code_exec: agent-controlled timeout (30s default, 300s max) + language metadata ...

The 30s timeout was a hard constant; long computations or test suites hit it
with no escape hatch. Expose it as a `timeout` param (clamped 1-300s, default
30) like terminal's. Also stamp `language: "python"` into metadata so the
TUI renderer can highlight without hard-coding, and record `timeout` on the
timeout result for a dedicated timeout status card.

tui: lower default visible chat window 200->60 (faster long-session resume) ...

The cold cost is dominated by the one-shot mount of the whole visible
window at load/switch time. 200 _ChatItemView widgets each parse Markdown,
build Content, compute height and run layout in one pass. 60 cuts that
peak ~3.3x with no loss for streaming (signature cache + throttle + render
cache keep per-token cost O(1) once mounted). Model cap (max_visible_items
* 3) follows automatically: 600 -> 180. Older items collapse into the
existing "... N earlier messages not shown" hint.

compress: raise budget 6000 + richer prompt detail + input headroom ...

Increase the compression summary budget and push the prompt toward more
detail and completeness, so the agent keeps more durable facts when old turns
are folded into a summary.

Budget (output):
- context_summary_max_tokens 4000 → 6000 (global default).
- navi_code compression_max_tokens 4000 → 6000 (profile override).

Prompt (base, all profiles):
- New 'Intermediate Findings' section: durable facts from read/grep/log/terminal
  that informed decisions (was forbidden by 'do not preserve intermediate
  reasoning'). Existing sections get richer instructions (Active Files: key
  symbol affected; Decisions: rationale + trigger; Completed: step + file +
  verification outcome; Errors: snippet + what was tried + outcome).
- Loosen 'tight bullet points' → 'specific and complete; prefer listing over
  generalizing — losing a fact costs more than a longer summary'.
- 'Do not write implementation code' → 'preserve short critical code verbatim
  (final signatures, one-line fixes, key config lines); drop long blocks'.

Input headroom (so the summarizer can SEE what it must preserve):
- Non-critical tool result preview 300 → 800 chars.
- _MAX_SUMMARY_INPUT_CHARS 24000 → 32000 (fits 65k–128k windows with 6k output).

Profile prompt (navi_code/compression_prompt.txt): preserve short diagnostic
snippets verbatim (exact error line, failing assertion, one-line fix) instead
of blanket 'do not preserve long output'; keep code signatures + short key
snippets, not just names.

Tests: Intermediate Findings section in base prompt; 'short critical code'
phrasing; non-critical preview keeps 800 chars; meta-summary threshold raised
(2×6000 chars). Fixed flaky test_attach_session_seeds_context_fill (relied on
_startup worker timing after Etap 4 lengthened attach; now explicit attach
like its sibling + mock get_todos to skip a 30s×2 real-network timeout).
Full suite: 1008 passed, 1 skipped.

todo: add op — append steps discovered mid-task (preserve statuses) ...

The todo tool only had set/view/update/clear. Adding a step that surfaced
mid-task meant either 'set' (which replaces the plan and resets every status
to pending) or 'replan' (1–3 LLM calls) — both heavy for a single new step, so
the agent usually skipped recording it. Add a cheap 'add' op:

- navi/tools/todo.py: op 'add' (tasks: list[str]) appends new _Task entries to
  the end of the existing plan in pending status; existing steps and their
  statuses are untouched. Requires an existing plan (points to 'set'
  otherwise). Schema enum + op description + tool description updated.
- system_prompt.txt: tell the agent to record new subtasks with 'todo add'
  right away (not held in head, not a replan); clarify the 'small todo edit'
  line — add via 'todo add', drop/merge/reorder via 'set' (re-apply statuses).
- renderers/todo.py: TodoStartedRenderer renders the add card
  ('→ todo · add (N)') listing the new pending steps, not a JSON dump.

Tests: add appends + preserves statuses / requires existing plan / requires
tasks; renderer add card + empty-tasks. Full suite: 1005 passed, 1 skipped.

tui: terminals picker — clearer two-line rows + Delete to close ...

- Each terminal row is now two lines: the name (bold, with status icon) on
  top, the description + pid + uptime as a dim meta line beneath, separated
  by '·' — easier to scan than a single run-on line.
- Close is bound to Delete/Backspace instead of Enter: Enter commonly means
  "open/select", Delete means "remove/close", which matches closing a
  terminal. Title updated: "Open terminals — Delete to close, Esc to cancel".
  No-op on an empty list.

