sessions: fold subagent runs into PR attribution#791
Merged
Conversation
iamtoruk
force-pushed
the
feat/subagent-pr-attribution
branch
from
July 21, 2026 00:54
6c1c2da to
ea5d518
Compare
added 6 commits
July 21, 2026 05:46
Sidechain (subagent) session cost never reached the by-PR view, so a session orchestrated on one model with subagent lanes on another showed only the parent model on every PR row. Fold each sidechain's cost, calls, models, and categories into the parent turn that spawned it, so it inherits that turn's PR set under the existing turn-level state machine. Linkage, in priority order: the spawn result's toolUseResult.agentId pairs the child's agent id with the Agent/Task tool_use id that launched it (recorded per turn), which is the true launch point and wins even when the child's first activity landed during a later turn; else the child's first-activity timestamp is bucketed into the containing turn span; else the child folds into the parent's unattributed spend. Children of parents that referenced no PR, and orphans whose parent is absent from the scan, contribute nothing, unchanged. Cache v6 to v7 (neither shipped, so one combined bump from v5): per-turn spawnToolUseIds, per-file parentSessionId and agentSpawnLinks; the validator and the append/compact paths thread them like prRefs. By-PR footers now count parents plus folded subagent runs and the payload gains an additive subagentSessions field. distinctCost now includes folded subagent spend, documented in the payload comment.
Rework child attribution to resolve each subagent to the PR its launching turn was working on, using the parent's UNFILTERED turn data, and enforce that every dollar is counted exactly once. - Mutual exclusion: a child that referenced its own PR attributes standalone and is never folded; a child with no links is folded only. Fixes a double-charge where a self-linking child was both folded and self-attributed. - Recursion: a fold aggregates a child plus its non-self-linking descendants (depth-first, cycle-guarded), so grandchildren spawned by subagents reach the PR report. - Global linkage: the subagent index keys by parentSessionId alone (UUIDs are globally unique), so a child whose worktree resolves to a different project still links. - Date-range correctness: spawn-to-PR sets are built at assembly from the full turn list, so a spawn in a pre-range turn attributes to the right PR; a PR-linked parent whose own turns fall out of range is kept as a 0-cost fold anchor so its in-range child is not lost. - Timestamp fallback compares epoch ms (mixed UTC offsets order right) and is end-bounded: a child active after the parent's last turn is unlinked (contributes nothing), matching orphan semantics. - Cache adoption tries the newest prior versioned file (v6 then v5) so the preceding build's expired-PR history survives the v7 bump; an invariant note requires the list to cover every version that can exist on disk. - Spawn-result pairing matches the tool_result block that carries the agentId, not the first block, when a record batches several results. - resolveSubagentAttribution is computed once and shared by aggregateByPr and prLinkedTotals. subagentSessions now counts folded subtrees (children plus descendants). Verified on real data: attributed + unattributed reconciles to cost, and parent-only cost plus folded-children cost equals the folded total to the cent (no double-count).
Address a second adversarial review of the new machinery. - ID collision: parents and a child's parent reference are keyed by provider + sessionId, not bare sessionId. When two distinct parents still share a key (true duplicate/imported data), the child folds into NEITHER (deterministic skip, stays standalone): correctness over coverage. - Recursion dedup is global: one claimed-set spans all of a parent's direct children, so a descendant reachable through two paths (a diamond or duplicate id) folds exactly once and a parent-link cycle terminates. - Cache adoption migrates every prior version oldest-to-newest and MERGES per source path (newer wins per entry), so a sparse or partial newer file no longer masks older-only expired-PR orphans. - Fold anchors (0-cost PR-linked parents kept only for attribution) live in a new ProjectSummary.subagentAnchors, never in `sessions`, so they no longer contaminate session counts, averages, or any per-session report. Folded PR rows take their date span from the contributing child activity rather than the anchor's empty timestamps. - One-pass buildPrAttribution computes rows and totals together; the payload builder and CLI call it once. Drops the identity-keyed memoization, which could return stale folds if the array was mutated. - Ambiguous multi-block spawn-result pairing leaves the spawn link unset on purpose; the child then folds via the timestamp fallback rather than pairing with the wrong id or disappearing. Every fix is mutation-verified. A fresh real-data drive re-proves the no-double-count identity to the cent (parent-only cost plus folded cost equals the folded total) and that the PR rows sum to attributedCost.
