| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| ... | |
| |
|
|
|
|
|
| |
Companion to PAGING_FINDINGS.md capturing the hard-won Textual behaviour
(bindings/mixins, ScrollView repaint & scroll-clamp, loading cover blurs focus,
mouse offset mapping, Input/Footer overlap, no cpp grammar) and the app patterns
(name-gen cache invalidation, cursor+scroll history, name+address follow).
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Right after a rename there's a window where the pseudocode cursor word is still
the OLD name while the symbol table already has the new one, so name-based decomp
follow (refs / resolve / _ref_on_line) all miss -> 'nothing to follow'. Disasm
never hit this because it falls back to an address xref.
Add the same address-based fallback to _follow_decomp: parse the line's /*0xEA*/
marker and follow a code xref from that statement — immune to a stale name.
Verified: with the old token 'sub_53D20' (renamed away) it still follows to the
right address. pilot suite 58/58.
|
| |
|
|
|
|
|
| |
Was 42% (≈100 cols on a 240-wide term). Now 30% clamped to min 42 / max 44, and
columns tightened to 10/21/7 -> a stable ~41-43 cols that fits the content: 46%
on a 90-col term but only 18% at 240 (code area 194 vs ~140 before). Ctrl+B still
hides it. pilot suite 57/57.
|
| |
|
|
|
|
|
| |
Fixed left gutter with right-aligned line numbers (dim; brighter on the cursor
line); the code scrolls horizontally past the gutter while the numbers stay put.
Column cursor, search-column, and mouse mapping account for the gutter offset
(_col_offset); virtual width includes it. pilot suite 56/56.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
When a decompile runs (tab to pseudocode, or an implicit refresh after rename),
DecompView.loading covers the pane with a centered, grayed '― decompiling… ―'
(custom get_loading_widget + .decomp-loading CSS), cleared when the body arrives.
The loading cover blurs the widget, which made Tab fall back to focus-nav instead
of toggling; made the app Tab/Shift+Tab binding priority so it toggles regardless
of focus, and restore focus to the pane when the decompile finishes.
pilot suite 55/55 (adds overlay-shown / overlay-cleared).
|
| |
|
|
|
|
|
|
|
|
|
| |
Textual's ScrollView.watch_scroll_y only repaints when the rounded scroll
changes, and an immediate scroll_to right after setting virtual_size clamps to 0
when the view's size isn't computed yet (e.g. just made visible). DecompView.show
had no deferred fallback at all, so it could stick at the top until a cursor move.
Add a shared ColumnCursor._apply_scroll: set the offset now AND re-apply it after
the next refresh with refresh(layout=True), forcing a repaint at the restored
offset. Used by DisasmView._on_primed and DecompView.show. pilot suite 53/53.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
The pane could keep showing the top (scroll 0) after jumping back even though
scroll_offset was correct: the deferred call_after_refresh set the scroll AFTER
the paint and didn't force a repaint. Now the deferred callback re-applies the
scroll AND refreshes, so the pane is repainted at the restored offset.
Crucially, the prior test only checked scroll_offset.y (which was correct even
when the bug was visible). Added a render-level assertion: trace render_line and
assert the last repaint happened at the restored scroll. (Also: my earlier
scratch repros used a non-existent function name so 'back' was a no-op -- the
real suite test uses a valid second function.) pilot suite 53/53.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
The restored scroll was correct in the final state but load() zeroed
virtual_size (snapping scroll to 0) and only a deferred call_after_refresh
jumped to the target -> a visible flash to the top before settling. Now: don't
zero virtual_size on load, and apply the target scroll synchronously in
_on_primed (plus the deferred call as a fallback for un-sized regions).
Also hardened the scroll test to use a MID-viewport cursor (rel>0). The prior
test had the cursor at the viewport top (rel=0), where scroll-into-view derives
the same scroll -> it would have passed even if scroll restore were broken.
pilot suite 52/52.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Root cause of 'Maximum idalib worker count reached': the app bumped idle_ttl to
~never on every open, so sessions never freed and hit the default cap of 4.
