BACKLOG.md

BACKLOG

Reconciled 2026-06-19. Status markers: DONE / PARTIAL / open (no marker). CLAUDE.md §7 build phases (0–5) are all complete; everything here is post-phase work.

Bug: Auto Height - DONE

IF the parent project has a lot of repos, then the herder bar can extend too far when trying to start a chat window. The auto heigh feature should not count unactive repos, or repos without any active agents in them. The auto hieght features should only make room for showing repos with active agents, leeaving the rest of the screen space to the change window

Resolved: stage_top_rows() counts only agent-bearing project rows (test stage_top_rows_ignores_idle_zero_agent_projects).

Feature: Tabbed Agents - DONE

When an agent is active in a project, I want it indnetd so I can clearly see that the agent is under that project. If an agent opens sub agents under it's perview, if possible I want those indended in that agent which is indented under the project.

Resolved: agent rows are now indented (AGENT_INDENT) under their project header in grouped view, and sub-agents nest one level deeper still (their ├─/└─ tree glyphs were already there), so the roster reads project → agent → sub-agent. Indent is suppressed in flat view (no header to nest under). ui/table.rs (session_row/subagent_row).

Bug: Status and Context is sometimes incorrect - DONE

I select a row in herdr, I press n to start an agent, it sarts and in herdr is report 15% context used up and processing already? I haven't even asked a question yet

"Processing already" half resolved: instantaneous CPU now seeds idle on a fresh agent's first sample instead of inheriting ps %cpu's lifetime average (fix: 1e54cd7). The "15% context" half is not a bug: that's real consumed context — the first API call already includes the system prompt, tool definitions, CLAUDE.md, and env, which on a 200k model lands ~15% before you type anything (monitor.rs:146 tracks the last call's input+cache tokens). Left as-is rather than hiding accurate data. See the herdr-status skill.

Feature: Better colors - DONE

The hightlight row color is too bright and hard to read. The colors are overall could use some work. Build a theme system were we can slot in different colors for the UI. The default color should resemble that of Dracula Dark

Resolved: the Dark theme is now built from the Dracula palette (DRAC_* consts in core/theme.rs), and the selected row uses a muted Dracula "current line" band (selection_bg/selection_fg) instead of full-brightness reverse-video — the "too bright" complaint. The theme system was already slottable (Dark/Light/None via ThemeMode); None keeps reverse-video for NO_COLOR. See theme.rs + ui/table.rs.

Bug: Slow response on Status and Activity - DONE (Status); Activity open

Both the Status and Activity readouts in herdr seem to lag behidn what is actually happening with the agent. Often times staying on processing while the task has compelted. Activity often lags behind or does not repreesnt the token consumotion or what is going on.

Status lag resolved: instantaneous CPU (ps time= deltas, not lifetime %cpu) + the user/tool_result decay branch, so a finished agent leaves Processing within ~1–2 ticks (fix: 1e54cd7, perf: 8b191cd). Still open: the Activity sparkline (activity_history) accuracy vs. real token throughput.

Bug: The $/hr is too small - DONE

I often see values of 6000/h or 4000/h. It seems to me that those should be the red values

Resolved with "$/h is wild" below: those 6000/h readings were the bug, not real rates — a cost delta divided by a hardcoded 2s tick during a sub-second event-loop burst. Now the rate is Δcost ÷ real wall-clock, so magnitudes are sane and the existing >10 → red threshold in ui/table.rs colors a genuinely hot agent red. Threshold coloring (low/mid/high) was already wired to the theme.

Bug: Context should reset on /clear - DONE

using /clear on an agent window should reflect in herdr that the context has been cleared

Resolved before this session: fix: e048a41 (attribute transcripts by session id; reset context on /clear) + test merge_resets_transcript_state_when_session_id_changes.

Feature: Claude window border - DONE

Can tmux add a border to the claude window just as herdr has it's own window title and border ie -herdr-----

Resolved by composing tmux (no rendering, §0.1): when an agent is staged, herdr sets pane-border-status top + a #{pane_title} border-format on its window and titles the agent pane "Claude — " (and its own "herdr"), plus a heavy border line and Dracula colors (pane-active-border-style purple = focused, pane-border-style comment = idle). The agent's top-edge border is the divider under herdr, so it reads like -herdr-. Options unset on unstage. terminals::set_stage_title/clear_stage_title → tmux::set_stage_border. Note: the left/right/bottom edges stay unbordered — they're the terminal's outer edge, where tmux draws nothing, and herdr can't render a box into Claude's pane without becoming a terminal emulator (§8). A full 4-sided frame isn't reachable within the design.