Third adversarial review pass. - Ambiguity counts ALL candidate parents (and anchors) sharing a provider+sessionId key, not just PR-bearing ones, and uses a per-record fingerprint: a key carried by more than one DISTINCT record folds its child/subtree into NEITHER (identical duplicates still fold once). This unifies the parent-collision and duplicate-descendant rules and is deterministic across input order. - Project-rebuilding filters (by day, by date range, by config source) now carry subagentAnchors through, and a date filter CONVERTS a spawn parent whose in-range turns are all filtered out into an anchor so a surviving in-range child still folds. Rebuilt sessions also keep their PR + subagent-linkage metadata (prLinks, parentSessionId, spawnPrSets, ...), which buildSessionSummary otherwise drops, so by-PR and folding work on a filtered slice (menubar/dashboard flow). - Fold anchors leave ProjectSummary.sessions entirely and folded PR rows take their span from the child, so 0-cost anchors never touch session counts or averages. - Ambiguous spawn pairing (parent named the agent but its exact launching tool_use could not be paired) is recorded per parent; a late child of such a pairing folds to the parent's last turn within a 30 minute grace window, else stays unlinked. A truly-absent pairing gets no grace. - Row session key is NUL-delimited and provider-prefixed, so a project name or session id containing a space no longer collides and undercounts distinct sessions. Every fix mutation-verified. A fresh real-data drive re-proves the no-double-count identity to the cent, and a day-filtered drive proves the filter fix end to end (anchors created, subagents fold, identity holds).
Fourth adversarial review pass. - sessionFingerprint now covers the COMPLETE linkage-relevant payload, not just headline stats: a canonical (sorted-key) serialization of agentSpawnLinks, spawnPrSets, prRefsAtRangeStart, ambiguousSpawnAgentIds, parent/agent identity, and the per-turn prRefs timeline. Two records that share an id and headline stats but map the child to different spawns/PRs now fingerprint DISTINCT, so the ambiguity rule fires and they fold into neither, deterministically rather than order-dependent first-wins. - A date/day filter recomputes prRefsAtRangeStart at the new slice boundary by replaying the original full turn sequence, instead of copying the wide range's value. A PR switch between the wide start and the slice start (July 1 A, July 10 B, slice July 20) now carries B, not a stale A; a turn exactly on the boundary stays in-slice and applies its own refs. The recompute selects by timestamp, so it is order-independent. Non-contiguous day selections are documented as treated contiguous from the earliest selected day (a single session-level seed cannot represent multiple segments; the menubar selection is a single day or a run). - The anchor-carry path drops an anchor that duplicates a surviving session id, so malformed merged input cannot double-count. Also: rebuilt filtered sessions were losing their PR/subagent-linkage metadata (a child its parentSessionId, a parent its prLinks), which carryLinkageFields now restores, so by-PR and folding work on any filtered slice. Every fix mutation-verified. A fresh real-data drive re-proves the no-double-count identity and reconciliation to the cent for both a lifetime scan and a day-filtered slice (anchors created, subagents fold through the filter).
Fifth adversarial review pass; closed-form fixes. - sessionFingerprint serializes the COMPLETE fold-determining state via a real recursive canonical encoder: session-level linkage AND, per turn in sequence, timestamp, prRefs, cost, calls, savings, and per-model cost. Object keys are sorted recursively and set-semantic arrays (PR-ref lists, ambiguous ids, spawnPrSets values) are sorted, while the turn list keeps order; the structure is emitted through JSON.stringify (no delimiter concatenation). Two same-id parents that differ only in a turn timestamp now fingerprint DISTINCT (fold neither), and records differing only in set-array order fingerprint EQUAL (no false ambiguity). - recomputeRangeStartPrRefs breaks an exact-same-millisecond tie deterministically by the lexicographically-last sorted-ref key, so the recomputed seed is stable regardless of turn order. - A day filter seeds EACH selected day's first ref-less turn by replaying the original full turn sequence up to that day's start (per-day seeding), so a PR switch on an UNSELECTED day between two selected days carries to the later day. Contiguous and non-contiguous selections are both correct. - The anchor dedupe drops an anchor only when a surviving session shares the full provider-aware, fingerprint-qualified identity (a proven duplicate): a different-provider or different-record same-id session no longer wrongly drops the anchor. Also fixed a double-count the fingerprint test exposed: two duplicate parent sessions share a key and the SAME resolved children, so folding is now done once per parent key. Every fix mutation-verified. Fresh real-data drives (lifetime, single-day, and a NON-CONTIGUOUS day selection) re-prove no-double-count and reconciliation to the cent.
