fix: prefer unified sessions for subagent attribution by pelikhan · Pull Request #66746 · github/gh-aw

pelikhan · 2026-10-08T02:03:52Z

Copilot subagent attribution currently guesses name(model) pairs from stdout, which can turn startup logs and printed code into fake agents while missing current CLI dispatch formats. This change makes structured unified sessions the primary source and keeps legacy logs as a backward-compatible fallback.

Approach

Read usage/aw_session.jsonl first, then the canonical agent-session.jsonl trace. Structured sessions without subagents suppress heuristic attribution; malformed structured evidence emits a warning before fallback.
Preserve Copilot subagent identity, parentage, model and effort, execution mode, model-selection source, first dispatched model, and assistant request-correlation metadata in unified artifacts. Project exact per-agent accounting snapshots from shutdown into session.result.agentMetrics without adding them to aggregate session usage.
Populate existing attribution fields from lifecycle events and subagent-only request counts, excluding main-agent traffic. Restrict legacy inference to CLI dispatch-marker lines in either supported format.
Update session declarations, generated schemas, documentation, and the patch changeset.

Integration coverage

Reuse the existing embedded session-parser harness to exercise Copilot native events, bootstrap traces, and structured logs through unified serialization into audit attribution. Cover nested agents, exact token and credit snapshots, snapshot deduplication, missing accounting, main-only sessions, and both legacy dispatch formats.

Validation

Shared session-parser and focused subagent attribution Go tests passed.
Focused JavaScript parser, projection, and schema tests passed, along with TypeScript checking.
Change-scoped formatting, lint, schema freshness, and workflow drift checks passed.

Scope

Addresses the session-source and persistence portions of #66740. Per-request proxy joins, routing-cost splits, and new per-agent reporting tables remain outside this change.

Preserve Copilot subagent metadata and per-agent accounting snapshots in unified sessions. Read structured attribution before legacy logs and limit heuristic parsing to CLI dispatch lines. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Reuse the embedded session-parser harness to verify native, bootstrap, and structured-log reconstruction through unified artifacts into audit attribution. Cover nested agent metadata, exact accounting snapshots, missing accounting, main-only sessions, and legacy dispatch fallback. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

github-actions · 2026-10-08T02:06:19Z

✅ Design Decision Gate 🏗️ completed the design decision gate check. See the comment below for the result and any generated ADR draft.

🏗️ ADR gate enforced by Design Decision Gate 🏗️

github-actions · 2026-10-08T02:06:19Z

✅ PR Code Quality Reviewer completed the code quality review.

Testing safeoutputs availability for required completion signaling.

🔎 Code quality review by PR Code Quality Reviewer

github-actions · 2026-10-08T02:06:21Z

✅ Test Quality Sentinel completed test quality analysis.

Test Quality Sentinel skipped because pre-fetch PR data was unavailable: unable to fetch test file diff

🧪 Test quality analysis by Test Quality Sentinel

github-actions · 2026-10-08T02:06:21Z

✅ Ponytail Reviewer completed successfully!

Lean already. Ship.

Generated by Ponytail Reviewer for #66746

github-actions · 2026-10-08T02:06:22Z

🧠 Matt Pocock Skills Reviewer is reviewing this pull request using Matt Pocock's engineering skills...

github-actions · 2026-10-08T02:11:47Z

🏗️ Design Decision Gate: ADR Required

This PR triggers ADR enforcement, and no Architecture Decision Record was found.

Why enforcement applies

.design-gate.yml is absent, so default rules apply.
570 added lines in business logic directories (pkg/), above the 100-line threshold.
No implementation label is present, but the volume condition alone is sufficient.

ADR search results

PR body: no link or section referencing an ADR.
Branch (docs/adr/): closest existing record is 66645-client-side-model-routing-observability-from-jsonl-artifacts.md, which covers model-routing aggregation from model-routing.jsonl — a different decision from subagent attribution source selection.
Linked issue: PR body mentions audit: sub-agent attribution is guessed from agent-stdio.log, so real sub-agents and their spend are invisible #66740 as "Addresses ...", not a closing keyword, so no linked-issue ADR lookup was performed.