Feature: Persistent cost - DONE

A value that reflects the cost over multiple (all) sessions. For the repo collection as a whole and individual projects.

Resolved: history::all_time_summary() rolls the full history CSV into collection-wide totals and a per-project map (ProjectTotal). The fleet strip shows all $X (tokens) for the collection; each project header shows its all-time spend as Σ$X (matched by name; shown even for idle projects with past sessions). Cached on App.all_time_summary, refreshed on the weekly-summary cadence (~30s). Limitation: per-project match is by display_name, so a custom session name won't fold into the dir.

Feature: Quickstart - DONE

A script to quickly launch tmux and start herdr tmux attach -t work cargo run --release -- "/mnt/c/Users/Ben Bracamonte/Work"

Resolved: ./quickstart.sh [PARENT_DIR] — rebuilds release (never a stale binary), then creates or reuses a tmux session (default work, override HERDR_SESSION) with herdr running over PARENT_DIR (defaults to the repo's parent dir; handles paths with spaces). Switches client if already in tmux.

Feature: Add Mouse Scrolling - DONE (tmux-native)

set -g mouse on Mouse scroll works on roster page but in agent window it scrolls the messages i eneted. Can we make the scroll wheel scroll the responses like a normal agent window

Resolved the tmux-native way: quickstart.sh now runs tmux set -g mouse on, so wheel-scroll over the staged agent pane enters tmux copy-mode and scrolls its responses. herdr deliberately does not call crossterm EnableMouseCapture — that would fight tmux's mouse handling in herdr's pane and lose native text selection (§0.1, tmux owns window management). Add set -g mouse on to ~/.tmux.conf to make it permanent.

Feature: Mouse Focus - DONE (tmux-native)

Is this possible without betraying our design doc? I want to be able to click between the herdr window and the chat window

Yes, without betraying the doc: it's tmux's set -g mouse on (now set by quickstart.sh) — click a pane to focus it. herdr stays out of mouse capture so tmux owns click-to-focus, fully aligned with §0.1. Pairs with "Add Mouse Scrolling" above.

Features: Charts - DONE (v1)

What kind of charts would be informative that we could display along side ond beneath the herdr nav bar?

v1 shipped: the Phase 5 fleet trend strip beneath the roster (G toggles) — status counts, live burn $/hr + a trend sparkline, today/week cost (feat: bb8ec94, ui/fleet.rs). Per-agent values stay in the roster (Context bar + Activity sparkline columns). Optional next: daily-cost BarChart or a dedicated graphs overlay if the strip earns the space.

Feature: wk token tracker - DONE

We need a way to track token usage over multiple sessions

Resolved: the fleet strip now surfaces wk tokens alongside wk cost, plus all-time tokens in the new all-time segment (ui/fleet.rs, fmt_tokens). Multi-session token totals come from history::weekly_summary.total_tokens and history::all_time_summary().total_tokens.

Bug: Asking for permission in roster doesnt work - PARTIAL

when waiting for an approavl to run a tool, there is no indicator in herdr roster view that it needs attention or can be arrpoved. Sometimes I see a purple Need Input notice, but then it kept going. maybe becaues I was on auto mode?

Partial: Phase 4c added the approval inspector (A) — tmux capture-pane shows the real permission dialog (invisible in JSONL) with approve/deny/interrupt from the roster (feat: 581c86d). NeedsInput detection also tightened via the status work. Persistent indicator now added (ui/table.rs): a bold ⚠N badge on any project header hosting N agents awaiting approval, plus a leading ⚠ on the agent's Status cell (visible in NO_COLOR too). Still open: the auto-mode "kept going" interaction.

Bug: $/h is wild - DONE

The nuimber is up and down and all over the palce. I feel like it should be the average of an hour? or er min?