- app: drop the immortal bump_idle_ttl(1e9); rely on the KeepAlive heartbeat
(120s) to keep the running session warm, and open with a moderate idle_ttl
(1800s) so the worker self-exits and frees its slot after the TUI closes.
- spawn.sh: set IDA_MCP_MAX_WORKERS (default 8) for headroom.
- doc: supervisor prunes unreachable workers before counting, so killing an idle
worker frees a slot on the next open.
(Verified separately: big/truncated decompile results propagate renames fine via
the earlier force_recompile fix.)
|
| |
|
|
|
|
|
|
| |
force_recompile takes 'items=[{addr}]', not 'addr' — my call raised and was
swallowed, so the Hex-Rays cache was never cleared and a caller's cached
pseudocode kept showing the old callee name on 'back' (disasm refreshed because
its names are live). Use the correct params. History test now also caches and
verifies the caller's pseudocode. pilot suite 52/52.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Visibility: #search/#rename/#status were dock:bottom like the Footer and got
placed on the SAME row (y=39) as the Footer, which drew over them, so the
typed text was invisible. Drop the docking; they now sit in normal flow just
above the Footer (y=38). Verified: input region is above the footer.
- Stale names on 'back': a rename only invalidated the current function, so
popping back to a caller (cached earlier) showed the old name. Add
Program.bump_names(): a rename bumps a name-generation, clears all disasm block
caches (disasm names are live), and decompilation is generation-checked and
force-recompiled lazily on next view. _after_rename now invalidates globally.
- verified: rename a callee, 'back' to the caller shows the new name.
- pilot suite 51/51.
|
| |
|
|
|
|
|
| |
PageUp/Down and Ctrl+U/D now move the scroll top and the cursor by the same
(clamped) delta, so the cursor keeps its relative screen row instead of snapping
to an edge; at the very top/bottom the cursor jumps to the extreme. Shared
_page_scroll in ColumnCursor, used by both views. pilot suite 49/49.
|
| |
|
|
|
|
|
| |
_goto_line now sets cursor_x to the first match range's start on the target line
(and _hscroll's it into view for the pseudocode view), so jumping to a search hit
positions the cursor exactly on the match, not the old column. Search cancel/empty
also restore the origin column. pilot suite 47/47.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Restoring only the cursor re-derived the viewport via scroll-into-view, so back
landed on the right line but scrolled differently. NavEntry now stores scroll_y
(disasm) and dec_scroll_y/dec_scroll_x (pseudocode); a single _save_current_pos()
snapshots cursor+scroll of both views right before navigating away, and load/show
restore the exact scroll. Disasm defers scroll_to via call_after_refresh (its
virtual_size is only set in _on_primed, so an immediate scroll_to clamps to 0).
Also routes rename's reload through the same save/restore path.
pilot suite 46/46.
|
| |
|
|
|
|
|
|
|
|
| |
_follow_decomp / _xrefs_decomp only matched the pseudocode ref list, so a
function name under the cursor that wasn't in refs (self-reference, or refs
truncated on big functions) did nothing — while disasm worked because it also
resolves the symbol. Both decomp paths now fall back to program.resolve(word)
like disasm does.
pilot suite 45/45.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- 'n' on the token under the cursor opens a rename prompt prefilled with the
current name. The symbol is classified and routed to the right rename batch:
* resolves to a function start -> func
* decompiler ref (global/string) -> data
* pseudocode token, not a ref -> local (Hex-Rays lvar)
* disasm var_/arg_ -> stack
tool errors (e.g. name collisions) are surfaced in the status line.
- after a rename: force_recompile + cache invalidation, reload the current views
(preserving cursor), update the renamed function's row in place (avoids racing
the initial stream), and update nav history names.
