werkregie/doc/prs/2026-10-14-PR#74-wake-session.md

PR#74: "Du bist dran" Shows the Author Session of a Stuck Agent Pull Request

WARNING: This document describes only the change applied in this PR. It may already be outdated once the next PR is merged. Historic PR documentation is not maintained along with new PRs; treat it as a snapshot.

Related Links

• werkregie PR#48, watch-pr: the watch of a coding agent ends with its session.
• werkcrappit PR#68 and werkator PR#161, stuck in a conflict with main on 2026-10-14.

Problem

On 2026-10-14, werkcrappit PR#68 had a conflict with main after werkcrappit PR#69 merged. werkrevisor wrote that the author rebases; its session had ended on 2026-10-12, so its watch had ended too. "Du bist dran" showed the human the stuck pull request, but not which session wrote it. The human had a coding agent search the transcripts for the branch and send the session the task; the session then rebased.

The data to find the session exists on the machine of the session: each line of a Claude Code transcript carries the gitBranch of the session, and the session file of Claude Desktop names the cliSessionId of its transcript. The relay reads both already, but reports only the sessions of the last 24 hours.

Non-Goals

• No automatic resume of the session; the human sends the message. A probe showed that Claude Desktop ignores a turn that claude --resume <id> -p adds from outside (see Solution).
• No new agent for the fix; the author session has the context of the pull request.
• No Codex threads; Claude Code only.

Specification

Feature: "Du bist dran" shows the author session of a stuck agent pull request

Scenario#74.01: a stuck pull request shows its author session

… so that the human finds the session that can fix the pull request.

• Given a pull request of an agent, stuck in a conflict with main, and a session that last worked on its branch in a checkout of its repository, 3 days ago
• When the human opens "Du bist dran"
• Then the pull request shows the title of the session with a link that opens it in Claude Desktop

Verified by

• AuthorsTest "a stuck pull request shows the session that worked on its branch, days ago"
• ClaudeCodeSourceTest "a session reports the branch it last worked on, older than the window too, unless it is archived"
• RelayClientTest "reports the sessions by their branch along"
• AuthorsTest "the snapshot attaches the authors of all sources"

Scenario#74.02: the message for a conflict asks for a rebase

… so that the human sends the task with one copy and one paste.

• Given a pull request of an agent, stuck in a conflict with main, with its author session
• When the human copies the message of the pull request
• Then the message names the repository, the pull request and its address, and asks to rebase onto main, take the next free version and push

Verified by

• AuthorsTest "the message for a conflict asks for a rebase onto main, the next free version and a push"

Scenario#74.03: the message for RED asks for the review round

… so that the agent answers werkrevisor's findings.

• Given a pull request of an agent, stuck with RED, with its author session
• When the human copies the message of the pull request
• Then the message asks to fix the findings or answer them with evidence, and push

Verified by

• AuthorsTest "the message for RED asks to fix the findings or answer them with evidence"

Scenario#74.04: the last session on the branch is the author

… so that a session that left the branch long ago is not shown.

• Given two sessions that worked on the branch of a stuck pull request
• When the human opens "Du bist dran"
• Then the pull request shows the session with the later line on the branch

Verified by

• AuthorsTest "the session with the later line on the branch is the author"
• TranscriptBranchTest "the branch of a session is the last one other than a detached HEAD, with the time of its last line"

Scenario#74.05: the same branch in another repository is no match

… so that a branch name that many repositories share, e.g. claude-code/sync-agentic-workflow, finds the right session.

• Given a stuck pull request of werkator and a session on a branch of the same name in werkcrappit
• When the human opens "Du bist dran"
• Then the pull request shows no session

Verified by

• AuthorsTest "the same branch in another repository is no match"

Scenario#74.06: an archived session is no author

… so that a closed piece of work is not woken.

• Given a stuck pull request whose only session on its branch is archived
• When the human opens "Du bist dran"
• Then the pull request shows no session

Verified by

• ClaudeCodeSourceTest "a session reports the branch it last worked on, older than the window too, unless it is archived"

Solution

• The relay reports each Claude Code session with the branch it last worked on, as authors of its report: each non-archived session of Claude Desktop and each running one in a terminal, older than the window too. The branch is the gitBranch of the last line of the transcript that names one other than HEAD, read from the end in parts of 64 KiB, at most 8 MiB, and cached until the transcript changes.
• The server attaches to each pull request with a stuck flag the author with the latest such line on its head branch, in a checkout of a repository of the same name; the flags of a pull request stay as they were.
• The message names the repository, the pull request, the reasons and the address, and asks for the work: after a conflict a rebase onto main with the next free version, after RED the findings, else to check where it hangs and report. It is German, as the human writes to the sessions.
• "Du bist dran" shows the author below the reason: a button that opens the session, on a phone its Remote Control, and a copy symbol for the message.
• A server older than 0.47.0 ignores authors in a report, and a relay older than 0.47.0 sends none; both stay compatible.
• Version 0.47.0.

The Probe of an Automatic Resume

The first plan was that the relay resumes the author session itself with claude --resume <id> -p "…". A probe on a throwaway session of Claude Desktop on 2026-10-14 showed:

• The resumed run worked in permission mode auto and wrote its turn into the same transcript.
• Claude Desktop did not show the turn, and its next turn did not know it: it attached to the line before the resumed run, so the transcript forked. A rebase by the resumed run would be real in Git and Gitea, but unknown to the session.
• The CLI has a login of its own; it had expired, while Claude Desktop worked.
• The auto mode of a coding agent refuses to start another agent in auto mode.

The human chose to send the message by hand; #orch.resume is dropped in the plan.

Open Questions

None.

Prerequisite PRs

• werkregie PR#75 brought Dashboard, RelayClient, ClaudeCodeSource and GiteaSource to a Mutation-CRAP of at most 15. Since werkregie PR#64, a branch fails on each method above 15 in a class it changes.

Follow-up PRs

• The demo data get an author session for their pull request in conflict, so that the public demo shows the button. This PR leaves Demo.kt unchanged: its method snapshot is above a Mutation-CRAP of 15 and would first need tests or a split, see the prerequisite PR above.
• A Codex thread counts as author of a branch, as a Claude Code session does.