feat: token and cost details on agents view rows

[snimu]

Every agents-view row can now answer "what did this session cost" at a glance, in one unconditional format: ↑12k ↓1.2k · $0.42 ($1.10 w/ subagents). ↑ is input/sent tokens, ↓ output/received; $0.42 is the session's own spend and the parenthesized total is the whole subtree. Every row renders the full cell, zeros included — a session that never spent reads ↑0 ↓0 · $0.00 ($0.00 w/ subagents).

The chosen semantic is whole-file spend. A row's own cost is everything the session ever spent — every branch, including forked-away ones, and the summarization calls themselves: compaction (both split-turn slices) and branch-summary responses persist their usage on the entry they produce, and both producers fold it in — with child_usage_attributed entries subtracted so a child's cost is never counted twice. That is the money answer ("what did this session cost in total"), deliberately broader than /usage, which answers the context question for the current branch. Resident and on-disk rows compute the identical number — pinned by a resident-vs-scan equality test on a file with a forked-away branch — so a row never shifts at passivation or revival.

Mechanics: resident rows use a memoized getOwnUsageSummary() over the in-memory entries (memo keyed on entry count and tail id; measured cold cost 0.27 ms per call at 10k entries). On-disk rows accumulate the same total inside the catalog's existing scanSessionInfo pass (already line-parsing every file, cached by size+mtime) — no new IO, no new cache. Recursive cost is own plus all descendants, summed in the view over the unfiltered hierarchy, so search and scope filters never shrink the parenthesized total; the traversal is iterative and pinned overflow-safe at 10k depth. Passivated descendants stay counted across restarts: both list_saved_sessions composition points (supervisor and worker daemon) append live spawn-ledger children the directory scan missed — their transcripts live in session-artifacts, which the scan never visits — through one shared walk (withPassiveRlmDescendantInfos: one readSessionInfo per descendant, parentSessionPath/rlmDepth backfilled from the edge, tombstoned edges excluded), so an inactive parent's subtree total is identical before and after a supervisor restart. On the wire: an optional usage field (inputTokens, outputTokens, cost) on session summaries and saved-session rows, DAEMON_SCHEMA_REVISION 26. Old daemons produce rows without the field; the view renders the same cell with zeros.

Deliberate changes and known limits: the "N turns" message-count detail on inactive rows is removed — the details cell is now usage · age. Entries above the scan's max line length keep their existing skip, so one multi-megabyte assistant message can undercount slightly (giant tool results carry no usage and were always skipped). A failed compaction persists no entry, so its billed tokens are knowingly unrecoverable.

Stacked on fix/agents-view-busy-descendant-indicator (#1986); draft until that merges, then retargeted to main.

Validation: every behavior across the review rounds keeps exactly one pin, on consolidated fixtures. One format test covers all four row shapes (and message-count removal); one forked-file fixture pins scan/resident equality, attribution subtraction, and compaction + branch_summary billing terms as one pair of exact totals; one hierarchy fixture pins the unfiltered rollup with own costs intact plus the 10k deep chain; one producer test covers active and saved rows; one supervisor-restart fixture pins restored catalog rows with ledger-edge topology beating a forked-away transcript header (plus tombstone exclusion); one two-family worker fixture pins restart rows through the worker handler, per-sessions-dir ledger isolation, and broken-ledger degradation; the view keeps the money-across-restart pin (recursive total identical for a live row vs a catalog-only row) and the no-duplicate merge pin; the real-summarizer compaction test pins that own spend grows by exactly what the entry recorded (and memo invalidation). All were proven fail-unfixed against the head that preceded their round — except the no-duplicate pin, whose input cannot exist pre-fix and which is mutation-verified against a broken alias merge; consolidation was re-verified by mutation on three survivors from different rounds (dropped parenthesized total, dropped scan summarization fold, header-first topology — each fails exactly its pin). Catalog-load cost of the ledger merge, measured at 300 saved sessions with 36 passivated descendants, cold ledger: scan 33 ms, merge +9 ms (~0.25 ms per descendant; zero descendants ≈ zero added cost). agents-view suites, rlm-ledger, agent-roster, daemon-agent-roster, subagent-summary-line, daemon-session-list, session-manager file-operations, daemon-protocol (schema sync), daemon-mode, daemon-supervisor, daemon-stop-confirm, daemon-launch/client/routed-client, package-self-update, agent-connection-daemon — green in a sanitized env; root npm run check green.

Net src: +267/−23 (test: +358/−2) — the formatter/producers/rollup plus compaction/branch-summary billing (RES-1258) and the catalog merge with its review-round hardening (RES-1262); all composition over existing scans, folds, and walks; one new exported helper, no new state or caches.

Linear: RES-1258 https://linear.app/primeintellect/issue/RES-1258
Linear: RES-1262 https://linear.app/primeintellect/issue/RES-1262

---

Note

Medium Risk
Touches session accounting, catalog scanning, and daemon protocol revision 26; incorrect usage rollup or ledger merge could misreport costs or omit passivated subagents from the roster.

Overview
Agents view rows now show input/output tokens, the session’s own dollar cost, and a parenthesized recursive total including subagents (e.g. ↑12k ↓1.2k · $0.42 ($1.10 w/ subagents)), replacing the inactive-row message count. Recursive totals are computed on the unfiltered session tree so search/scope does not shrink parent totals.

Billing semantics treat a row’s “own” cost as whole-file spend: assistant usage on the current branch, plus compaction and branch-summary LLM calls (usage is captured and stored on those entries), with child usage attribution subtracted so subagent spend is not double-counted. Live sessions expose this via memoized getOwnUsageSummary(); catalog scans compute the same totals during the existing line walk.

Daemon / wire: optional usage on session and saved-session summaries; schema revision 26. Saved-session listing merges passivated RLM descendants from the spawn ledger (withPassiveRlmDescendantInfos) so artifact-only child sessions and their costs still appear after restart.

^{Reviewed by Cursor Bugbot for commit af309b8. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Show token and cost details on Agents View rows

• Agents View rows now display input tokens, output tokens, direct session cost, and recursive subagent cost, replacing the previous message-count detail
• Usage tracking flows through the full pipeline: compaction and branch-summary entries now persist their LLM Usage, scanSessionInfo aggregates assistant plus summarization usage minus child attribution into a SessionUsageSummary, and the daemon protocol (revision 26) carries it to clients
• computeRecursiveCosts walks the unfiltered unified-session hierarchy to produce descendant-inclusive cost totals for each row
• Daemon list_saved_sessions now merges passive RLM descendants using a per-directory ledger (rlmSpawnLedgerFor) instead of always using the default ledger
• Behavioral Change: rows without usage data now render zero-valued token and cost fields where the message count previously appeared; daemon schema advances from revision 25 to 26

^{Macroscope summarized af309b8.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 03e09d4. Configure here.}