Tests: close-on-delete (was close-on-enter), empty-list delete is a no-op.
Full suite: 1000 passed, 1 skipped.

tui: terminal tool_started renderer (action-aware card) ...

terminal tool_started cards used to fall through to the generic
ToolStartedRenderer (a raw JSON dump of the args — a wall of quoted strings for
a command or input). Add an action-aware started renderer, mirroring the result
renderer and the filesystem started renderer:

- Title: → terminal <action>.
- run: the command as a shell prompt in accent (the headline) + optional cwd.
- open: terminal name (accent) + description + background + command preview.
- close/status: terminal name. send_input: name + input preview.
- list: empty body (no per-call headline; the result card carries the table).

Registered before ToolStartedRenderer (first accepting wins).

Tests: started accepts gating + run/open/close/send_input/list cards;
render_plain covers the borderless reading-mode path. Full suite: 1000 passed,
1 skipped.

tui: /terminals command + modal to view and force-close terminals (Etap 5) ...

The final piece of the terminals plan: a modal to list the session's open
persistent terminals and force-close one.

- screens/terminals_picker.py: a modal (mirror of sessions_picker). Seeds the
  list from api.list_terminals on open, kept live by refresh_from_model — the
  app calls it as terminal_opened/closed events arrive while the modal is open
  (closed ones drop out, newly opened append). Each row shows status emoji +
  name (bold) + description + pid + uptime. Up/Down navigate, Enter force-closes
  the highlighted terminal via api.close_terminal (in a worker), Escape cancels.
  Empty list is a no-op for Enter (no crash, no close call).
- commands/builtin.py: TerminalsCommand (/terminals) opens the picker for the
  active session. Registered in the command registry.
- tui_app._open_terminals_picker pushes the screen; on_ws_event's terminal
  branch now also refreshes the picker live if it is the active screen.

Tests: list seed, Enter closes the highlighted terminal (api.close_terminal
called, drops from list), Escape cancels (dismiss None), live refresh filters
closed + adds opened, empty list Enter is a no-op. Input-box hint tests
adjusted for the new /t match order (terminals, thinking, themes). Full suite:
993 passed, 1 skipped.

Plan docs/terminal_tool_plan.md — Etap 5 of 5 (complete).

tui: handle terminal lifecycle events + status-bar count (Etap 4) ...

terminal_opened/output/closed WS events used to fall through to ChatModel's
unknown-event path (raw dict dumped into chat). Now they feed a dedicated
terminals state and a status-bar count, with no chat bubble:

- ChatModel: terminals dict (name → {description, pid, background, status,
  output_tail, closed}) updated by terminal_opened/output/closed in
  handle_ws_event (return None — no chat item). open_terminal_count property;
  seed_terminals(summaries) populates from a REST list (on attach). output_tail
  capped at 200 lines.
- tui_app.on_ws_event: a terminal-events branch forwards to the model (no _sync
  walk — background output can be frequent) and updates the StatusPanel count.
  attach_session seeds via api.list_terminals and sets the count (graceful 0 on
  failure).
- StatusPanel: the "Ctrl+P palette | /help commands" hint line is replaced with
  a live "Terminals: N" count (set_terminals_count). Key combos still live in
  the bottom StatusBar.
- Tests: mock_tui_api fixtures across 5 client test files gain an async
  fake_list_terminals so attach_session never hits a real network in tests.

Tests: ChatModel opened tracks state (no bubble) / output appends tail / unknown
ignored / closed decrements count / seed from REST; tui_app terminal events
update the status count with no chat bubble. Full suite: 988 passed, 1 skipped.

Plan: docs/terminal_tool_plan.md — Etap 4 of 5.

tui: terminal tool-call renderer (action-aware) (Etap 3) ...

terminal tool_call results used to fall through to the generic
ToolResultRenderer (flat Text, no awareness of the action). Add an
action-aware renderer registered before the generic one, mirroring filesystem:

- run: command echoed as a shell prompt + dim output + an exit-code anchor
  (green on 0, red otherwise). A failed run still shows what ran and the exit
  code, not just the raw error.
- open: terminal name (accent) + description + background flag + pid.
- list / status: dim, truncated (200-line cap like the generic tool renderer).
- send_input / close: dim echo. open failure surfaces the reason (already
  exists / max reached). Other action failures show the message in red.