Resolved: burn rate is now Δcost ÷ Δ(real wall-clock), EMA-smoothed (smooth_burn, α=0.3), sampled no faster than every 2s (MIN_BURN_SAMPLE_MS) — mirroring the CPU instantaneous-rate derivation. The old delta * 1800.0 assumed a fixed 2s tick, but the Phase-2.5 event loop fires at irregular, sub-second intervals, so a JSONL burst spiked $/hr into the thousands and quiet ticks decayed it erratically. New prev_cost_sample_ms field on the session carries the sample baseline. Tests: smooth_burn_steadies_a_spiky_signal, smooth_burn_decays_toward_zero_when_idle. (app.rs refresh).

Feature: Model identification - DONE (detail view)

Each agetn should have a indicator about which model they are current running.

Resolved via the existing roster detail (Enter → ui/detail.rs shows "Model: …", already shortened by models::shorten_model). Per user decision, deliberately not added as a roster column to keep the row uncluttered.

Bug: Reszing the window changes the agent window size permenantly - DONE

When reszing the window smaller they resize properlly. When resize the window larger with both the roster and the agent window open, then the agent window stays the compressed size while herdr gets bigger. Upon getting larger the agent window should take more of the screen space when possible while still maintaiing the min-height for the roster

Resolved: Event::Resize only repainted — resize_stage_top was never re-issued, so tmux kept the old height and a grown window left the agent pane compressed. The event loop now calls App::refit_stage() on resize, which re-fits herdr to its roster height (capped to preserve the agent's min) when an agent is staged. Gated on "staged" so it never fights a manual Ctrl-b resize. (src/main.rs, app.rs.)

Bug: Roster selection bar should stay with last repo selected when launching an agent. - DONE

Launhcing an agent in the roster view with n. The repo will sort to the top with the now active agent, leaving the roster selection on an unintended repo. The selection bar should follow the now sorted repo.

Resolved: selection is now sticky by stable identity, not row index. refresh() captures a RosterSelKey (project path for a header, PID for an agent) before the re-sort and re-anchors the cursor onto it afterward (selection_key / reselect_by_key). So after n the cursor follows the launched repo to the top — and the roster no longer drifts under you on any status-driven re-sort. Test reselect_by_key_follows_the_same_entity_across_a_resort.

Bug: Reoster bar is too small with only one or two active agents - DONE

The roster bar should be a minimum for 12 rows when there are a small number of active agents. Ruight now there is one active agent and an open chat window and I see 7 rows. Header, Columns names, 2 repos, session overview and fleet bar.

Resolved: stage_top_rows() floor raised 6 → 12 rows. terminals::cap_stage_top_rows still trims it on short terminals so the staged agent pane keeps its minimum, so the floor never starves the agent.

Feature: Split waiting states - DONE

If the agent is waiting for a repsonse from the API, the roster displays Waiting. The a task is complete the roster also dispalys Waiting. When a task is complete and the agent is waiting for the next task from the usr, then I want that task to be Job Done. So we can tell when we are waiting on the API and when we are waiting ont he user

Resolved: new SessionStatus::JobDone ("Job Done", cyan). infer_status now splits the two: assistant end_turn (turn complete, your move) → Job Done; a user/tool_result line where Claude still owes output (request in flight) → Waiting. (monitor.rs; tests assistant_end_turn_reads_as_job_done_not_waiting + the existing user-branch tests.)

Feature: Error Display - DONE

Somtimes we hit an API error. The Roster should display Error in the stats when this is encounted

Resolved: new SessionStatus::Error (red, sorts just below Needs Input). Claude Code writes API errors as a <synthetic> isApiErrorMessage line ("API Error: 529 Overloaded…"); transcript.rs emits a TranscriptEvent::ApiError, monitor.rs keeps it sticky (last_was_api_error) until a newer message supersedes it, and infer_status reports Error when CPU is low (an active retry stays Processing). Tests api_error_flag_reads_as_error_when_idle, active_retry_after_error_reads_as_processing.

Bug: Sparklines - DONE

The per agent sparklin looks like it only has two states, on and off The fleet sparkline has all sorts of varitaion. Make the roster sparkline more granula like the fleet sparkline

Resolved: activity_history now stores a continuous CPU% per tick (Vec<f32>) instead of a discrete status level that only ever produced 1–2 bar heights. format_sparkline scales each bar against a 100% ceiling (a multi-core spike clamps to full, idle ticks blank), giving fleet-strip-like granularity. Test sparkline_scales_cpu_to_block_heights. (session.rs.)

Bug: Opus 4.8 (and unrecognized models) over-priced ~3× - DONE

Surfaced by the comparable-tool research (docs/COMPARABLES.md, 2026-06-20). models.rs::shorten_model collapses any …opus-4-8… (or 4.5/4.7) to "opus", which resolves to the opus-4.6 profile = $15/$75 per M. Real Opus 4.8 is ~$5/$25 per M. Every model that isn't opus/sonnet/haiku also lands on the Opus fallback profile, over-counting cost. herdr is the binary you run on Opus 4.8 and it mis-costs itself.

Resolved: rewrote built_in_profile to current Anthropic list pricing (Opus 4.5–4.8 $5/$25, Sonnet 4–4.6 $3/$15, Haiku 4.5 $1/$5; cache read 0.1× / 5m write 1.25× input — verified against the claude-api reference), matched by model family prefix so every version/date/[1m] suffix resolves. Replaced shorten_model's opus-only special-case with a boundary/version-aware matcher (family-major.minor, split on non-alphanumerics so date stamps and [1m] don't confuse it). The fallback now prices at mid-tier (Sonnet) not Opus, still flagged cost_estimate_unverified. Tests opus_48_priced_at_current_rate_not_legacy, shorten_model_extracts_family_and_version (models.rs).

Bug: Token over-count from streaming-duplicate JSONL lines - DONE

Surfaced by the same research. Claude Code writes the same assistant message's usage on multiple physical JSONL lines during streaming; herdr accumulates every usage line (no dedup), so multi-line turns inflate token + cost totals. Both ccusage (mod.rs:292) and tokscale (claudecode.rs:498) document and handle this.

Resolved: transcript.rs now captures message.id + top-level requestId; monitor.rs keys each usage line by (message.id:requestId) and max-merges per tier (merge_usage), adding only the increase over any prior emission to the running totals — on both the own-agent and subagent paths. Dedup state persists across incremental ticks (carried by merge_discovered_session), resets on /clear (fresh session) and on file truncation. Tests merge_usage_dedups_streaming_duplicates, merge_usage_keyless_lines_add_in_full (monitor.rs). Deferred (lower value once the table above is correct): CostMode::Auto (trust costUSD — present only sporadically in transcripts, and our computed cost now matches it); the 5m/1h cache-creation split (real but ~1.6× on a minor tier); a >200k long-context tier (N/A — current Opus/Sonnet 1M context is flat-priced, no premium). User-configurable price overrides need a config-file subsystem herdr doesn't have yet (models::set_overrides is wired but unused).

Feature: Inbound permission-prompt hooks (status from fact, not guess) - DONE

herdr infers the invisible permission-prompt NeedsInput state (<2% CPU + stale tool_use). Claude Code's Notification hook fires exactly when an agent needs permission/input — make it authoritative. (COMPARABLES §7 item 5; whiteroomed from claude-tui's hooks-as-push-signal.)

Resolved (Phase B): opt-in, merge-safe. herdr hook install merges Notification/Stop hooks into ~/.claude/settings.json (idempotent; preserves the user's other hooks; herdr hook uninstall reverts). The hook runs herdr hook notify, which reads the payload on stdin and atomically writes ~/.claude/herdr/<session_id>.json (Notification→NeedsInput, Stop→JobDone). herdr's notify watcher now also watches ~/.claude/herdr/, and monitor::apply_hook_override treats a fresh status file (newer than the transcript mtime, <5min old) as authoritative — self-clearing once the agent writes new JSONL, so no one has to delete the file. Works without hooks installed (pure no-op). New core/src/hookstate.rs + src/hookcmd.rs; tests state_and_settings_under_temp_home, read_missing_is_none; verified end-to-end against a temp HOME. Usage: run herdr hook install once, then restart Claude Code sessions. Deferred: agent-deck's "acknowledged-downgrades-waiting" (stop nagging once you've attended an agent) — the override already self-clears on the agent's next JSONL line, so this is a minor polish; and stale-file cleanup for dead sessions.

Feature: Phase C tmux orchestration — INVESTIGATED, mostly N/A for herdr

COMPARABLES §7 items 6–7 (tmux -C control-mode refresh + low-latency send-keys; orphan-pane recovery) were borrowed from agent-deck. On investigation against herdr's actual architecture, both are obviated — herdr is not agent-deck:

• Orphan recovery — already covered. ~/.claude/sessions/<pid>.json is written by Claude Code itself (v2.1.183: sessionId/cwd/startedAt/status/name/…), so herdr's discovery::scan_sessions already finds every running agent — herdr-launched or manual — and maps each to its tmux pane via TTY (ps → pane_for_tty). agent-deck needs orphan recovery only because it owns named agentdeck_* tmux sessions; herdr owns none (§0.1/§8) and discovers via a restart-independent registry. No gap.
• tmux -C control mode — doesn't earn its complexity. herdr's refresh signal is JSONL writes, caught reliably by the notify watcher (transcripts live on the Linux fs ~/.claude/projects, not the flaky /mnt/c mount), and herdr sends whole prompts, not keystrokes — so neither the %output refresh nor the persistent-send-keys benefit applies. A persistent tmux -C subprocess
  • protocol parsing would add real complexity (and edge toward §8) for ~zero gain. Skipped per the plan's "commit only if it earns its complexity" gate.

Shipped the one real adjacent win the investigation surfaced: herdr parsed past Claude Code's session name (it read only pid/sessionId/cwd/startedAt). discovery::scan_sessions now lifts name into a new ClaudeSession.cc_name, shown read-only in the detail view under Project (e.g. "audit-model-pricing-bugs") so multiple agents in one repo are tellable apart. Kept separate from display_name()/session_name so it never disturbs history's per-project cost folding. Follow-up (UX decision, not done): promote cc_name to the roster agent-row label and the staged- pane border title.

Feature: Phase D graphs — daily activity heatmap (D1 done; D2/D3 deferred)

COMPARABLES §7 items 8–10. Applied the plan's own gate ("each graph reflects real data and is something we actually glance at; defer any that don't earn the space").

Shipped D1 — daily activity heatmap on the fleet strip (G). New history::daily_cost_series

• history::intensities (the 5-level GitHub-contributions bucketer vendored from tokscale's calculate_intensities: ≥0.75→4 … >0→1, else 0) → history::daily_activity(ACTIVITY_DAYS=14). Cached on App.daily_activity, refreshed on the weekly_summary cadence (~30s) so it never reads history.csv on the render path. ui/fleet.rs renders one cell per UTC day (· ░ ▒ ▓ █, shaded by that day's spend vs. the busiest day) — the "days" time axis the live burn sparkline (~30s) can't show. Tests intensities_bucket_by_ratio_to_busiest_day, daily_cost_series_length_matches_window, heatmap_maps_levels_to_shades. (Populates as sessions complete — same history.csv the all-time/ weekly totals already use.) Deferred with rationale:

• D2 (P90 personalized limit + 5h-block projection) — would introduce a 5-hour "block" entity in tension with the project-first model (CLAUDE.md §2); the plan-limit tables are community guesses; and the JSONL limit-notice regexes can't be verified without a real "limit reached" sample. Heavy + speculative for a narrow quota-prediction payoff.
• D3 (turns-until-compaction / wasted-context) — value dropped sharply with 1M-context models (auto-compaction is now rare), and the detail view already shows context %/bar. Marginal vs. effort. Follow-up if wanted: a full multi-row GitHub-style heatmap overlay (the intensities bucketer is already in place and tested, so it's mostly render plumbing).

Feature: Phase E roster UX cheap wins (E1 + E2 done; E3 deferred)

COMPARABLES §7 tail.

