, so 'hermes update' wouldn't re-seed it.
Add PROTECTED_BUILTIN_SKILLS ({plan}) enforced at the master gate
is_curation_eligible() (covers archive_skill + the transition walk) and
in the candidate enumerator (so the LLM consolidation pass never sees
them). Immune to prune_builtins, pin state, and LLM judgment.
Closes#33617. Adds additive _meta.hermes.sessionProvenance to ACP session
surfaces so clients can detect compression-driven internal session rotation
without parsing status text, guessing from token drops, or reading state.db.
Derived on demand from the existing compression chain (parent_session_id /
end_reason) — no new persisted state, no schema change, no ACP protocol change.
ACP session_id stays the stable client handle.
- acp_adapter/provenance.py: derive provenance from SessionDB
- server.py: attach _meta to new/load/resume responses; emit a
session_info_update when the internal head rotates during a prompt
Salvage follow-up for PR #33221 — the cherry-picked commit is authored
under martin.alca@gmail.com (not the draixagent@gmail.com already mapped),
which would fail the CI author-attribution gate.
VolcEngine's api/plan endpoint occasionally leaks raw XML attribute
fragments into tool_use.name when its protocol-translation layer
converts the model's native XML-style tool emission to Anthropic
Messages tool_use blocks, producing names like:
terminal" parameter="command" string="true
execute_code" parameter="code" string="true
session_search" parameter="session_id" string="true
The corruption happens server-side at the provider, but it breaks
every tool call for affected users — no normalization rule in
repair_tool_call can rescue them, so each request runs through three
retries and then aborts as partial.
Add an early sanitizer in agent_runtime_helpers.repair_tool_call that
trims at the first ' " ', " ' ", '<', or '>' character (idx > 0
only) so the rest of the existing repair pipeline (lowercase /
snake_case / fuzzy match) can resolve the cleaned name normally.
Whitespace is deliberately NOT a separator — the legitimate
"write file" -> write_file repair path (covered by
test_space_to_underscore) must keep working.
Tests: 11 new regression cases in TestVolcEngineXmlPollution
covering all three observed polluted names, CamelCase + pollution
mix, single-quote variants, angle-bracket variants, clean-name
passthrough, and the whitespace-preservation guard. All 18 pre-
existing repair tests still pass (29 total in the file).
Replace the ACP-local prefix/suffix matcher + helper with a single
startswith() check against INTERRUPT_WAITING_FOR_MODEL_PREFIX, now
defined once in conversation_loop.py where the sentinel is produced.
Keeps the source of truth in one place so the guard cannot drift if
the status string changes. Net -17 LOC in server.py.
Also add lsaether to release.py AUTHOR_MAP.
The auxiliary Codex adapter maintained its own chat->Responses conversion
loop that forwarded every non-system message's role verbatim into
Responses input[]. When flush_memories()/compression replayed session
history containing assistant tool_calls + role=tool results, those tool
messages leaked into the request and the Responses API rejected them with
HTTP 400: Invalid value: 'tool'.
Route _CodexCompletionsAdapter.create() through the same shared converter
the main agent transport uses (_chat_messages_to_responses_input), so tool
calls become function_call items and tool results become function_call_output
items with a valid call_id. Single conversion path means no future drift.
Also remove the now-dead _convert_content_for_responses() helper — its only
caller was the private conversion loop this change deletes.
Co-authored-by: ProgramCaiCai <techxacm@gmail.com>
Batch extraction of every remaining subcommand whose handler is top-level and
whose parser block is pure argparse: model, setup, postinstall, whatsapp, slack,
login, logout, auth, status, webhook, hooks, doctor, security, dump, debug,
backup, import, config, version, update, uninstall, dashboard, gui, logs,
prompt-size.
Each becomes hermes_cli/subcommands/<name>.py with build_<name>_parser() and an
injected handler (no main import). dashboard also injects cmd_dashboard_register
for its nested 'register' action.
Behavior-neutral: all 25 subcommands' --help output (and nested subaction help)
diff-verified byte-identical to pre-extraction. Two RawDescriptionHelpFormatter
epilogs (debug, logs) needed their multi-line string interiors preserved at
column 0 — caught by the --help diff, not compile.
main() 3297 -> 1798 LOC across this PR; add_parser calls in main.py 179 -> 89.
Validation: tests/hermes_cli/ 6476 passed / 0 failed under per-file process
isolation; new test_subcommands_batch.py smoke-tests all 25 builders + the
dashboard two-handler case.
Follow-on to the cron extraction in the same Phase 2 PR. Same pattern:
per-group build_<name>_parser() functions with injected handlers, no main
import.
- subcommands/profile.py: build_profile_parser (190-line block out of main()).
- subcommands/gateway.py: build_gateway_parser (gateway + proxy, 238-line block;
they shared one inline section). Imports argparse for SUPPRESS defaults.
- main(): two more inline blocks become single builder calls.
Behavior-neutral: 'profile [sub] --help' and 'gateway/proxy [sub] --help'
byte-identical to pre-extraction (diff-verified).
main() now 2723 LOC (was 3297 at Phase 2 start); add_parser calls in main.py
179 -> 141.
Validation: tests/hermes_cli/ 6476 passed / 0 failed under per-file process
isolation; new builder unit tests cover subactions, aliases, dispatch, flags.
Phase 2 of the god-file decomposition plan. main()'s argparse tree is 179
inline add_parser calls in one 3,297-line function. This establishes the
hermes_cli/subcommands/ package and extracts the first group (cron) as the
proof-of-pattern:
- hermes_cli/subcommands/_shared.py: shared parser helpers (add_accept_hooks_flag),
re-exported from main.py for backwards compat.
- hermes_cli/subcommands/cron.py: build_cron_parser(subparsers, cmd_cron=...).
Handler injected so the module never imports main (cycle avoidance).
- main()'s ~155-line inline cron block becomes one build_cron_parser() call.
Behavior-neutral: 'hermes cron create --help' output is byte-identical to
origin/main. main() 3297 -> 3143 LOC.
Validation: tests/hermes_cli/ 6466 passed / 0 failed under per-file process
isolation; new test_subcommands_cron.py covers subactions, aliases, options,
no-agent tristate, injected dispatch, and --accept-hooks.
Phase 1 of the god-file decomposition plan. run_conversation's ~470-line
once-per-turn setup block (stdio guarding, retry-counter resets, user-message
sanitization, todo/nudge hydration, system-prompt restore-or-build,
crash-resilience persistence, preflight compression, the pre_llm_call hook, and
external-memory prefetch) is moved verbatim into build_turn_context(), which
returns a TurnContext dataclass the loop unpacks.
Behavior-neutral move-and-name refactor: the builder mutates `agent` exactly as
the inline code did; only the locals the loop reads back are returned.
- run_conversation: 4602 -> 4217 LOC (-385)
- agent/conversation_loop.py: 4965 -> ~4580 LOC
- new agent/turn_context.py: focused, dependency-injected, unit-tested in isolation
Tests: tests/run_agent/ 1570 passed / 0 failed under per-file process isolation.
Relocation follow-ups: 413_compression mocks now patch both module references;
nudge/on_turn_start source-inspection guards point at the extracted module.
