[jiffy-2389]

Issue #717 — run.json counts subgraph agents' tokens twice

Root cause

Two sibling handlers stamp a node id onto events, and they disagreed:

• pipeline.NodeScopedPipelineHandler (pipeline events) prefixes:
  evt.NodeID = parentNodeID + "/" + evt.NodeID.
• agent.NodeScopedHandler (agent events) overwrote: evt.NodeID = h.nodeID.

Scopes nest by wrapping. A subgraph's registry factory wraps the parent's agent
handler with a handler scoped to the subgraph node ("Task02"); the child's
codergen handler wraps that again with its own node id ("Implement"). Because
the agent stamp overwrote, the agent's call-level lines (llm_finish,
turn_metrics) landed under "Task02" (or "Implement"), while every pipeline
event for the same node — including the decision_outcome rollup that carries
token_input/token_output — landed under "Task02/Implement".

pipeline/runusage.go keys per-node usage by node id and takes each node's most
direct tier, so it counted the call tier under one id and the rollup tier under
the other. .totals came out at exactly twice combinedUsageForBudget's
snapshot. Reproduced before the fix on a two-agent child subgraph: engine and
the last cost_updated both said 240 tokens; .totals said 480 (ratio exactly
2.00), with the container node Task02 also carrying its children's 200 input
tokens. The same defect showed up for manager_loop children.

Fix

agent/events.go — the agent stamp now composes exactly as the pipeline one
does (new scopeNodeID helper): adopt the scope when the event names no node,
otherwise prefix "<scope>/<existing>". One node now has one id across the agent
and pipeline streams, so runusage.go's per-node tier selection sees a single
node and counts its work once. A top-level agent node is unchanged ("Gen").

Supporting changes: docs/architecture/agent.md (the NodeID invariant now
states the composition rule) and a CHANGELOG ### Fixed entry.

Tests (all written first, all fail before the fix and pass after)

• agent/events_test.go — TestNodeScopedHandlerComposesWithAnEnclosingScope
  (the registry-factory + codergen nesting), …StampsABareNodeID,
  …MatchesPipelineScoping (table pinning the rule shared with
  NodeScopedPipelineHandler, including the degenerate "Review/Review" case),
  …ComposesANestedSubgraphPath ("L1/L2/Leaf").
• pipeline/handlers/expand_integration_test.go —
  TestSubgraphAgentEventsShareTheirNodesScopedID: runs a real registry/engine
  with a subgraph and asserts the agent's usage-bearing events and the matching
  decision_outcome carry the same "Sg/Child" id.
• tracker_capture_totals_test.go —
  TestRunManifestSubgraphTotalsCountEachAgentOnce: end-to-end run with capture,
  comparing run.json against three independent oracles — the engine's own
  accumulated usage, the run's last cost_updated event (the exact discrepancy
  the issue reports), and per-node attribution (agent spend must land on
  Task02/Implement / Task02/Review, not on the container Task02).

Verified by reverting only agent/events.go and re-running: the end-to-end test
then reports run.json totals = 480 tokens, but the last cost_updated says 240.

Verification actually run

• go build -o /dev/null ./... — pass.
• go test -vet=off -count=1 ./... — pass, no failures.
• go test -count=1 ./cmd/tracker-conformance -run TestGoldenTraces — pass (the
  golden traces include node_id and usage; no fixture needed regeneration).
• go vet ./agent/ ./pipeline/handlers/ . — clean.
• make complexity — complexity gate OK (135 grandfathered, 0 new).
• Repo hooks on exactly the changed files (per instruction, not the frozen
  prek run --from-ref FETCH_HEAD --to-ref HEAD command):
  prek run --config .pre-commit-config.yaml --files CHANGELOG.md agent/events.go agent/events_test.go docs/architecture/agent.md pipeline/handlers/expand_integration_test.go tracker_capture_totals_test.go
  — all hooks pass (incl. gofmt, goimports, make test); no files were modified
  by the hooks.

Limitations / notes

• Archived runs keep their inflated totals. The double count is recorded at
  emission time; a pre-fix log still contains both node ids, so re-assembling it
  with tracker run-json does not recover the correct figure. Only new runs are
  fixed.
• Out of scope: parallel fan-out shows the same aggregate-vs-children
  shape (the component node's decision_outcome rollup restates its branch
  targets' spend, which is booked under the branches as well). That is a
  separate construct from this issue's subgraph nesting and was left untouched.
• A bare agent.Completer emits no trace events, so llm_calls is unavailable
  in the end-to-end test (a limit already documented by
  TestRunManifestTotalsMatchEngineUsage); the test therefore asserts token and
  cost totals rather than a call count.

Starting base: c697043bf705b9cfec28326d7b6183cfee2a3832
Verified base: c697043bf705b9cfec28326d7b6183cfee2a3832
Verified commit: 4815e57b29b722f6237e70f773c7d3cf7e5631e8

Verification commands and exit codes:

[{"Argv":["prek","run","--from-ref","FETCH_HEAD","--to-ref","HEAD","--config",".pre-commit-config.yaml"],"ExitCode":0}]

Closes #717

Final review: pass (pi/lunaroute/deepseek-4.1-flash), job 1, round 1.

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

[coderabbitai]

Important

Review skipped

This PR was authored by a bot without an assigned CodeRabbit review seat. Ask an organization administrator to assign a seat to the bot. See the seat-assignment guide. After the seat is assigned, request a review again.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: d2fc3616-33bc-4198-bddf-f4157f04402c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn