/job

name: job
description: >
  Daily routines for a corporate software engineering job. Start mode
  triages the inbound review queue, your own open PRs, and the issue
  tracker's plan for today. End mode clears the outbox, surfaces review
  debt, sweeps worktrees for unpushed work, and tidies tracker state for
  tomorrow. Use via /job, /job start, /job end, or /job setup.
argument-hint: "[start | end | setup]"
disable-model-invocation: true
allowed-tools:
  - Read(${CLAUDE_SKILL_DIR}/references/*)
  - Bash(claude agents:*)
  - Bash(claude --bg:*)
  - Bash(herdr agent list:*)
  - Bash(herdr agent focus:*)

Job

Run a daily work routine: gather state read-only, present one prioritized brief, then execute confirmed bulk actions through installed skills.

Config

!`cat ~/.config/claude-job-skill/config.json 2>/dev/null || echo CONFIG_MISSING`

If the output above is `CONFIG_MISSING` and the requested mode is not `setup`, run the interview in [references/setup.md](references/setup.md) first, then continue into the requested mode.

Mode

Requested mode: `$0`. Local time: !`date "+%A %H:%M"`.

• `start`: read [references/start.md](references/start.md)
• `end`: read [references/end.md](references/end.md)
• `setup`: read [references/setup.md](references/setup.md)

With no argument, suggest a mode from the local time above: before the midpoint of the configured working hours (about 13:00 by default) suggests start, after suggests end. On a weekend, suggest nothing and ask. Confirm the suggestion with AskUserQuestion before proceeding.

Arguments beyond the mode are a focus hint ($ARGUMENTS). Weight the brief toward what they name.

Delegation

This skill never names platforms. The config declares which version-control platform, issue tracker, worktree tool, messaging platform, email account, and optional personal capture inbox the user works with. For each task, find the installed skill covering the configured tool's task and load it for mechanics, falling back to the configured CLI or MCP directly. Stay read-only until the user confirms actions.

Platforms, hostnames, and usernames come only from the config or the user, never from this skill.

Brief, then act

Both modes follow this contract.

Gather

Read-only first. The sources are independent (review queue, own PRs, tracker, messaging inbox, email inbox), so dispatch parallel read-only sub-agents and merge their results. Merge on shared identifiers: when items from different sources name the same issue or MR, they are one piece of work and become one brief entry carrying every source's state. The join is the orchestrator's job, keyed on the cross-references each sub-agent returns.

Reporting

End every gather prompt by naming the report as the agent's final action: send it to the orchestrator before going idle. An agent that goes idle carrying no report gets one nudge, never a second. When the nudge produces nothing, treat that source as timed out.

Gather runs against a five-minute deadline from dispatch. Past it, stop the outstanding agent (`TaskStop` for an `Agent` dispatch) and run its source inline.

Surface each source's findings as it lands, so a stalled source delays one entry instead of the brief. Hold the merge and the brief entries themselves until every source has reported or been run inline.

Agents

Background Claude sessions may already be working items in the brief. Run `claude agents --json --all` inline in the orchestrator rather than in a sub-agent: it is one command, and the join needs the raw records that a sub-agent summary would flatten. Every record carries `sessionId`, `name`, `cwd`, `kind`, and `startedAt`. The rest is keyed to `kind`: a `background` record adds `id` and `state` (`working`, `blocked`, `done`, or `failed`, plus `waitingFor` when blocked), and an `interactive` record adds `pid` and `status` instead. Join on `sessionId`, the one field both kinds share.

When `HERDR_PANE_ID` is set, a second inline call, `herdr agent list`, maps live sessions to the terminal panes running them. Join its agents to the `claude agents` records on `sessionId == agent_session.value` and record the matched `pane_id` on the brief item. That join is exact, so never fall back to a title or `cwd` comparison. When `HERDR_PANE_ID` is unset there is no herdr server to ask, so skip the call and the resume command below carries the handoff. Load the `herdr:herdr` skill for the mechanics and for which commands answer with JSON.

Join each record to a brief item in this order:

1. A `name` matching the `job-<identifier>` convention below, with the identifier in its sanitized form. Exact and structural, so it is the only join that never guesses. 2. A full PR URL in `name`, or an issue key or `#`-prefixed PR number bounded by non-alphanumeric characters and confirmed against the repo `cwd` resolves to. This covers sessions launched by hand. Do not match a bare substring: `#42` occurs inside `#142`, and every repo has a PR numbered 42. 3. `cwd` resolved to a repo, narrowing the candidate items, plus a semantic match of the name text against item titles.

`cwd` is the directory the session launched from. A session that entered a worktree mid-run still reports its launch directory, usually a main checkout, occasionally a path that has since been pruned. Use it to scope candidates by repo. Never derive a branch from it, and never report it as the work.

When an item's session is still unresolved and the user asks about that specific item, fall back to the `claude-code:session` skill's `search` query with the identifier as `query` and the repo as `project`. Never during gather, since it costs a query per item.

Handle three record classes explicitly. Skip any record whose `sessionId` is the current session. Surface `blocked` records, which are waiting on the user and are the reason this source exists. Surface `failed` records, since abandoned work reads as done work otherwise.

Brief

One prioritized brief, g