- 'Ctrl+S' persists the IDB (idb_save); edits mark the session dirty.
- 'n'/'N' search-repeat dropped for 'n'=rename (repeat stays on '/'+Enter / '?').
- verified live: func rename (list + disasm update), local var rename (pseudocode
update), save. pilot suite 44/44; domain 17/17.
|
| |
|
|
|
|
|
| |
- on_data_table_header_selected sorts the cached function list by the clicked
column (address / name / size); clicking the same header reverses. Header
label shows a ▲/▼ arrow. Sorting composes with the active filter.
- pilot suite 41/41.
|
| |
|
|
|
|
|
|
|
|
| |
NavEntry now also stores dec_cursor/dec_cursor_x. DecompView posts a CursorMoved
(with the line's /*0xEA*/ address) on move/click; the app keeps the top-of-history
pseudocode position current and DecompView.show restores it when returning to a
function. So Esc after a jump lands on the exact pseudocode line+column too,
independent of the disasm position.
pilot suite 38/38.
|
| |
|
|
|
|
|
|
|
| |
NavEntry now stores cursor_x too; the top-of-history position is kept current
(line + column) on every disasm cursor move, and DisasmView.load restores the
column (clamped once the line is cached). So Esc after a jump lands exactly on
the symbol you left from, not the line start.
pilot suite 37/37 (adds 'back restores the exact line + column').
|
| |
|
|
|
|
|
|
|
|
|
| |
- ColumnCursor.on_click maps the click to a virtual (line, column) via
event.get_content_offset (handles padding/border) + scroll_offset, then places
the line+column cursor (disasm posts CursorMoved for the status line).
- double-click (event.chain>=2) = place cursor + follow the symbol under it,
identical to selecting it and pressing Enter.
- works in both views (verified: disasm click on 'call sub_XXXX' lands on the
exact token; pseudocode click maps columns with no padding).
- pilot suite 36/36.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
- ColumnCursor mixin for both views: h/l/left/right, w/b word-hops, 0/$ line
ends; block cursor cell + word-under-cursor highlight rendered on the cursor
line (region-refresh so it stays cheap). vertical moves clamp the column.
- follow/xrefs now use the token under the cursor, so a line with multiple refs
follows the one you're pointing at (proven: 'return sub_A(sub_B,..)' -> cursor
on sub_B follows sub_B). decomp matches a decompiler ref by exact name; disasm
resolves the symbol, falling back to the line's from-xref (address column /
hex operands are excluded so the default position still follows the target).
- horizontal scroll follows the cursor in the pseudocode view.
- dropped the f/b paging aliases (b is now word-back).
- pilot suite 34/34.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Enter follows the reference under the cursor: disasm uses xref_query(from) to
find the code target; pseudocode matches a decompiler ref name on the line.
Navigates to the containing function at the exact line (address-history push).
- x opens an xrefs popup (XrefsScreen modal): xrefs_to the subject under cursor
(call target if any, else the current item); Enter jumps to a referencing
site, Esc closes.
- domain: Program.function_of (mid-addr -> containing func via lookup_funcs),
xrefs_from/xrefs_to (xref_query), Xref model, DisasmModel ea->index map for
landing on an exact address.
- fix: goto to a mid-function address now lands on the right line (was opening
it as a bogus function start).
- pilot suite 32/32; domain 17/17.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
- Incremental client-side filter over the cached function list (no per-keystroke
server round-trips): narrows + highlights the matched substring in each name
as you type, debounced ~80ms. Substring or glob (*/?), smartcase.
- Clear the filter: Esc in the filter box, OR Esc on the focused list, OR empty
Enter. Status shows 'filter <q>: N/total'.
- streaming pauses while a filter is active; re-applied over the full set once
load completes. Row selection resolves the name from the index (cells can now
be styled Text).
