[peyton-alt]

https://entire.io/gh/entireio/cli/trails/1602

An OpenCode session whose background subagent is still running can now be credited for a commit that another session also claims, whether the parent is IDLE or still in its turn (ACTIVE).

A running subagent's edits reach the parent's tracked files only when it stops. Until then, post-commit's read-only gate asks the live task records whether they touched a committed file (liveTaskFilesInCommit). It read only transcripts on disk, and a running OpenCode child has none (its transcript exists only through opencode export), so the parent was skipped as read-only and the commit lost its link to that session.

Changes

• New agent.LiveSubagentFileReporter: reports the files a running subagent call has modified so far. OpenCode implements it by exporting the child to a private staging file, cutting it to the running call, and removing it. Only completed tool parts count: a pending, running or failed edit is no evidence. It writes nothing a capture reads, so a report that finishes after the call's stop cannot replace the stop's complete transcript.
• liveTaskFilesInCommit uses the reporter for a live record with no transcript on disk. A report that fails, or runs out of budget, is no evidence.
• Fetch budgets: the co-authorship checks of all a commit's sessions share 10s, counted from the start of post-commit's session loop, so a hung export there cannot hold the commit. Condensation's re-exports of completed tasks keep their own 30s per session (agent.SubagentFetchBudget), as on main.
• The gate is memoized per session (sync.OnceValue), since each consultation may export.
• prepare-commit-msg's check of a foreign session's running subagents shares the same 10s deadline, so a slow or hung export cannot hold up git commit either.
• The gate's evidence applies to an ACTIVE parent too (hasFreshLiveTaskRecord: any phase but ENDED), for every agent. A parent keeps working after launching a background subagent. For OpenCode this is what credits it at all, since its live transcript does not include its children; for Claude and others it applies when the parent's own transcript does not yet show the subagent's files and its on-disk transcript names a committed file. When to skip the overlap check is unchanged (idleWithLiveTaskRecord, IDLE only).
• Docs: OpenCode AGENT.md and sessions-and-checkpoints.md, including that a still-running subagent credits its session again on a later commit of the same file (as Claude's on-disk transcripts do on main).

A commit made outside the parent's agent process (a terminal, another window) goes through #2668's foreign-session checks. #2668 has merged, and this PR switches both of them (the prepare-commit-msg staged-file overlap and post-commit's liveTaskCoauthored) to hasFreshLiveTaskRecord, so an ACTIVE parent's running subagent counts there too, as #2668 already allowed for an IDLE one.

This redoes the co-authorship part cut from #2390, rebuilt on what landed there: readTaskTranscript no longer re-exports live records, so the check uses its own non-writing probe instead.

Tests

• Live-tested with OpenCode 1.18.33 (opencode serve, gpt-5.6-luna): an IDLE parent whose background child rewrote a file the coder session also claimed, committed while the child was still running: main 1 session, this PR 2; with Keep a running agent's session off commits it did not make #2668 merged in, also 2 (both sessions foreign). With Keep a running agent's session off commits it did not make #2668 merged in, also: an ACTIVE parent (still in its own turn) is credited (2 sessions), and a parent alone gets the trailer when its running child's file is committed.

• TestOpenCodeSubagentRunningChildCoauthorsCommit (integration): an IDLE parent with a running child that wrote a file another session claims is condensed into the commit. Fails without the change.

• TestLiveTaskFilesInCommit_AsksTheAgentForARunningSubagent: committed file, other files, failed report, spent budget.

• TestCommitFetchBudgets_CoauthorshipGetsAShare.

• TestLiveSubagentModifiedFiles_WritesNothingACaptureReads: reports only the running call's files and leaves nothing under .entire/tmp.

• TestOpenCodeSubagentRunningChildCoauthorsCommit_ActiveParent (integration): the same with the parent mid-turn. Fails without the change.

• TestHasFreshLiveTaskRecord_AnyPhaseButEnded.