Shipped:

• E1 — Error badge on project headers. The header already flags blocked agents (⚠N); now it also flags errored ones with ✕N (red, t.status_error), shown only when a project hosts an agent in SessionStatus::Error — same tasteful, only-when-relevant pattern. (Chose this over the full per-status count breakdown agent-deck shows on collapsed groups: herdr always shows the agent rows, so a full breakdown is redundant clutter — and the user keeps the roster lean.) ui/table.rs.
• E2 — status filter quick-keys. !=needs-input @=processing #=waiting $=idle jump straight to a filter (vs. f which cycles); pressing the same key again toggles back to All for a clean on/off (App::set_status_filter). Documented in the ? help overlay. app.rs. Test status_quick_key_sets_then_toggles_back_to_all. Deferred: E3 (freshness [Nm ago]) — herdr's roster is event-driven (notify watcher + 2s safety tick), so it's near-always fresh; a staleness indicator (valuable for vscode-claude-status's network panel) would read "now" almost always here. Low value.

Bug: If an anget process Job Done then the fleet count should be idle +1 - DONE

what is the difference between idle and waitign anyway? And jobs done? I feel like I am conflating terms

Resolved: fleet_counts (app.rs) lumped Job Done into the wait bucket alongside WaitingInput; it now counts as idle — a finished-turn agent isn't doing anything, it's your move. So wait is only a genuine in-flight API request (WaitingInput). The taxonomy: Processing = working · Waiting = blocked on the API (request in flight) · Job Done = turn complete, your move · Idle = no recent activity (stale Job Done / parked tool call) · Needs Input = approval · Error = API error. Added this glossary + the !!/!F/REC/⚠N/✕N indicator legend to the ? help overlay (ui/help.rs) so the terms are discoverable. Test fleet_counts_job_done_reads_as_idle_not_waiting.

Bug: Approve / Deny not visible on roster view

I think we have the functionlailty but I don't see any notification that an agent is asking for approve deny or some other reponse from the user before proceeding. If its just approve deny then I would like to be able to anserw that from the roster view. Other questions the user needs to go to the claude window for more details

Feature: Color the agents in roster

The agent names, which seem to be just the same as the project that they were stated on are just regualr gray. Can we use a more poppy dracula themeed color? I want those to stand out more

Bug: Is that hooks thing installed? - PARTIAL (install works; bug fixed)

I thgouth that we had a hook to tell when an agent does stuff so we could more relaiably determine when an agent needs direction or is done with the job etc. How do we install that? target/release/herd install ?

Install it with herdr hook install (then restart Claude Code sessions); it merges opt-in Notification/Stop hooks into ~/.claude/settings.json and herdr hook status shows the live state. Fixed a blocker found here: hook_command wrote the binary path unquoted, so Claude Code's /bin/sh -c word-split a path with spaces (/mnt/c/Users/Ben Bracamonte/… → /bin/sh: /mnt/c/Users/Ben: not found) and the hook silently failed. Now shell-single-quoted (shell_single_quote); the bare hook notify suffix lets re-install replace the old broken entry. Re-run herdr hook install once to repair an existing install. Tests quotes_paths_with_spaces, escapes_embedded_single_quotes, command_keeps_bare_hook_notify_suffix. Also wired groundwork for the (deferred) session-limit notice: transcript::maybe_capture_limit_notice logs a real "usage limit" line under HERDR_LOG so a verified regex/badge can follow.

Bug: Sub-agents never appeared in the roster + their tokens vanished - DONE

When an agent spawns sub-agents (e.g. Explore agents in a code review), I want them tabbed under the orchestrating agent with their own cost/context, their token spend rolled into the parent's counters (kept even after they finish and disappear), and reflected in the today/wk/all fleet counters — which were also constantly 0.

Root cause: the sub-agent structs (SubagentRollup/subagent_breakdown/subagent_row) were all built and correct, but fed by a dead discovery path. discovery::scan_subagents scanned /tmp/claude-{uid}/{slug}/{sessionId}/tasks/*.jsonl ("Feature #29", inherited from claudectl) — but Claude Code v2.1.x stores only .output scratch there. Real sub-agent transcripts live at ~/.claude/projects/{slug}/{sessionId}/subagents/**/agent-*.jsonl (+ agent-*.meta.json). So active_subagent_count was always 0; no rows, no token attribution. Measured on live data: 409M tokens across 491 sub-agent transcripts (2.45M of it output) were being silently dropped. Fixes:

