CHANGELOG.md

Changelog

[Unreleased]

[0.13.0] - 2026-07-31

Added

• Oh My Pi (omp) session support (#56, by @Michelh91) — discover, preview, resume (omp --resume <id>), and delete Oh My Pi sessions under ~/.omp/agent/sessions/, reusing the pi JSONL parser and excluding nested subagent transcripts.
• Yolop session support (#55, by @chaliy) — discover, preview, resume, and delete sessions from Yolop's platform-native session store. Current metadata supplies session titles, canonical repository names, timestamps, and concise worktree labels; older sessions fall back to first prompts and repository-name recovery.
• Contributor guide for adding an agent — docs/adding-an-agent.md documents the full wiring checklist, including the three Vec registrations the compiler can't enforce.

Fixed

• Bulk delete could delete the wrong sessions (#58) — multi-select stored sessions Vec indices captured at toggle time, but a background scan landing mid-selection reorders that Vec, so confirming a bulk delete removed different sessions than the ones checked. Selections are now keyed by (agent, session_id) and resolved at delete time.
• session_id is shell-escaped in resume commands (#58) — the id was wrapped in raw single quotes, so an id containing ' broke the eval'd command and a crafted transcript filename could inject shell. It now goes through the shell-aware quoter (POSIX and PowerShell).
• Cursor and pi sorted by creation time, not last activity (#58) — a session created long ago but used today sorted as old and sank below stale ones. Both now use the transcript's last-modified time (pi: the max of its header timestamp and mtime), matching the other agents. Oh My Pi inherits the pi fix.
• Grouped view ordered projects by the first session, not the newest (#58) — correct only under Time sort; grouped view now keys on each group's max session timestamp.
• Codex timestamp overflow on corrupt data (#58) — updated_at * 1000 now saturates instead of wrapping to a garbage sort key.
• Streaming startup keeps the initial cursor at the top (#57, by @MilkClouds) — when a fast scanner (commonly OpenCode) returned before a slower one (commonly Claude Code), the later merge pushed the cursor down the newly sorted list. The top row now follows incoming results until the user moves away from it, while explicit selections remain anchored.

Performance

• Faster cold scan on large ~/.claude/projects trees (#58) — the Claude metadata scan read whole transcript bodies (up to 272 KB each) just to reach the tail for the off-by-default recap. The tail window is now 32 KB; measured read I/O dropped 62 MB → 29 MB (−53%) on a 714-file tree, with the recap unchanged.

Changed

• Dropped panic = "abort" (#58) — restores scan_all's per-thread panic isolation so one malformed session file can't abort the whole listing.

[0.12.0] - 2026-06-11

Added

• Grouped view: Ctrl+L previews the selected session (#50, by @soomtong) — the first grouped→preview path; child rows jump straight to the session detail pane.
• j / k navigation in the action menu, Ctrl+P / Ctrl+N in the help screen (#50, by @soomtong) — consistent with the other modes; the footer advertises Tab/jk.

Changed

• List scrolling keeps a 3-row margin below the cursor (#50, by @soomtong) — browse, grouped browse, and agf watch all scroll before the cursor hits the bottom edge. Review fixups hardened the arithmetic (see Fixed) and added the missing end-of-list clamps so the margin can't overscroll into blank rows.
• Summary cycling [ / ] now wraps around (#50, by @soomtong) — cycling past the last summary returns to the first; the help screen reads [ or ] to avoid /-key confusion.
• Deletes only disappear from the UI when the delete actually succeeded — both delete paths previously did let _ = delete_session(...) and removed the row unconditionally, so a permissions failure looked like success until the session reappeared on next launch. Failed deletes now stay visible.
• Rust edition 2024, rust-version = "1.88" declared (#52) — the floor is set by let-chains; the code already used 1.87 APIs (usize::is_multiple_of), which clippy::incompatible_msrv surfaced the moment an MSRV was declared.
• rusqlite 0.32 → 0.40, toml 0.8 → 1 (#52) — both verified zero-code-change bumps; the bundled SQLite moves to libsqlite3-sys 0.38. sha2 0.11 (breaks {:x} digest formatting) and superlighttui 0.21 (adds 29 #[must_use] call sites) were evaluated and deliberately deferred.