Registered before ToolResultRenderer in default_registry (first accepting
renderer wins). ChatPanel._item_msg already spreads the WS payload into meta,
so the renderer reads args/result/success/metadata directly.

Tests: accepts gating, run success/failure with exit code, open name/desc/pid,
list/status/send_input/close, run truncation; render_plain covers the
borderless reading-mode path. Full suite: 982 passed, 1 skipped.

Plan: docs/terminal_tool_plan.md — Etap 3 of 5.

terminal: REST list/close endpoints + async client helpers (Etap 2) ...

The /terminals modal and the status-bar count need a server source for the
session's persistent terminals (background dev servers, etc.) and a way to
force-close one. Add REST endpoints over the existing TerminalManager:

- GET /sessions/{id}/terminals → terminal_manager.list(session_id) →
  {session_id, terminals: [summary,...]}. Access-checked (navi.sessions.read_all).
- POST /sessions/{id}/terminals/{name}/close → terminal_manager.close(...)
  → {session_id, terminal_name, closed: bool}. Name is URL-encoded by the
  client (terminal names may contain spaces).
- clients/terminal/api.py (async): list_terminals(session_id),
  close_terminal(session_id, name) — for the TUI modal + status-bar seed.

Tests: list empty / list returns active / close routes to manager / 404 on
unknown session (fake TerminalManager injected into the container, mirroring the
kv_store injection pattern). Full suite: 972 passed, 1 skipped.

Plan: docs/terminal_tool_plan.md — Etap 2 of 5.

terminal: session-broadcast sink + lifecycle events (Etap 1) ...

Persistent terminals emitted their stream only into a per-tool-call sink, which
closed once the open action returned — so a background dev server's output
stopped reaching the client immediately after open. TerminalClosed was never
emitted at all (the handler existed, the emitter did not).

Replace the per-tool event_sink with a single callback the orchestrator wires to
the session's WebSocket(s):

- TerminalManager: set_event_callback(cb) + _emit(session_id, event). Reader
  tasks and open/close deliver TerminalOutputDelta / TerminalOpened /
  TerminalClosed through the callback, which lives for the terminal's (and
  session's) lifetime — not the tool call's.
- New TerminalOpened event (terminal_opened: name/description/pid/background).
  TerminalClosed now actually emitted from _close_one (explicit/idle/shutdown/
  session_ended). Background output keeps streaming after open returns.
- terminal.py: drop the per-tool event_sink from _tm.open (no longer needed).
- orchestrator._on_terminal_event → _notify_session (per-session WS fan-out,
  same path recall/mcp updates use). container wires the callback after building
  the orchestrator.
- Events imported lazily inside terminal_manager methods to avoid a circular
  import (navi.core.events → navi.core.__init__ → agent → registry → navi.tools).

Tests: open emits TerminalOpened; background output streams via the callback
after open returns (gap-3 fix); close emits TerminalClosed; no callback does not
break readers. Existing terminal tests unchanged. Full suite: 968 passed, 1 skipped.

Plan: docs/terminal_tool_plan.md (5 etaps, this is Etap 1).

tui: bash-style message history (Up/Down recall, per-session, in-memory) ...

On an empty prompt with the hints list closed, Up recalls the previous
submitted message and Down moves toward the most recent; Down past the newest
restores the empty draft. Like a shell history:

- InputBox owns the per-session history (capped at the 10 most recent, with
  consecutive-duplicate suppression) plus the browsing index and saved draft.
- _PromptInput._on_key enters history mode on Up only when the field is empty
  and the hints list is not open; while hints are open Up/Down still navigate
  command matches, and a non-empty field uses TextArea's multiline cursor.
- Only plain user messages are recorded — slash commands and !shell are not
  remembered (they are not messages to the agent). The raw typed text is stored.
- on_user_submitted records a plain submit; attach_session resets the history
  on session switch/resume (the history is in-memory per session, not synced
  to the server, by design).

Tests: append cap/dedup/skip-commands, up/down navigation, reset, Up-recall /
Down-draft / Up-non-empty-no-recall over pilot, submit records plain (not
/commands). Full suite: 964 passed, 1 skipped.

tui: minor correctness/perf fixes (Etap 5) ...