- FunctionIndex.all_loaded(); pilot suite 27/27.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
- Fix overlap: #search, #status and Footer all docked bottom, and Input defaults
to a 3-row bordered widget squeezed into 1 row. Now the search bar is a clean
1-row (border:none) that REPLACES the status line while active (status hidden),
so the query is always visible.
- Incremental search: highlight all matches + preview-jump from the origin on
every keystroke (Input.Changed), not just on submit. Disasm indexes once then
each keystroke is instant. Enter keeps; empty Enter repeats last; Esc cancels
and restores the origin cursor + status line.
- pilot suite 23/23 (adds: bar-visible/no-overlap, incremental-before-Enter,
Esc-cancel).
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
- SearchMixin shared by both views: '/' search forward, '?' backward, empty query
repeats last; n/N next/prev. Smartcase. All matches highlighted (substring
overlay), cursor lands on current match; status shows k/N.
- disasm indexes line text lazily in a worker (handles huge funcs); pseudocode
has text in hand. Substring highlight via crop/join Strip overlay.
- app: bottom search prompt, SearchRequested routing; '/' is context-sensitive
(filters when the function table is focused, searches in code views).
- gotcha fixed: Textual only merges BINDINGS from DOMNode subclasses, so mixin
bindings are ignored -> each view lists SEARCH_BINDINGS explicitly.
- pilot suite 20/20 (adds search finds/hl/repeat-fwd/repeat-back).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Highlighting: Textual has NO C/C++ tree-sitter grammar (and the syntax extra
wasn't installed), so language='cpp' was a silent no-op. Replace TextArea with a
virtualized, Pygments-highlighted ScrollView:
- idatui/highlight.py: CLexer -> Rich styles, per-line Segment lists
- DecompView: lines highlighted ONCE at load, cached as Strips; O(1) scroll
Jank: profiling showed our code is ~3.6ms/move (render_line 0.018ms) -- the cost
was terminal repaint volume, since every cursor move refreshed the whole ~43-line
viewport. Now:
- cursor reactive repaint=False (no implicit full refresh)
- region-limited refresh: in-place moves repaint only the 2 changed rows
(measured 43 -> 2 render_line calls/move), full refresh only on scroll
Applied to both DisasmView and DecompView.
pilot suite 15/15 (adds highlighting assertion).
|
| |
|
|
|
|
|
|
|
|
|
| |
- domain.decompile now fetches the FULL body: server caps responses at 50KB and
clips strings to 1KB, but caches the full output and exposes it at
_meta.ida_mcp.download_url -> we GET that so pseudocode is complete
(main: 43257 chars, not a 1KB stub). include_addresses gives per-line /*0xEA*/.
- DecompView (read-only TextArea, cpp highlighting best-effort, line numbers)
- Tab and Shift+Tab toggle the code pane; lazy decompile in a worker, cached;
hard-failures ('decompilation failed') shown gracefully with disasm fallback
- pilot suite 14/14 (adds tab->pseudocode, full-body, shift+tab->disasm)
|
| |
|
|
|
|
|
| |
- app: when open_path is set, idb_open (idle_ttl=1e9, long timeout) and pin the
returned session; clean status on failure (e.g. non-writable dir)
- tui.py: --open flag + usage docs (attach vs open)
- verified: --open targets/cat -> new session, 161 funcs loaded
|
| |
|
|
|
|
| |
- action_toggle_functions: collapse/restore left pane, disasm takes full width;
focus follows (disasm when hidden, table when shown)
- pilot test covers the toggle (11/11)
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- idatui/app.py: keyboard-first Textual app
* FunctionsPanel: lazy-streamed function list (DataTable) with glob filter
* DisasmView: line-VIRTUALIZED ScrollView -- paints from domain block cache,
background-fetches misses, prefetches neighbors. Proven: 52,120-insn func,
jump-to-bottom in 0.31s with only ~6 blocks (~1.5k lines) ever materialized.
* address-history nav stack (Enter follow / Esc back); goto by name or 0xADDR
* bump_idle_ttl + KeepAlive on startup so the session never idles out
* all MCP work on Textual worker threads; UI never blocks
- idatui/tui.py: launcher (python -m idatui.tui / console script 'idatui')
- tests/test_tui.py: headless Pilot test, 9/9 (boot, load 10k funcs, open,
scroll, bg-cache, goto-bottom, filter)
- pyproject: textual extra + idatui script entry
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Root cause: idalib workers run an idle watchdog (worker_lifecycle.py) that
self-exits after idle_ttl_sec (default 600s) of no requests -- deliberate
resource hygiene for a shared/ephemeral supervisor, wrong for an interactive
TUI. CLI-opened startup binary always gets 600s (no flag).
Fix (both verified against a 12s-TTL session):
- IDAClient.bump_idle_ttl(1e9): re-open (idempotent) re-applies TTL, no upper
cap -> effectively immortal. Primary fix, call once at startup.
- IDAClient.keepalive()/KeepAlive: background server_health heartbeat resets the
watchdog; covers adopted sessions too. Safety net.
tests/test_keepalive.py: 4/4 (heartbeat + bump both keep a short-TTL worker
alive past 2x TTL). docs updated with why + fix.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
- Program: model registry + small prefetch pool; cache invalidation hooks
- FunctionIndex: lazy pagination (advance by len, clamp page=500), filter globs,
viewport windows; full 10,092-func enumerate matches survey in ~3.1s
- DisasmModel: block-cached windowed disasm (256/block), forward+back prefetch,
cached instruction totals; deep window revisit 196ms -> 0ms (cache proven)
- Decompilation: truncation-aware, hard-failure surfaced as data (monster funcs)
- resolve(): int/hex/symbol via lookup_funcs (string-array query shape)
- tests/test_domain.py: 17/17 on both ls (~290 funcs) and libcrypto (10k+52k-insn)
- smoke_client no longer assumes 'main' exists (works on libraries)
- doc: idle_ttl worker self-exit finding ("not reachable" needs re-open)
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Measured against libcrypto.so.3 (10,092 funcs, biggest 52,120 insns):
- list_* count cap ~700, disasm max_instructions cap ~500, then SILENT
collapse to 10 (not clamped) -> must clamp client-side (use 500).
- pagination must advance by len(data); next_offset=offset+count skips data.
- disasm offset paging is O(offset): 6ms@0 -> 180ms@50k, no resumable cursor
-> domain layer must cache windows + prefetch + over-fetch.
- include_total scans whole func (~217ms on monster) -> fetch once, cache.
- decompile hard-fails on huge funcs as a soft error (code=null) -> handle.
- idb_open needs a writable path for the .i64.
tests/stress_paging.py: durable paging benchmark harness
docs/PAGING_FINDINGS.md: constraints that drive the paging layer
|
| |
|
|
|
|
|
|
|
|
|
| |
- auto-recover stale db ("Session not found" after server restart): drop pin,
re-resolve sole session, retry once. Only when db was auto-injected; never
silently switches an explicitly-pinned db (auto_recover_session gate).
- tests/serverctl.sh: start/stop/kill9/ready control over spawn.sh
- tests/stress_client.py: 6 adversarial scenarios, 24 assertions:
dead-port fail-fast, tight-timeout recovery, worker crash, full restart
(naive auto-recovery + no-autoswitch), 300/16 concurrency, kill-under-churn.
All clean: no hangs, no deadlocks, no non-IDAError leaks.
|
|
|
thread-safe
- IDAClient: single handshake, ~7ms warm calls (vs ~60ms cold CLI)
- keep-alive connection pool over http.client; concurrent-safe (40 threads OK)
- grounded error taxonomy: IDAToolError on isError, soft per-item errors as data
- session-expiry recovery (404 -> re-handshake -> retry), bounded transport retries
- auto db-injection + resolve_db (ignores stale empty-id sessions)
- live smoke test: 13/13 pass
|