Fixed

• Windows without the shell wrapper: bare agf no longer crashes with program not found (#49, by @MilkClouds) — with AGF_SHELL unset the shell defaulted to POSIX on every OS, so native Windows exec'd cd '…' && agent via a nonexistent sh. The default is now OS-aware via a pure, unit-tested default_shell (native Windows → PowerShell; MSYSTEM/unix-style SHELL → POSIX). POSIX hosts are bit-for-bit unaffected.
• Upgrades no longer serve stale per-session cache data (#51, closes #37) — mtime freshness can't see "the data didn't change but its interpretation did" when scanner logic changes within one CACHE_VERSION. The cache now stamps the writing binary's version and any mismatch forces a one-time rescan; AGF_DEBUG=1 logs cache built by X, current Y → rescanning. The write-side carry-over path is gated the same way, and per-release CACHE_VERSION bumps are no longer part of the release checklist.
• Gemini scanner: UTF-8 char-boundary panic in extract_summary_partial (#52) — a multi-byte (e.g. CJK) character straddling the 1 KiB partial-read window made the byte slice panic and silently killed the entire Gemini scan. The window now backs up to a char boundary; CJK regression test added.
• agf resume --list 0 <query> no longer panics (#52) — --list 0 underflowed top_n.len() - 1; it now shows the top result instead.
• Cursor scanner: hex_decode panic on non-ASCII store.db metadata (#52) — a corrupt meta blob could index mid-char; an is_ascii guard rejects it and the scan continues.
• Scroll-margin arithmetic underflow (review fixups on #50) — the margin branch can fire while selected < visible, where selected - visible + 1 + margin underflows usize: debug builds panicked on routine down-navigation (release builds wrapped to the right value by accident). All three sites now use saturating arithmetic; regression tests pin no-underflow, margin, and end-clamp behavior.
• agf watch: selection clamped when a refresh shrinks the list (#52) — a background refresh that removed sessions could leave the cursor past the end, blanking the viewport.
• Ctrl+H no longer leaks into the fuzzy-search textarea (#50, by @soomtong).

Internal

• Audit-driven modernization (#52) — 31 verified findings applied: three shared scanner helpers (collapse_whitespace, project_name_from_path, push_concat_titles) replacing 5–6 duplicated sites each; Agent derives serde and the hand-rolled string maps are deleted (on-disk cache format pinned byte-identical by a round-trip test); let-else/let-chain flattening across scanners, delete, and TUI dispatch; per-frame allocation cuts (render_chunks takes its chunks by value, update_filter/apply_sort stop cloning, detect_editor caches config for the process lifetime); claude history.jsonl pre-filters orphaned sessions the way codex has since v0.11.4; read_first_line gets the same 512 KiB byte budget as the other bounded readers. Tests grew 53 → 75.
• CI hardening (#52) — clippy runs --all-targets (tests were unlinted in CI) including a new Windows lint job that caught a real platform gap on its first run; Swatinem/rust-cache (Test job 35s → 17s); per-ref concurrency cancellation; --locked everywhere cargo resolves. The release workflow gains a tag↔Cargo.toml version guard, loud checksum failures, and a graceful publish skip when CARGO_REGISTRY_TOKEN is unset instead of a red run on every release.

[0.11.4] - 2026-05-29

Fixed

• scanner/pi: a single invalid-UTF-8 line no longer drops the rest of the file (or the whole session) — parse_session used reader.lines().map_while(Result::ok), which stops at the first Err and silently truncates every prompt summary that follows. Worse, when the bad line lands before the session header (a real failure mode for crash-truncated or rotation-racing writes), the header is never captured and the entire session is dropped from agf list. This is the same bug class as the extract_first_prompt .ok()? regression fixed in v0.11.3 for the Cursor scanner. Replaced with let Ok(line) = line_result else { continue; }; so each bad line is skipped individually. Regression test: parse_session_skips_invalid_utf8_lines.
• scanner/codex: ~/.codex/history.jsonl no longer scales with the user's lifetime codex usage — read_history_summaries streamed the entire file, parsed every line into a HistoryEntry, and accumulated (f64, String) tuples for every session_id ever seen — including thousands of sessions that no longer have a rollout JSONL on disk. For power users who run codex daily, the file reaches tens of MB; v0.11.3's CACHE_VERSION bump to 6 forces a cold rescan on every upgrader, which would otherwise pay this cost on first launch. The function now takes the same live_session_ids set already collected by collect_live_session_ids and short-circuits any line whose session_id is not in it. When live_session_ids is None (transient I/O on the sessions tree), the legacy "keep everything" behavior is preserved, mirroring scan_sqlite's same-condition fallback. Regression tests: read_history_summaries_pre_filters_against_live_session_ids and read_history_summaries_keeps_all_when_live_set_is_none.