The last batch from the code review — real bugs and edge cases, not cosmetics:

- tui_app._resolve_session: read session_id with .get and return None when a
  malformed server response omits it, instead of KeyError slipping past the
  api-call except (5.B3).
- cli /switch: surface the original error cause (404 vs network vs server) on
  a non-unique prefix match instead of swallowing it into a generic "not found"
  (6.B2).
- input_box _complete_command: a bare "/" with no hint list available no longer
  auto-picks the first command in the registry (3.B3).
- filesystem _render_grep: when the grep ran in regex mode, highlight with the
  actual regex so marked spans match what the server found — a literal highlight
  of a regex pattern would mark nothing (2.B2).
- file_refs _collect_files: filter sensitive entries out in the generator before
  sorted() so a huge directory is not fully materialized only to be mostly
  discarded (5.P1).
- renderers/tool ToolResultRenderer: cap a generic tool result at 200 lines
  (tail + marker) so a non-filesystem tool with huge output does not flood the
  chat bubble — filesystem has its own renderer (2.P3).

Tests: malformed-response /switch error surfacing, bare-slash no-auto-pick,
grep regex highlight, sensitive-subdir skip, tool-result truncation. Full
suite: 953 passed, 1 skipped.

tui: theme repaцвет chat + bounded history + tail purge (Etap 4) ...

- apply_theme now re-renders already-drawn chat bubbles in the new palette:
  ChatPanel.refresh_content() rebuilds each visible item's rich renderable (the
  item renderers read get_active_theme(), already switched) and drops the
  per-widget Content/height caches so the next paint rebuilds instead of
  returning the old-theme cached Content. Theme switch / live preview no longer
  leaves the conversation in the previous palette (5.B1).
- theme_picker live preview is debounced (~100 ms coalesce via set_timer):
  apply_theme now re-renders every bubble, so applying it on every highlight
  while scrolling the theme list would stutter — rapid moves collapse into one
  apply, with the freshest theme winning. Escape/cancel dismiss the pending
  timer so it cannot fire after a restore and re-apply the highlighted theme
  over the restored one (4.P2).
- ChatModel.items is now bounded: a cap (visible window * 3) trims the oldest
  items off the front via a single _append/_trim path, so a long autonomous
  session no longer grows items without limit. The visible window is much
  smaller than the cap, so trimming only drops already-off-screen items and the
  truncation hint still reads correctly (3.P3).
- stream_end purges empty assistant/thinking bubbles off the tail (while-pop)
  instead of a full-history list-comprehension copy — empty bubbles are created
  at the end of the turn, so O(tail) suffices (1.P1).

Tests: apply_theme re-renders bubbles; theme picker debounces + escape cancels;
chat_model cap trims front / cap=None keeps everything; stream_end keeps
non-empty assistant + purges trailing empty thinking. Full suite: 946 passed, 1 skipped.

tui: async REST api client — stop blocking the event loop (Etap 3) ...

The terminal client's REST helpers were synchronous (httpx.Client); every
caller inside the Textual event loop (attach/switch/resume/profile/stop) blocked
the whole TUI for the duration of each network round-trip — the spinner froze,
input stopped responding, the screen did not repaint. Move the network layer to
async and a shared pooled client:

- api.py: all 7 helpers are now `async def` over a single lazily-created
  httpx.AsyncClient (connection pooling) — awaits yield the loop while a request
  is in flight, so the TUI never blocks on the network.
- tui_app.py + commands/builtin.py: every `api.*` call site is now `await`-ed
  (all already ran in async workers / execute). _stop_stream_worker drops the
  dead `iscoroutine` check (api.stop_session is async, the await is real).
- cli.py: _resolve_session_id and _handle_command await api.*; _run_raw wraps
  the resolve + run in asyncio.run(_run_raw_async) so the sync click entrypoint
  stays unchanged. Removed the now-unused _run_one_shot wrapper.
- sessions_picker.on_mount: list_sessions runs in a worker (run_worker) instead
  of blocking on_mount — a slow/unreachable backend no longer hangs the modal;
  the input is focusable immediately and the list populates when the request
  returns.

Tests: every api mock across 6 client test files is now an `async def` fake
(test_tui_app, test_chat_panel, test_sessions_picker, test_terminal_client,
test_tui_export, test_input_box); the sessions picker tests use an async
_raise_not_found helper for the 404 path. Full suite: 939 passed, 1 skipped.

tui: throttle streaming markdown rebuild + cache height (Etap 2) ...