iamtoruk
force-pushed
the
feat/subagent-pr-attribution
branch
from
July 21, 2026 03:47
de00ecd to
2803040
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Subagent (sidechain) session cost never reaches the by-PR view: each sidechain transcript becomes a standalone session with no
prLinks, so it is dropped from the by-PR aggregation. A session orchestrated on one model with implementation lanes on another therefore shows only the parent model on every PR row, and the by-PR totals understate real spend.What
Fold each sidechain session's cost, calls, per-model cost, and per-category cost into the PR its launching parent turn was working on, so subagent spend inherits that PR under the existing turn-level attribution. Children remain standalone rows in the sessions list; each is counted exactly once.
Attribution model
resolveSubagentAttributionresolves every subagent, once, against its parent's UNFILTERED turn data.buildPrAttributionproduces PR rows and totals in a single pass (soaggregateByPrandprLinkedTotalsnever disagree; the payload builder and CLI call it once), folding each parent key exactly once so duplicate parents cannot double-count shared children.Resolution of a folded subtree to a PR:
toolUseResult.agentId) pairs the child's agent id with theAgent/Tasktool_useid that launched it; each spawn id maps to the PR active at its emitting turn over the FULL (pre-slice) turn list, so a pre-range spawn still yields the right PR.Correctness rules:
ProjectSummary.subagentAnchors, never insessions; a folded row takes its span from the child.Filters (menubar / dashboard)
The project-rebuilding filters carry
subagentAnchorsthrough; a date filter CONVERTS a spawn parent whose in-range turns are all filtered out into an anchor. Rebuilt sessions retain their PR + subagent-linkage metadata.prRefsAtRangeStartis RECOMPUTED at the new slice boundary by replaying the original full turn sequence (order-independent, with a deterministic same-millisecond tie-break). The day filter additionally seeds EACH selected day's first ref-less turn from the full sequence up to that day's start, so a PR switch on an UNSELECTED day between two selected days carries correctly -- contiguous and non-contiguous selections are both exact. An anchor is dropped only when a surviving session shares its full provider-aware, fingerprint-qualified identity (a proven duplicate), so a different-provider or different-record same-id session does not wrongly drop it.Cache
CACHE_VERSION6 -> 7 (v6 never shipped; combined v5 -> v7 bump). Adoption migrates EVERY prior versioned file oldest-to-newest and MERGES per source path (newer wins), so a sparse newer file cannot mask older-only expired-PR orphans; an invariant note requires the list to cover every prior version on disk. New cache fields (spawnToolUseIds,parentSessionId,agentSpawnLinks,ambiguousSpawnAgentIds) are threaded through the validator and the append/compact paths likeprRefs, with an append-vs-cold parity test.Footer / payload
Totals include folded subagent spend, so CLI and app footers read "across N PR-linked session(s) + M folded-in subagent runs"; the payload gains an additive
subagentSessionscount (folded subtrees).distinctCostcovers folded spend, documented in the payload type.attributedCost + unattributedCost === distinctCostholds and rows are summable.Verification
cli-durable-totalsprovider-parity) is pre-existing on main and unrelated; a couple of CLI integration tests that scan the live environment flake under parallel contention and pass 3/3 in isolation.subagentSessionsequals the folded subtree count -- for a lifetime scan, a single-day slice (anchors created, subagents fold through the filter), and a NON-CONTIGUOUS day selection.