Docs

• README: Kiro row now surfaces "no per-session resume — always opens the latest session for the cwd" — kiro-cli ignores session_id, so selecting a specific older Kiro entry in the TUI silently launches a different session. The caveat was previously only in Agent::Kiro::resume_cmd's inline comment; it now appears in the top agents table where users actually read.
• README: Hermes row now surfaces "cwd-independent — resumes in your current shell directory" — documented in the expanded Full session storage paths section but missing from the discoverable top table.

Internal

• .gitignore: /.claude/ added — every contributor running Claude Code locally was seeing ~/.claude/ show up as untracked in git status, with the latent risk of an accidental git add . committing a personal agent state directory.

[0.11.3] - 2026-05-27

Fixed

• Cursor scanner: walk both legacy .txt and current Composer 2+ .jsonl layouts (#45, by @rooty0 / Stan) — current Cursor stores transcripts at ~/.cursor/projects/*/agent-transcripts/<uuid>/<uuid>.jsonl (depth 4) rather than the legacy ~/.cursor/projects/*/agent-transcripts/<uuid>.txt (depth 3). The scanner walked depth 3 with a .txt-only filter, so on current Cursor it returned zero sessions. Verified against live data: ~/.cursor/projects had 2 JSONL transcripts at depth 4 and 0 TXT, and agf list --agent cursor-agent returned No sessions found. before this release. Closes #35.
• Cursor scanner: read chat metadata from the right table (#45, by @rooty0) — the previous code queried SELECT value FROM cursorDiskKV WHERE key = 'composerData', which is the IDE's state.vscdb schema, not the CLI's store.db. Cursor CLI's store.db actually exposes a meta(key TEXT PRIMARY KEY, value TEXT) table with a single key = '0' row whose value is a hex-encoded JSON containing agentId, name, createdAt, mode, and lastUsedModel. Verified via sqlite3 against a real store.db on disk.
• Cursor scanner: skip JSONL transcripts whose store.db is missing (#45, by @rooty0) — cursor-agent --resume only surfaces sessions that have BOTH a transcript and a ~/.cursor/chats/<workspace>/<session_id>/store.db entry; reporting orphaned transcripts that the CLI itself refuses to resume just confuses the listing. Legacy .txt sessions are unaffected (they predate the chats/ directory).
• Cursor scanner: fall back to the first user prompt when store.db has no usable metadata (#45, by @rooty0) — the JSONL is parsed for the first role: user text part, with <user_info> system injections skipped and <user_query> wrappers stripped.
• extract_first_prompt no longer panics on inverted <user_query> tags — str::find returns the FIRST occurrence of each substring independently, so a text part where </user_query> byte-precedes <user_query> (e.g. a pasted log or AI-generated code sample) gave start > end and text[s+12..e] panicked with begin > end. Confirmed via a standalone rustc reproducer. The closing tag is now searched after the opening one. Regression test: extract_first_prompt_does_not_panic_on_inverted_tags.
• extract_first_prompt no longer aborts on the first malformed or non-UTF-8 line — both the per-line IO read and serde_json::from_str used .ok()?, which propagates None out of the whole function on the first error instead of skipping the bad line. A single corrupted/truncated/non-UTF-8 line at the top of the JSONL silently disabled the blank-summary fallback for the rest of the file. Replaced with let Ok(...) else { continue; }; (matching scanner/pi.rs). Regression tests: extract_first_prompt_skips_malformed_json_lines and extract_first_prompt_skips_invalid_utf8_lines.
• extract_first_prompt now bounded by a 512 KiB byte budget — pi.rs added this safeguard in v0.11.2 after large Claude logs stalled the TUI; Cursor transcripts can carry multi-MB tool-result blobs, and the CACHE_VERSION bump in this release forces a cold rescan for every upgrader, so the same precaution applies.
• Cursor delete: legacy .txt transcripts now actually get removed — delete_cursor_agent_session called remove_dirs_matching_name(&projects_dir, &session.session_id), but that helper filters on path.is_dir() AND file_name == name. Legacy sessions live at agent-transcripts/<uuid>.txt (a file named <uuid>.txt), so it never matched. Delete returned Ok(()), the orphan file persisted on disk, and the next scan resurrected it. A new sibling helper remove_files_matching_name removes the file form alongside the directory form. Regression test: delete_cursor_agent_removes_legacy_txt_transcript.
• Cursor scanner: enforce stem == parent UUID invariant on the .jsonl arm — the previous check only required the grandparent to be named agent-transcripts. A stray agent-transcripts/<uuidA>/<uuidB>.jsonl would produce session_id = uuidB, which mismatches both the store.db lookup and what cursor-agent --resume expects. Real Cursor always writes them equal, but the invariant is now explicit. Regression test: scan_from_rejects_jsonl_with_stem_mismatched_to_parent.
• decode_dash_path test coverage for hyphenated project segments — added decode_dash_path_resolves_hyphenated_segments which places agent, agent-tui, and agent-tui-finder as sibling directories and asserts the backtracking decoder resolves to the longest existing match. This is the load-bearing case for this very repo's path.

Changed

• CACHE_VERSION bumped to 6 — the new orphan-skip rule fires only on fresh scans; cached 0.11.x cursor entries persist with their old summaries until each transcript's mtime changes. Bumping the version forces a one-time rescan on upgrade so the "35 orphans → 0" effect actually lands for upgraders.

Docs

• README: Cursor CLI doc link + transcript paths updated (#45, by @rooty0) — docs.cursor.com/agent no longer resolves; switched to cursor.com/docs/cli/overview. Storage column now lists both the current JSONL layout and the legacy .txt form.

[0.11.2] - 2026-05-23

Fixed

• pi: resume the selected session by id (#43, by @shellus) — pi sessions were resumed with a bare pi --resume, which opens an interactive picker / the latest session for the cwd and ignored the session the user actually selected in agf. The command is now pi --session '<id>'. Verified against the pi-mono source (resolveSessionPath in packages/coding-agent/src/main.ts): an argument with no /, \, or .jsonl is treated as a session-id prefix and matched against the session store; because agf wraps resume as cd '<project_path>' && pi --session '<id>', pi resolves it in the session's own project directory and resumes directly without a fork prompt. (Confirmed on pi 0.53.0.)
• pi: keep every session from a project selectable (#43, by @shellus) — the scanner deduped to the most recent session per project directory, a workaround for the old "only resumes latest" assumption. Now that resume-by-id works, all sessions are listed and individually resumable.
• pi: show prompt summaries and full history (#43, by @shellus) — the scanner now extracts user-message text from each session's JSONL (collapsing whitespace, capped at 120 chars per line) so the listing shows what a session was about, and the Preview History: pane lists every prompt — parity with the other agents instead of bare project (model) rows.

Changed

• cache: CACHE_VERSION bumped to 5 (#43, by @shellus) — the pi payload now carries prompt summaries, so 0.11.x cache entries written before this release would otherwise keep rendering only the project name until each session's source mtime happened to change. The bump forces a one-time rescan on upgrade.

Internal

• pi scanner: bound the per-file read with a 512 KiB budget — collecting every prompt requires reading the whole JSONL, but pi transcripts can grow to several MB and the CACHE_VERSION bump forces a cold rescan for everyone on upgrade. The read loop now stops after 512 KiB (the header is the first line, so it is always captured), mirroring the read_head_tail cap added in v0.10.1 after large Claude logs stalled the TUI.
• model: document Kiro's resume-latest behavior — restored the note (dropped in #43) that kiro-cli chat --resume has no per-session flag and ignores session_id, now placed on the Agent::Kiro arm of resume_cmd where the command is defined.

Reverted

• TUI default-to-cwd search query (from #43) — the merged PR also pre-filled the search box with $PWD when agf was launched without a query. That changed the no-arg behavior for every agent (empty list in non-project directories) and re-introduced the cwd special-casing deliberately removed in v0.11.0 (the cwd-match sort boost made time-sort look broken). Reverted to keep this release scoped to pi; a cwd default can land later as its own opt-in setting.

[0.11.1] - 2026-05-04

Changed

• superlighttui bumped 0.17 → 0.20.1 — picks up 8 patch/minor releases of the underlying TUI library since v0.10.0 was tagged. Drop-in: every public API agf uses (Context, Color::Rgb, KeyCode, KeyModifiers, RunConfig, TextareaState, container/col/row builders, help_colored, separator_colored, consume_key, mouse_down, quit) is unchanged; agf does not touch any of the v0.20 breaking APIs (gauge/line_gauge/breadcrumb chainable builders, scrollable_with_gutter GutterOpts, Constraints WidthSpec/HeightSpec redesign, f32 → f64 ratio unification on SplitPane). Notable improvements inherited for free:
  • 3–5× flush-path speedup on redraw-heavy frames (SLT 0.18.2 #62) — flush_buffer_diff now coalesces consecutive same-style cells in a row into a single Print(run) instead of per-cell, dropping queue! calls roughly 12000 → 2000 on a 200×60 redraw.
  • ~1000× speedup on static frames (SLT 0.20.0 #171) — per-row hash skip in flush_buffer_diff short-circuits cell iteration on rows whose contents and style didn't change. The agf list view, which is mostly static while the user reads it, gets the full benefit.
  • rgb_to_ansi256 u8 overflow fix at r=g=b=248 (SLT 0.19.1 #104) — agf's color palette (Rgb(229, 229, 229), Rgb(245, 158, 11), etc.) sits below the threshold so the user-visible bug never fired in agf, but downstream installs on color-limited terminals now route grayscale tones to the correct ANSI 256 cell instead of silently mapping to Indexed(0) (Black).
  • stdout BufWriter (SLT 0.19.1 #172) — every frame now batches dozens of write_all ANSI commands behind a 64 KiB BufWriter, ending with a single flush(). Reduces syscalls per frame on every TUI mode without any agf-side change.
  • Bordered title CJK truncation + overdraw fixes, image command count drop, treemap/textarea panic fixes (SLT 0.19.x), textarea undo/redo, modal tab_trap opt-in, Anchor enum + modal_at (SLT 0.20.0) — not used by agf today; available for future TUI work.

Fixed

• DeleteConfirm: arrow Up/Down (and j/k, Ctrl-j/Ctrl-k) now toggle Yes/No (#38) — previously only Left/Right worked, which was surprising on the v0.11.0 release because every other modal-style picker in agf accepts both axes. The toggle is the same direction-agnostic flip (Yes ↔ No) regardless of which arrow key fires.
• ActionSelect / AgentSelect / PermissionSelect / ResumeSelect: arrow keys cycle through the menu instead of stopping at edges (#39) — Up on the first item now jumps to the last; Down on the last wraps to the first. Tab/BackTab already wrapped — Up/Down/Ctrl-p/Ctrl-n/Ctrl-j/Ctrl-k now match.
• Footer shows agf v<version> on the leftmost cell of every mode's status bar (#40) — version is read from CARGO_PKG_VERSION at compile time so it stays in sync with the published crate. The 10 per-mode status bars are now routed through a single render_footer helper to keep layout uniform.

[0.11.0] - 2026-05-04

Added

• Hermes Agent support (#34, by @SHL0MS) — adds Hermes Agent (Nous Research's self-improving agent) as a first-class scanner. Reads top-level sessions from ~/.hermes/state.db (SQLite); aggregates child-session titles as additional summaries (same pattern as OpenCode subagents); cascade-deletes messages → child sessions → parent → on-disk JSON dumps under ~/.hermes/sessions/session_<id>.json. Resume via hermes --resume <session_id>.

Fixed

• scanner/hermes: pull first user message as preview when title is NULL — Hermes lazily auto-generates titles after the first exchange, so short sessions stay un-titled and the listing fell back to bare api_server session (model) — N msgs. The scan query now selects the earliest messages.content WHERE role='user', collapses whitespace, strips <user_query> wrapper tags, and caps to 160 chars so the detail pane shows what the conversation was actually about.
• scanner/hermes: empty project_path so resume stays in the user's cwd — Hermes is cwd-independent; the original PR set project_path = ~/.hermes, which made agf emit cd ~/.hermes && hermes --resume <id> and yanked the shell out of whatever project the user was working in. shell::cd_and now skips the cd when the quoted path is empty ("" / '' / \"\"), and the Hermes scanner leaves project_path empty. display_path() renders empty paths as — so the TUI doesn't show a blank cell.
• delete/hermes: wrap cascade in a single SQLite transaction — the original four DELETEs ran independently; a mid-cascade failure (disk full, DB locked, etc.) could leave orphan messages whose sessions row was already gone, surfacing as ghost rows on the next scan. The four DELETEs are now BEGIN/COMMIT-wrapped via Connection::transaction.
• TUI: time sort not applied on first frame when config.sort_by is unset (regression visible since the cache was grouped per-agent) — main.rs only called app.apply_sort() inside an if let Some(sort_by) = config.sort_by, so the default-sort path silently rendered the cache's per-agent grouping order. A session from 11 minutes ago could land below sessions from a month ago on the first paint. apply_sort() is now always called; sort_mode falls through to SortMode::Time when config.sort_by is None.
• TUI: cwd-match boost removed — the secondary sort that pushed sessions whose project_path == $PWD above everything else made the time-sorted listing look broken when agf was launched from inside a project: 11 sessions from "this project" surfaced first, then suddenly a 2-minute-old session from another project, then a 31-minute-old Hermes session, and so on. The boost was implicit (no on-screen indicator), so users read it as "time sort is wrong." Pinning is still honored — it's an explicit user action — but cwd is no longer special-cased.
• cache: bump CACHE_VERSION to 3 — Hermes Agent gained a new entry in the per-agent cache map, and the Hermes session payload now carries first-user-message previews instead of bare source/model fallbacks. Without bumping, 0.10.x cache files would surface as stale "cli session (...)" summaries on first 0.11.0 launch until the underlying DB mtime happened to change. The bump forces a one-time rescan on upgrade.