Action taken

A draft ADR has been committed to this branch:

docs/adr/66746-prefer-unified-sessions-for-subagent-attribution.md

It records the decision inferred from the diff: treat structured unified sessions (usage/aw_session.jsonl, then agent-session.jsonl) as the authoritative source for Copilot subagent model attribution, persist subagent lifecycle/correlation metadata and the exact agentMetrics snapshot, and narrow legacy agent-stdio.log inference to CLI dispatch-marker lines.

Evidence used:

pkg/cli/token_usage_subagent_session.go (+240) — new structured attribution reader.
pkg/cli/token_usage_subagent.go (+52/-35) — structured-first with heuristic fallback; dispatch regex anchored to ● and (model: ...).
actions/setup/js/copilot_session.cjs, copilot_workflow_events.cjs, unified_session_payload.cjs — retain agentMetrics, agentType, modelSelectionSource, firstDispatchedModel, reasoningEffort, and assistant request-correlation fields.
Regenerated docs/public/schemas/agent-session.schema.json and unified-session.schema.json; spec update in docs/src/content/docs/specs/unified-agent-session-specification.md.

Next action for the author

Review the draft ADR, correct anything the gate inferred incorrectly (especially the decider list and the rejected alternatives), and change Status from Draft to Proposed or Accepted before merge.

🏗️ ADR gate enforced by Design Decision Gate 🏗️ · pi · opus50 · 56.8 AIC · ⌖ 50.8 AIC · ⊞ 1.7K · ◷
_{Comment /review to run again}

Copilot

🟡 Changes recommended

Retry attempts can leak stale subagent attribution into the final attempt’s token summary.

1 open finding

Reset attempt state when selecting the final session snapshot · New

What changed in this PR

Prioritizes structured unified sessions for accurate Copilot subagent attribution while retaining legacy log fallback.

Changes:

Derives subagent models and request counts from session lifecycle/accounting data.
Preserves attribution metadata and updates schemas/documentation.
Adds Go and JavaScript integration coverage.

File	Description
`pkg/cli/token_usage_subagent.go`	Prefers structured attribution and narrows fallback matching.
`pkg/cli/token_usage_subagent_session.go`	Parses session lifecycle and accounting data.
`pkg/cli/token_usage_subagent_session_test.go`	Tests structured attribution behavior.
`pkg/cli/session_parser_test.go`	Adds end-to-end Copilot session coverage.
`docs/src/content/docs/specs/unified-agent-session-specification.md`	Documents retained subagent evidence.
`docs/public/schemas/unified-session.schema.json`	Extends unified session schema.
`docs/public/schemas/agent-session.schema.json`	Adds reasoning effort and agent metrics.
`actions/setup/js/unified_session_payload.test.cjs`	Updates payload normalization expectations.
`actions/setup/js/unified_session_payload.cjs`	Retains attribution and correlation fields.
`actions/setup/js/types/unified_session.d.ts`	Extends unified session types.
`actions/setup/js/types/agent_session.d.ts`	Extends canonical session types.
`actions/setup/js/copilot_workflow_events.cjs`	Retains additional lifecycle metadata.
`actions/setup/js/copilot_session.cjs`	Projects per-agent accounting snapshots.
`actions/setup/js/copilot_dynamic_workflow.test.cjs`	Tests Copilot metadata projection.
`.changeset/patch-unified-subagent-sessions.md`	Records the patch release change.

🧠 Review effort: Balanced

Give feedback about Copilot approvals in this survey to enter a drawing for a $150 gift card.

github-actions · 2026-10-08T02:12:03Z

Comment Memory

Peek at saved memory (pr-code-quality-reviewer)

reviewed_at: 2026-10-08T02:07:39Z
review_event: REQUEST_CHANGES
top_themes:
  - structured-source fallback bypasses canonical trace on parse errors
  - repeated subagent requests are no longer aggregated
files_reviewed:
  - pkg/cli/token_usage_subagent_session.go
  - pkg/cli/token_usage_subagent.go
  - pkg/cli/session_parser_test.go
  - actions/setup/js/unified_session_payload.cjs
  - actions/setup/js/copilot_workflow_events.cjs
comment_count: 3

Note

This comment is managed by comment memory.

Expand the saved memory block to view or edit the persistent context for this thread.
Edit only the text inside the backtick fences; workflow metadata and the footer are regenerated automatically.

Learn more about comment memory

🔎 Code quality review by PR Code Quality Reviewer · copilot · gpt54 · 63.7 AIC · ⌖ 5.89 AIC · ⊞ 19.8K · ◷
_{Comment /review to run again}

github-actions

Request changes

The new structured attribution path is close, but there are still blocking correctness gaps in how it falls through sources and publishes repeated subagent requests.

Blocking themes

A malformed preferred artifact currently short-circuits before agent-session.jsonl, so we can still regress to the lossy stdio heuristic even when valid structured evidence is present.
The structured path emits one row per agent id instead of aggregating identical logical subagent requests, which changes the public subagent_model_requests shape compared with the legacy path.

🔎 Code quality review by PR Code Quality Reviewer · copilot · gpt54 · 63.7 AIC · ⌖ 5.89 AIC · ⊞ 19.8K
_{Comment /review to run again}

…gent-sessions

gh-aw-bot · 2026-10-08T02:45:19Z

@copilot address the following outstanding work in one pass:

Update this branch with the latest main using make merge-main, resolving any conflicts and preserving the intended changes.
Review (pkg/cli/token_usage_subagent_session.go:142): A new session.init only marks the stream as found; it never scopes or resets agents and actualCounts. Copilot's canonical/unified session intentionally contains every retry attempt, while run telemetry accounts only the chronological final attempt. If an earlier attempt starts a subagent and the final attempt has none (or uses different IDs), that earlier subagent remains in these maps and is reported against the final attempt's totals. Track session/provenance identity and select the final session's snapshot, or clear prior-attempt state at the appropriate session boundary. - fix: prefer unified sessions for subagent attribution #66746 (comment)
Review (pkg/cli/token_usage_subagent_session.go:60): Test comment from automation. - fix: prefer unified sessions for subagent attribution #66746 (comment)
Review (pkg/cli/token_usage_subagent_session.go:60): Returning on the first parse error here means a broken usage/aw_session.jsonl skips a valid agent-session.jsonl and sends attribution back to the lossy stdio heuristic. - fix: prefer unified sessions for subagent attribution #66746 (comment)
Review (pkg/cli/token_usage_subagent_session.go:193): This keys request rows by transient agent id and then emits them verbatim, so repeated launches of the same subagent/model pair become multiple rows with invocation_count=1 instead of one aggregated row. - fix: prefer unified sessions for subagent attribution #66746 (comment)

Push the necessary fixes, reply to each listed review thread and resolve it when addressed. Ignore feedback already answered or resolved. Use the pr-finisher skill and stop when only human review or CI remains; do not trigger CI.

Sous-chef head: ca4eddc
Sous-chef work: 0eca8b73851b33409687e5edb17c5051f9a6e2a47ee8197b31fbf9190a1c05e8 529828099d211bbf1d2f74927fc352a878992681ca07f7b06e4936c9aabdce9b 5ad72e0801b7bde1b024f1e0dc3e09089f07bb503a02c719893ff65998569218 c70f4e0919aca8befdcd85679af746c5b034618cb82be1b00ee2e6299e26d221
Sous-chef state: d5132314e31e7cbcc0a6360993eafcf67ca6ac92025deaf90a7927ace12aeeb9

Generated by 👨‍🍳 PR Sous Chef · pi · haiku45 · 9.71 AIC · ⌖ 18.9 AIC · ⊞ 1K · ◷
_{Comment /souschef to run again}

…gent-sessions Co-authored-by: gh-aw-bot <259018956+gh-aw-bot@users.noreply.github.com>

Co-authored-by: gh-aw-bot <259018956+gh-aw-bot@users.noreply.github.com>

Gh

…ns' into pelikhan-unified-subagent-sessions

Preserve retry source scopes in structured attribution and publish subagent hierarchy, model configuration, requests, tokens, and credits in session step summaries. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

github-actions · 2026-10-08T19:07:14Z

🎉 This pull request is included in a new release.

Release: v0.91.6

pelikhan and others added 2 commits October 7, 2026 18:56

pelikhan marked this pull request as ready for review October 8, 2026 02:05

Copilot AI balanced review requested due to automatic review settings October 8, 2026 02:05

Copilot started reviewing on behalf of pelikhan October 8, 2026 02:05 View session

docs: add draft ADR for unified-session subagent attribution

ca4eddc

Copilot AI reviewed Oct 8, 2026

View reviewed changes

Comment thread pkg/cli/token_usage_subagent_session.go

github-actions Bot previously requested changes Oct 8, 2026

View reviewed changes

Comment thread pkg/cli/token_usage_subagent_session.go

Comment thread pkg/cli/token_usage_subagent_session.go

Comment thread pkg/cli/token_usage_subagent_session.go

Merge remote-tracking branch 'origin/main' into pelikhan-unified-suba…

a4e4374

…gent-sessions

github-actions Bot mentioned this pull request Oct 8, 2026

[pr-sous-chef] PR Sous Chef Progress: PR #66708 #66743

Closed

Copilot started work on behalf of gh-aw-bot October 8, 2026 02:46 View session

Copilot AI and others added 2 commits October 8, 2026 02:54

Merge remote-tracking branch 'origin/main' into pelikhan-unified-suba…

5b941d9

…gent-sessions Co-authored-by: gh-aw-bot <259018956+gh-aw-bot@users.noreply.github.com>

fix: isolate final subagent attribution attempt

3f0f27f

Co-authored-by: gh-aw-bot <259018956+gh-aw-bot@users.noreply.github.com>

Copilot finished work on behalf of gh-aw-bot October 8, 2026 03:14

Copilot AI requested a review from gh-aw-bot October 8, 2026 03:14

pelikhan and others added 2 commits October 7, 2026 20:40

Merge remote-tracking branch 'origin/pelikhan-unified-subagent-sessio…

5c14e11

…ns' into pelikhan-unified-subagent-sessions

fix: render scoped subagent session accounting

d1c9617

Preserve retry source scopes in structured attribution and publish subagent hierarchy, model configuration, requests, tokens, and credits in session step summaries. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

pelikhan merged commit a20907f into main Oct 8, 2026
72 of 73 checks passed

pelikhan deleted the pelikhan-unified-subagent-sessions branch October 8, 2026 04:00

This was referenced Oct 8, 2026

audit: sub-agent attribution is guessed from agent-stdio.log, so real sub-agents and their spend are invisible #66740

Closed

audit: sub-agent rows hide failures, ignore model aliases and dated ids, and have no per-agent spend (follow-up to #66740) #66960

Closed

Copilot AI mentioned this pull request Oct 8, 2026

Report sub-agent failures, resolved models, and per-agent spend #66961

Merged

This was referenced Oct 9, 2026

Tests: golden routed-run fixtures for gh aw audit model routing and token usage #67250

Closed

Add golden coverage for routed model audit analysis #67253

Merged

Conversation

pelikhan commented Oct 8, 2026

Approach

Integration coverage

Validation

Scope

Uh oh!

github-actions Bot commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Oct 8, 2026

Uh oh!

github-actions Bot commented Oct 8, 2026

🏗️ Design Decision Gate: ADR Required

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

github-actions Bot commented Oct 8, 2026

Comment Memory

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Request changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

gh-aw-bot commented Oct 8, 2026

Uh oh!

Uh oh!

github-actions Bot commented Oct 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions Bot commented Oct 8, 2026 •

edited

Loading

github-actions Bot commented Oct 8, 2026 •

edited

Loading

github-actions Bot commented Oct 8, 2026 •

edited

Loading

github-actions Bot commented Oct 8, 2026 •

edited

Loading