The assistant bubble re-parsed rich Markdown on every stream_delta and rich-rendered it
twice per delta (once to paint, once to measure height) — O(N^2) over the answer length.
Fix the cost while keeping live markdown formatting:

- _ChatItemView.maybe_update now throttles assistant_message rebuilds to ~one per 150 ms
  (coalesced via set_timer). The ChatItem content still updates per delta; only the
  Markdown parse + rich render is deferred. Plain renderables (thinking/planning are
  Text) bypass the throttle — they are cheap. The first mutating delta rebuilds
  immediately for a snappy start; rapid bursts coalesce.
- get_content_height is cached by (renderable identity, width) so the layout pass no
  longer re-renders the rich renderable to measure it between throttled rebuilds.
- stream_end flushes any pending throttled rebuild so the final chunk of streamed text
  appears immediately instead of after the throttle window.
- A deferred rebuild re-anchors scroll to the bottom (via ChatPanel._stick_to_bottom)
  so the freshly rendered streaming text stays in view after its height grows.
- ThemedMarkdownRenderable.__rich_console__ now streams segments straight through
  instead of materializing them into a list first (no per-paint O(N) allocation +
  double walk); the link rewrite is inline in the loop.

Tests: adapted test_per_token_refresh_renders_only_the_streaming_item to the throttled
semantics (a burst yields a couple of rebuilds, not one per token, all for the assistant
item); added throttle-coalesces and stream_end-flushes tests. Full suite: 939 passed, 1 skipped.

tui: code-review Etap 1 — crash/injection/reconnect fixes + drop artifact renderer ...

Point fixes from the full TUI/terminal-client code review (docs/code_review_tui.md),
each with a covering test:

- command_palette: guard on_list_view_selected against IndexError when the
  filter is empty and the placeholder row is selected.
- cli /profile: read session_id (not the absent "id") from the server — the
  handler no longer KeyErrors.
- todo_list: escape task text/validation/index/milestone before interpolating
  into rich markup so a task like "fix [bug]" no longer breaks the line styling.
- builtin /help: drop bold markup tags from status content — StatusRenderer
  builds a literal Text() that would show the tags as literal square brackets.
- tui_app on_user_submitted: always enqueue when a bridge exists (the input loop
  buffers across reconnect) instead of dropping the message on connected=False.
- chat_model stream_start: also reset _current_thinking so a dropped turn (no
  thinking_end) does not glue the next turn's reasoning onto the old block.
- todo renderer update: coerce validation to str before .strip() so a non-string
  JSON value does not AttributeError.
- renderers: remove the unused ArtifactRenderer (no production path emits
  type=artifact — TUI shows files via the OS); keep the diff renderer test.

Full suite: 937 passed, 1 skipped.

tui: unify syntax highlighting under Theme.code_theme; highlight filesystem read output ...

All syntax-highlighted code in the TUI now resolves one Pygments style per
theme from a single point: Theme.code_theme (gnexus-dark=dracula,
gnexus-light=paraiso-light). Previously the dracula/github-light name was
duplicated in artifact.py and markdown_content.py, and the light theme's
"github-light" is not a real Pygments style (only github-dark exists), so
rich silently fell back to "default" and light-mode code was never
highlighted — fixed by switching to paraiso-light.

New renderers/syntax.py::highlight_code() is the shared Syntax factory
(always theme.code_theme + background_color=theme.surface.hex); artifact.py
and the new filesystem read path build through it, and
_theme_aware_code_theme now reads Theme.code_theme via ThemeRegistry.

filesystem FilesystemToolResultRenderer._render_read: the file body is now
syntax-highlighted via highlight_code with the language guessed from the
read path (guess_language). The server's "{num:>width}: {line}" prefix is
stripped (_strip_number_prefix) so Syntax renders its own line-number column
and multi-line constructs (triple-quoted strings, block comments) highlight
correctly across line boundaries; numbered=False renders without numbers.
The header plaque (accent path) and the large-file warning stay as Text;
body is Group(header, [warning,] Text, Syntax). Unknown extensions fall back
to the "text" lexer (plain).

Diff content keeps marker-only coloring (combining Pygments token styles
with the +/- tint is non-trivial); grep/list/info/find unchanged.

Tests: tests/clients/test_code_theme.py pins the single-point contract
(markdown/artifact/read all resolve theme.code_theme, dark != light);
test_filesystem_renderer read tests rewritten for the Group+Syntax body,
plus language-guess and numbered=False coverage. Full suite: 932 passed.

mcp: dedicated runner task per client (fix cross-task cancel-scope on shutdown) ...

MCP client SDK transports (stdio/sse/streamable_http) + ClientSession are
anyio task groups whose cancel scopes require __aenter__/__aexit__ in the
SAME asyncio task. McpClient entered the transport in one task (lifespan
connect / health-check reconnect / request retry) and exited it in another
(lifespan teardown) -> RuntimeError: Attempted to exit cancel scope in a
different task than it was entered in.

Refactor McpClient to own a single long-running runner task that holds the
AsyncExitStack and performs ALL transport enter/exit + list_tools/call_tool.
The public async API (connect/disconnect/list_tools/call_tool/mark_disconnected)
just enqueues a _Cmd and awaits a Future, so callers from any task no longer
cross cancel-scope boundaries. connected/instructions mirror from the runner
onto the instance to stay sync-readable. disconnect() enqueues a stop command
and awaits shield(runner) so teardown isn't interrupted by lifespan cancel.

Also call mcp_manager.stop_health_check() BEFORE disconnect_all() in
AppContainer.shutdown() so the health-check task cannot enqueue onto a
client whose runner is being torn down. mark_disconnected() is now async
(queued) and its manager caller updated.

Regression test: connect in one task, list_tools in a second, disconnect in
a third — the exact scenario that raised the RuntimeError before.

tui: copy chat to system clipboard via reading mode + selection fixes ...

Two ways to copy chat content out of the TUI now work:

1. In-app selection (Ctrl+C, OSC 52): _ChatItemView.render() now builds a
   selection-aware Content from the rich renderable instead of letting Static
   wrap a Panel into a non-selection-aware RichVisual — so the magenta
   drag-select highlight actually renders and get_selection returns clean,
   chrome-free text. rich Segment styles are converted to Textual Style
   (Style.from_rich_style) so the selection_style merge doesn't crash, and
   None segment styles are guarded. get_content_height is overridden to
   compute auto-height straight from the rich renderable; without it the
   width-keyed Content churned new Visual objects per width change and spread
   the layout-settling cascade across refreshes, leaving a deferred scroll_end
   at an intermediate max_scroll_y (chat no longer stuck to bottom). Theme
   defines screen-selection-background/foreground so the highlight is visible.

2. Reading mode (Ctrl+R) for terminals without OSC 52 (e.g. GNOME Console,
   TERM_PROGRAM=kgx, which silently ignores OSC 52 so Ctrl+C never reaches the
   OS clipboard). Toggling the app's reading-mode class hides the chrome around
   the chat (right status/todo column, input prompt, status bar, chat outer
   border) and re-renders every message without its Panel borders via the
   registry's new render_plain — so Shift+drag + Ctrl+Shift+C copies just the
   conversation text to the system clipboard. ContentRenderer.render_plain
   defaults to unwrapping Panel/Padding(Panel) to its body; assistant messages
   and plan_ready override it to raw Text(content) so rich Markdown doesn't
   re-wrap fenced code blocks in their own bordered Panels. ChatPanel tracks
   reading mode and re-renders visible widgets in place (scroll preserved);
   new streamed items pick up the mode on mount.

Help (/help) now lists keys and documents both copy paths.

Tests: render_plain parametrized across all item types, message renderer
plain tests, chat panel reading-mode toggle (borders off/on, scroll kept,
new items borderless), tui_app reading-mode chrome hide, plus the existing
selection suite. Full suite green (922 passed, 1 skipped).

tui: make chat content selectable/copyable without bubble chrome ...

Chat items render rich Panels (rounded borders + Navi/You titles), but the
base Widget.get_selection only extracts text from widgets whose render is
Text/Content -- Panels return None, so a drag-select over a message or tool
output copied nothing (the highlight showed the borders, but Ctrl+C yielded
no message content).