Notes

• Cursor CLI scanner regression tracked separately in #35 — recent cursor-agent installs write transcripts as <id>/<id>.jsonl (depth 4) instead of the legacy <id>.txt (depth 3) the scanner expects, so agf list --agent cursor-agent returns 0 sessions on those installs. Out of scope for this release.

[0.10.3] - 2026-05-01

Fixed

• scanner/claude: skip sessions whose per-session JSONL is missing (#27, #29, by @Mert-coderoid) — ~/.claude/history.jsonl accumulates session IDs forever; Claude Code never trims it when the per-session JSONL under ~/.claude/projects/*/<id>.jsonl is deleted. Those orphan IDs surfaced in the listing and resume failed with No conversation found. scanner::claude::scan now filters its output through the set of session IDs that have a JSONL on disk; the directory walk is shared with scan_session_metadata so the tree is still read once per scan.
• delete/codex: also remove the SQLite threads row (#28, #30, by @Mert-coderoid) — since the Codex scanner moved to state_*.sqlite as primary source, deletes that only removed the rollout JSONL and history.jsonl entry came back on the next scan. delete_codex_session now walks every state_*.sqlite in ~/.codex/ and runs DELETE FROM threads WHERE id = ?1; older-schema dbs without threads are skipped silently.
• scanner/codex: prune orphan threads rows on scan (#31, #32, by @Mert-coderoid) — state_*.sqlite rows whose rollout JSONL no longer exists kept dominating agf list / agf stats (e.g. 324 ghost rows for an empty ~/.codex/sessions/) and could not be resumed. scan_sqlite now builds the live session-id set from ~/.codex/sessions/, excludes orphans from the listing, and DELETEs them from every state_*.sqlite. A walker error with no live IDs collected falls back to the legacy "surface every row" behavior so a transient I/O failure cannot wipe the table. Note: Codex scans are no longer strictly read-only — they will hard-delete unrecoverable orphan rows.

[0.10.2] - 2026-04-25

Fixed

• Cache write race on early TUI exit (regression in 0.10.1) — v0.10.1 introduced streaming background scans; if the user exited the TUI before a worker finished, cache::write_cache persisted that agent as (empty session list, fresh mtime), hiding its sessions on the next launch until file mtime changed again. write_cache now takes the still-scanning agent set and carries the prior cache entry verbatim for those agents — only completed scans replace cache state.

[0.10.1] - 2026-04-25

Fixed

• TUI hangs / never opens on heavy Claude logs — scanner::claude::scan_session_metadata line-iterated every per-session JSONL to EOF to find the latest away_summary recap. On a directory with multi-MB session files (10+ MB jsonl is common with long Claude Code sessions), rayon would parse hundreds of MB in parallel and the TUI never showed up. Fixed by capping per-file I/O to 16 KB head + 256 KB tail via a new scanner::read_head_tail helper: cwd and aiTitle are extracted from the head, the latest away_summary from the tail, and small files (≤ 272 KB) still read in full. Cold scan on a ~1.6 GB Claude log directory dropped from 48 s to 0.7 s in local testing.

Changed

• Background scan + streaming TUI ingest — cache::start_stale_scan returns an mpsc::Receiver<ScanResult> and the TUI now drains it on every render frame. With a warm cache, the TUI opens instantly on cached sessions and the scanning agents stream results in as they finish (footer shows • scanning N… until every worker reports). With a cold cache, the TUI still opens immediately on the first agent that completes instead of blocking on the slowest one. Final cache write happens at TUI exit.
• Per-agent scan timing under AGF_DEBUG=1 so users can locate the slow agent on their machine.