• TestShouldCondense_RunningSubagentCreditsAnActiveParent: the read-only gate for an ACTIVE parent of any agent.

• TestCommitFetchBudgets_CondensationKeepsItsOwnBudget: condensation still re-exports after the co-authorship deadline has passed.

• TestCompletedToolParts_DropsUnfinishedEdits.

• TestRunOpenCodeExportToFile_CancelKillsTheLaunchersChild: a launcher-style fake returns at the deadline (30s without the fix).

• TestForeignCommit_RunningOpenCodeSubagentLinksActiveForeignParent (integration): an ACTIVE OpenCode parent, foreign to the commit, is credited through its running child. Fails without the switch to hasFreshLiveTaskRecord.

• TestCommitFetchBudgets_PrepareCommitMsgIsBounded.

Diff: +675/−45 across 16 files.

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches post-commit session linking and runs live opencode export during commits with time budgets; incorrect attribution or budget starvation could mis-link checkpoints or skip co-authorship, but failures degrade to “no evidence” rather than hard errors.

Overview
Post-commit can now treat an IDLE OpenCode parent as a co-author when a still-running background subagent has already edited files in the commit, even though those edits are not in FilesTouched until the child stops.

Adds LiveSubagentFileReporter and OpenCode’s LiveSubagentModifiedFiles, which runs a scoped opencode export into a private staging file (no writes to .entire/tmp cache or the call’s declared transcript). liveTaskFilesInCommit falls back to that probe when a live task record has no on-disk transcript; failed or budget-exhausted reports count as no evidence.

Subagent fetch budgeting is refactored (WithSubagentFetchDeadline, shared subagentFetchContext); each commit gets a 30s export budget with 10s reserved for co-authorship checks. The live-task gate is memoized per session via sync.OnceValue.

Docs (AGENT.md, sessions-and-checkpoints.md) and tests cover the running-child commit case, budget splitting, and the no-persistent-writes guarantee.

^{Reviewed by Cursor Bugbot for commit 6c5727f. Configure here.}

[Copilot]

🟡 Changes recommended

Live probes can incorrectly credit unsuccessful edits and bypass repository-resolution safety checks.

2 open issues

• Avoid current-directory fallback on repository resolution errors · New
• Ignore pending or failed tool calls when detecting modified files · New

What changed in this PR

Enables Entire CLI to credit an idle OpenCode parent when its running subagent modified files committed alongside another session.

Changes:

• Adds a private, call-scoped live subagent probe.
• Shares export budgets across post-commit processing and memoizes co-authorship checks.
• Adds regression tests and updates documentation.

File	Description
docs/​architecture/​sessions-and-checkpoints.md	Documents live-task co-authorship.
cmd/​entire/​cli/​strategy/​manual_commit_task_fetch_test.go	Tests reporting and budget limits.
cmd/​entire/​cli/​strategy/​manual_commit_hooks.go	Integrates budgeted, memoized probes.
cmd/​entire/​cli/​integration_test/​opencode_subagent_test.go	Tests running-child co-authorship.
cmd/​entire/​cli/​integration_test/​hooks.go	Adds configurable transcript timestamps.
cmd/​entire/​cli/​agent/​subagent_fetch.go	Shares deadline handling between fetch operations.
cmd/​entire/​cli/​agent/​opencode/​subagent_export.go	Implements private live exports.
cmd/​entire/​cli/​agent/​opencode/​subagent_export_test.go	Tests call scoping and staging cleanup.
cmd/​entire/​cli/​agent/​opencode/​lifecycle.go	Asserts reporter interface implementation.
cmd/​entire/​cli/​agent/​opencode/​AGENT.md	Documents probe behavior and budgets.
cmd/​entire/​cli/​agent/​capabilities.go	Adds built-in reporter capability lookup.
cmd/​entire/​cli/​agent/​agent.go	Defines the live file-reporting interface.

_{Review effort: Balanced}