Override _ChatItemView.get_selection: re-render the item's renderable via
the app console at the widget width (line layout identical to the screen),
strip the outer panel border lines and the │ / │ gutters, and remap the
mouse offsets onto the clean content so a partial-line selection lands on
the right characters. Plain renderables (status, turn_meta, filesystem tool
output) have no chrome and map 1:1. Ctrl+C (Screen.copy_text) then copies
the clean text via OSC 52.

tui: restore tool args/metadata, assistant-then-tool order, turn_meta on resume ...

ChatModel.load_history rebuilds the resumed session so the chat renders the
same way as a live turn:

- Recover tool-call args via tool_call_id (the persisted tool message stores
  only the result, not the call args) and forward the stored metadata. Without
  this the filesystem renderer saw args={} -> action=None and fell back to a
  plain, un-highlighted result -- the resume "no syntax highlighting" symptom.
- Emit assistant_message BEFORE tool_started when an assistant message has both
  text and tool_calls, matching the live stream where stream_delta precedes
  tool_started; previously the answer was stranded under its own tool cards.
- Append a turn_meta item when the persisted assistant message carries
  elapsed_seconds, reproducing the single duration line stream_end appends.

is_subagent for resumed tool calls is intentionally deferred (needs storage).

tui: let the user scroll up while the agent streams + keyboard scroll bindings ...

ChatPanel._sync called scroll_end on every WS event (every stream_delta),
yanking the view back to the bottom on each token — impossible to scroll up
and read earlier messages during a response. Stick-to-bottom: capture
is_vertical_scroll_end before any DOM change and only scroll_end when the
user was already at the bottom; auto-follow resumes when they scroll back.

Add app-level keyboard bindings so the chat scrolls without leaving the
input box: Ctrl+Shift+Up/Down (one line), Ctrl+End (jump to bottom). These
keys are not claimed by TextArea (it uses shift+up/down for select and
ctrl+e/end for line end), so they bubble from the input to the app.

compressor: target hysteresis — shrink to 65% after compression, not just below the trigger ...

Compression triggered at 90% but left the context just under the trigger, so a
fixed keep_recent (navi_code: 12 turns ≈ 104 messages) re-triggered a couple
of messages later — the context yo-yoed at the trigger line instead of gaining
headroom. Add context_compression_target (0.65): in turn-based (preturn) mode
compress_context shrinks keep_recent until the verbatim kept region fits 65% of
the window, folding the extra turns into the same single summary LLM call; a
safety net in compress_session token-budget-truncates if the kept region alone
still exceeds the target (midturn kept a huge in-flight turn, or the
keep_recent floor can't fit). Trigger 90% → target 65% leaves real headroom.

filesystem: strip trailing newline in _number_diff so diff lines aren't blank-separated ...

_unified_diff (used by edit_lines / smart_edit) feeds difflib with
"line + \n" so difflib does not emit a "\ No newline at end of file" marker.
With lineterm="" difflib keeps that trailing "\n" on content lines, so the
final "\n".join produced "\n\n" between every diff line — a blank line
between each, bloating the model-facing output and rendering with double
spacing in both clients. Strip the trailing "\n" in _number_diff before
processing; no-op for the _diff action, whose lines come from splitlines()
without trailing newlines.

renderers: colour diff line numbers + read line numbers, keep content neutral ...

Standard for diff and read output: the line number (and the +/- marker for
diffs) is the coloured anchor; the content itself reads plainly.

diff: "{num} {marker}  {content}" with num+marker in the marker colour
(green/red), content neutral. Fixed the number-column regex to "^( +)(\d+)|"
so files with more than 9 lines (width > 1) still match — the old
single-space pattern silently fell back to whole-line colouring.
read: "{num}: {content}" with the number in the accent colour, content
neutral (was dim number).

Applied to both clients: the TUI diff/filesystem renderers and the webclient
ToolCard (renderDiff / renderRead). The model-facing text from the server is
unchanged — this is display-only.

tui: sort milestone groups by status rank so finished ones sink ...

Todo widget already sorted steps within a milestone by status, but the
milestone groups themselves stayed in plan order — a done milestone sat
mid-list while its steps were pushed down. Group milestones by their
worst step status (in_progress < pending < failed/skipped/done) with a
stable sort that preserves plan order within equal rank; the unnamed
(legacy, no-milestone) group participates in the same ranking.

Автономность мелких моделей: OUTPUT DISCIPLINE, milestone-todo, перехват финала хода ...

Два последовательных захода над одной проблемой — мелкие модели (12–30B) в navi_code
теряют автономность на трудных шагах: «остановился поболтать» вместо действия и
слишком большое расстояние между пунктами плана.