[0.10.0] - 2026-04-25

Added

• Windows / PowerShell support (#22, by @MilkClouds) — agf init powershell (alias pwsh) emits a wrapper compatible with Windows PowerShell 5.1 and PowerShell 7+; agf setup auto-detects PowerShell on Windows and writes to $PROFILE.CurrentUserAllHosts. A new CommandShell (Posix / PowerShell), selected via the AGF_SHELL env var the wrapper sets, routes action::* and the TUI new-session preview through shell-specific quoting ('' vs '\'') and cd_and (Set-Location ...; if ($?) { ... } vs cd ... && ...). POSIX behavior is unchanged when AGF_SHELL is unset.
• Windows agent detection — is_agent_installed now matches %PATHEXT%-aware stems on Windows, so claude.exe / claude.cmd / claude.ps1 resolve correctly. Previously every agent read as "not installed" on Windows and the TUI showed "No agent sessions found" even when sessions existed on disk.
• UTF-8 round-trip in PowerShell wrapper — wrapper reads AGF_CMD_FILE with -Encoding UTF8, fixing CP949/CP1252 mojibake on Windows PowerShell 5.1 (Korean Windows etc.) so non-ASCII project paths Set-Location correctly.
• CI runs on Windows — windows-latest job added so PATHEXT stemming, PowerShell command synthesis, and the wrapper UTF-8 contract cannot silently regress.
• Release ships Windows x86_64 binary — release.yml matrix gains x86_64-pc-windows-msvc; tagged releases now include agf-x86_64-pc-windows-msvc.zip alongside the existing macOS/Linux tarballs.

Changed

• CommandShell::from_env cached via OnceLock — the wrapper sets AGF_SHELL once before exec'ing agf, so the value is immutable for the process lifetime; caching collapses repeated env lookups on the TUI render path and at every action::* call.

[0.9.0] - 2026-04-21

Fixed

• UTF-8 panic on non-ASCII paths — gemini.rs hashed-dir slicing, codex.rs title truncation, list.rs/stats.rs/watch.rs column truncation used byte-based slicing and crashed on Korean/emoji/CJK input. All paths now char-safe via shared scanner::char_prefix + local char-based truncate helpers.
• Scanner panics silently dropped — scanner::scan_all replaced unwrap_or_default on JoinHandle::join with AGF_DEBUG=1-gated stderr logging. Crashes are still non-fatal but now diagnosable.
• Cache staleness broken for nested sources — cache::get_max_mtime now recurses via walkdir (max depth 4). Previously, file writes inside ~/.codex/sessions/<date>/, ~/.gemini/tmp/<dir>/chats/, and ~/.cursor/{chats,projects}/* did not bump the top-level dir mtime, leaving the cache permanently stale for Codex/Gemini/Cursor.
• Cache/config write corruption — atomic write-then-rename for sessions.json and config.toml prevents truncation on concurrent agf invocations or ^C.
• Gemini 64KB UTF-8 cut — String::from_utf8_lossy over the hard-capped buffer could slice a multi-byte char. Buffer is now trimmed to the last valid UTF-8 boundary before decode.
• Selection lost on sort — apply_sort now snapshots the selected session_id and restores cursor position after reorder.
• Resume bypassed PermissionSelect via number key — ActionSelect number keys 1-9 (and mouse click) now route Resume through the PermissionSelect/ResumeSelect flow, matching the Enter path.
• DeleteConfirm Up/Down toggled Yes/No — only Left/Right, h/l, and Ctrl-h/l now toggle the horizontal choice.
• Preview dismissed on any key — only Esc and Left dismiss; Up/Down (and Ctrl-p/n/k/j) now cycle to prev/next session without leaving preview; other keys are no-ops.
• watch thread leak — refresh thread now gated by an AtomicBool; slow scans no longer accumulate threads every interval.
• agf setup loose detection — precise sentinel # agf - AI Agent Session Finder replaces contains("agf init"), avoiding false positives from user comments.
• agf setup <unknown-shell> exit code — now returns non-zero instead of silently Ok(()).
• Silent config parse failure — Settings::load now prints [agf] config parse error at <path>: <err> — using defaults instead of discarding pins silently.
• Cache version/parse failures — AGF_DEBUG=1 now logs why cache was discarded (version mismatch or parse error).