• scan_subagents now derives the dir from jsonl_path (resume-safe) and recursively collects agent-*.jsonl (skips .meta.json), splitting all-discovered (subagent_jsonl_paths, drives rollup so a sub-agent that came and went is still counted) from the fresh-by-mtime active subset (active_subagent_jsonl_paths, < SUBAGENT_ACTIVE_SECS=25s → individual live rows; older collapse to "completed (N)"). discovery.rs.
• monitor::refresh_subagent_rollups iterates the full set; update_subagent_rollup resolves a human label once from the agent-*.meta.json sidecar (agentType · description, e.g. "Explore · Map launch→discovery"). Rollup tokens/cost already fold into session.cost_usd/totals via finalize_usage, so the parent agent row + fleet live-cost already include sub-agents.
• Fleet today/wk/all were CSV-only (history.csv, written only at Finished) → 0 for never-finished agents. New App::fleet_totals() folds live (active-session) cost+tokens into today/wk/all (skips Finished to avoid double-count); wired into ui/fleet.rs strip + ui/table.rs title. Tests: scan_subagents_finds_agent_jsonls_and_splits_active_by_mtime, subagents_dir_derives_from_resolved_transcript_path, scan_subagents_clears_when_no_subagents_dir, resolve_subagent_label_reads_meta_sidecar, truncate_label_*, fleet_totals_fold_live_spend_and_skip_finished. Deferred: per-sub-agent context % (the rollup accumulates deltas, not last-call absolute context, so the sub-agent rows show cost/tokens but - for context); a cap on simultaneous active rows if a wide fan-out ever floods the roster (YAGNI).

Bug: Total cost over day/week and heatmap do not persist across sessions - DONE (already shipped)

Those totals should be stored somewhere locally and loaded in when the app starts. I would like to keep that tally running

Already implemented in 77437e5 ("Bug fixes", 2026-06-21): a session-keyed ledger at ~/.local/share/claudectl/ledger.json (core/history.rs, struct Ledger). It's loaded on startup (App::with_parent → Ledger::load, seeded once from the legacy history.csv on first run), upserted from every live session each tick, and atomically saved (tmp+rename) every ~30s and on quit (save_tally, wired at main.rs:193). today/wk/all-time and the 14-day heatmap all derive from it (weekly_summary/all_time_summary/daily_activity), so the running tally survives restarts. If you still saw it reset, you were on a stale binary — rebuild via ./quickstart.sh or cargo run --release. Verified by the workspace test suite (245 green).

Bug: No more auto sort active agents - DONE (already shipped)

Once a repo has an active agent in it, don't auto store the roster by statss or whatever. Sort it alphabetically by repo name once there is an active agent in one and it has been brought to the top of the roster. Don't auto sorts those at the top over and over, just alphebecially

Already implemented in 77437e5: App::project_groups (app.rs:4189) sorts agent-bearing repos to the top, then stable alphabetical by name — and deliberately not by status/urgency/cost, so a repo's position no longer re-floats as its agent churns NeedsInput→Processing→Idle. Ordering only shifts when a repo gains or loses its first/last agent. (Default view is grouped; sticky selection via RosterSelKey keeps the cursor on the same repo across any re-sort.) Same stale-binary caveat as above if you still saw rows jumping.

Bug: A see a diaplay of Needs Input then it changes to processing - DONE

Needs Input somteimes shows up in the status. I would think that we have a better way of telling for sure if an agent needs input or not. Right now it seems to infer it incorrectly

Resolved: the flicker was the tool_use branch in infer_status (monitor.rs) flipping to NeedsInput the instant a tool was outstanding with low CPU — which is indistinguishable from an auto-approved tool still executing (the Claude process is idle while a Bash/web call runs), so it flashed NeedsInput then snapped back to Processing when the ToolResult landed. Now an outstanding tool must outlive a normal round-trip (NEEDS_INPUT_STALE_SECS = 10) before the heuristic calls it NeedsInput; within that window it stays Processing. To keep accuracy when the inbound hook is installed, apply_hook_override now lets a fresh Notification (NeedsInput) override that low-CPU Processing too (deferring only to a visibly-busy CPU>5 agent or an active Error), so a real permission prompt still surfaces instantly. Tests fresh_pending_tool_reads_as_processing_not_needs_input, aged_pending_tool_reads_as_needs_input, busy_pending_tool_reads_as_processing. The authoritative path remains herdr hook install (status from fact, not guess).