Заход 1 (З1–З3):
- З1: OUTPUT DISCIPLINE в системном промпте navi_code — «act, don't announce»,
  без few-shot антипримеров. Контракт хода: ответ без tool_calls = конец хода, поэтому
  объявление намерения текстом убивает автономность; правила заставляют вызывать
  инструмент в том же ходе.
- З2: плоский todo + метка группы milestone + декомпозиция. _parse_plan_steps
  возвращает list[tuple[milestone, text]]; milestone — метка группировки (не сущность,
  без статуса), «done» вычисляется при рендеринге; подшаги = больше плоских шагов
  (без вложенности). TUI side-panel группирует по milestone (плоский фолбэк при пустом
  milestone). Plan depth: max 15→20 + правило декомпозиции.
- З3: adaptive re-plan «длинный шаг» — nudge «разбей шаг» при in_progress ≥ порога
  итераций без смены todo (порог 4, раньше общего anti-stall warning на 8).

Заход 2 (шаги 1–3, после cloud-теста 31b vs 12b):
Корневая структурная причина: весь спасательный механизм (anti-stall warning с явным
предложением reflect, adaptive re-plan) живёт только внутри tool-цикла — nudge
инжектируется в pre_turn *следующей* итерации, которой при «остановился поболтать»
нет (ход закрылся по return до post_turn). 31b застревала через «продолжаю
tool-итерации» → дожала до warning → спаслась; 12b — через «замолчала текстом» → мимо
всех nudge.

- Шаг 1: перехват финала хода. Если модель выдала bare-text, но в todo есть открытые
  шаги (pending/in_progress) и лимит не исчерпан — НЕ эмитить StreamEnd, а сохранить
  ассистентский текст в session.context, поставить системный nudge и continue
  (без StreamEnd, без workers — консистентно с multi-iteration tool-турами). Счётчик
  final_interceptions на AgentTurnContext, лимит final_intercept_limit (default 2),
  эскалация жёсткости (мягкий → «second stop»). has_open_steps в todo.py: пустой
  todo → False (защита casual-сообщений), failed/skipped терминальны. Профильные
  флаги final_intercept_enabled/limit в base.py + loader + admin.
- Шаг 2: жёсткий reflect-триггер в промпте — «~3 tool attempts on the same step
  without progress → call reflect IN THIS TURN (tool call, not reasoning aloud)».
- Шаг 3: открыть replan для застревания — «call replan when reflect showed the whole
  approach is dead (not one failed step, but the approach itself)».

Тесты: 874 passed, 1 skipped. Новые — has_open_steps (5), final intercept (5),
milestone-группировка, adaptive long-step nudge, парсер шагов с milestone-маркером.

cloud: navi_code model → gemma4:31b-cloud для тестирования догадки (31b признала
застревание, 12b — нет); .env cloud-host уже gitignored.

Co-Authored-By: Claude <noreply@anthropic.com>

config: raise context compression threshold 0.70 -> 0.90 ...

Midturn compression now splits the in-flight turn correctly, so it's safe to
let the context grow closer to the limit before compressing. 0.90 gives the
model more raw context per turn and trips the compressor less often.

Worker test context_tokens bumped 50_000 -> 60_000 (76% -> 91.5%) to stay
above the new threshold; other compression tests pass the threshold as an
explicit argument and are unaffected.

Co-Authored-By: Claude <noreply@anthropic.com>

filesystem: show real file line numbers in unified diffs ...

_number_diff parses @@ hunk headers and prefixes each content line with its
real file line number: removed/context lines use the old (from) number,
added lines use the new (to) number, in `{marker} {num}│ {content}` form
(marker first so existing startswith highlighting stays valid; column
right-aligned to the largest line number). Hunk/file headers and the
`\ No newline` marker are left unchanged.

Applied in _unified_diff (covers edit / edit_lines / smart_edit) and the
`diff` action — a single server-side change, so both the TUI and the web
client receive numbered diffs in the tool result text.

TUI highlight_unified_diff renders the `{marker} {num}│` column dim and the
content in the marker color, so the number reads as meta, not as part of the
added/removed text. Lines without the prefix (standalone `diff` event, older
callers) are still colored whole.

Co-Authored-By: Claude <noreply@anthropic.com>