Changed

• Startup flash eliminated — main.rs enters the alt-screen (and hides the cursor) via a RAII guard before cache load and scan. Previously, cold-cache first-runs showed the shell prompt for 200ms–3s while scanning; now the terminal switches immediately and the scan runs under the TUI surface.
• which fork storm removed — is_agent_installed replaces 7 per-launch subprocess calls with a single cached $PATH directory walk via OnceLock. ~50ms saved on every startup.
• No Session clone per keystroke — FuzzyMatcher::filter signature is now (&[Session], &[usize], query, ...); the TUI passes the agent-filtered indices directly instead of cloning Session values into a subset vec.
• name_col_width cached — computed once in update_filter and invalidated on sort/delete, not recomputed per-frame.
• agents_with_sessions uses agent_counts — avoids an O(N) walk through every cycle of the agent filter.
• Cache + scan honor installed_agents() — uninstalled agents no longer burn a thread, syscalls, or cache slot on every launch. Filter applied uniformly at cache, scanner, and TUI layers.
• scan_stale_agents dispatch — direct match agent → scanner::*::scan() instead of iterating plugin::all_plugins() inside each spawned thread.
• Stats labels — Today / This week / This month → Last 24h / Last 7d / Last 30d to match the actual rolling-window semantics.
• Stats comment drift — "most common agent for color" corrected to reflect the first-seen behavior it actually implements.
• watch process detection — pgrep -f → pgrep -x so running agents are matched by exact binary name, not by any cmdline containing the string (editors/greps no longer false-positive).
• Redundant Settings::load — App::new now accepts a Settings parameter instead of re-reading the config file.
• Settings::save_editable + cache writes — both use atomic tmpfile + rename.

Added

• Shared scanner helpers in src/scanner/mod.rs: char_prefix, read_first_line, first_line_truncated. Removed duplicated file-reading and truncation logic from codex.rs/pi.rs/others.
• AltScreenGuard (RAII) in main.rs — ensures the alt-screen is left and cursor restored even on early-exit paths.
• decrement_agent_count() helper in TUI — keeps agent_counts consistent after single/bulk delete, which fixes the agent filter showing deleted agents.
• debug_assert! in delete_session — defense in depth against session_id values containing / or ...

Removed

• Per-project git-branch thread + 100ms timeout — claude::read_git_branch now just fs::read_to_strings the ~30-byte .git/HEAD. The timeout-per-project was overhead, not safety.
• any_key_pressed helper — preview no longer dismisses on arbitrary keys; helper deleted.

[0.6.4] - 2025-03-20

Changed

• SLT upgrade: v0.6 → v0.15 — major TUI library upgrade bringing 9 minor versions of improvements.
• Rounded borders — filter bar and bulk-delete header now use Border::Rounded with colored borders for a modern look.
• Native separators — replaced manual "─".repeat() with SLT's separator_colored().
• Native help bars — all footer keybinding hints now use SLT's help() widget for consistent styling.
• Responsive breakpoints — compact layout now uses ui.breakpoint() instead of manual width checks.
• Inline text with line() — preview details, headers, and info rows use line() for proper inline text rendering.

Added

• Agent filter badges — agent filter indicator now uses badge_colored() / badge() widgets.
• Empty state — shows a friendly "No sessions found" message when search/filter returns no results.
• Section dividers in Help — help screen uses divider_text() for section headers.
• Key hints in Help — keybindings displayed with key_hint() widget for visual distinction.
• Terminal title — window title set to "agf" via RunConfig::default().title().

Removed

• Dead legacy files — deleted unused tui/input.rs and tui/render.rs (ratatui/crossterm remnants).

Fixed

• 49 #[must_use] warnings — all unused Response returns from SLT 0.11+ properly handled.
• Clippy clean — resolved 17 clippy suggestions (collapsible if-statements, redundant imports).

[0.6.0] - 2025-03-14

Breaking Changes

• TUI engine: ratatui → SuperLightTUI (SLT) — complete rewrite of the rendering layer from ratatui's retained-mode to SLT's immediate-mode architecture. Same look, fewer dependencies, simpler code.
• Shell wrapper updated — the wrapper now uses a temp file instead of stdout capture. Run agf setup again after upgrading, or restart your shell.
• Keybinding change: summary cycling changed from Shift+↑/Shift+↓ to [/].

Fixed

• Scanner hang on unreadable .git/HEAD — read_git_branch() now has a 100ms timeout per path, preventing infinite blocking when a git directory is on an unresponsive filesystem.

Changed

• ratatui and crossterm dependencies removed; replaced with superlighttui v0.6.
• TUI source consolidated from 3 files (~2,350 lines) into a single tui/mod.rs (~1,850 lines).
• Shell wrappers (zsh, bash, fish) use AGF_CMD_FILE temp file for command passing instead of stdout capture.

[0.5.5] - 2025-03-10

• Resume mode picker with Tab on the action menu.
• Parallelize worktree scanning for faster startup.