From c783c10f5a574084dc0e6f4dfeb8201442a6e741 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Oscar=20S=C3=B6derlund?= Date: Thu, 19 Nov 2026 10:38:57 +0100 Subject: [PATCH 1/2] docs(agent-team): require verified URLs from web research - Web research agents work in a vacuum and never read this skill, so the orchestrator must put the citation rules in their unit prompt. - Require a URL for each key reference, fetched in the same session. - Drop dead, unrelated, or unsupported URLs, and flag outdated sources. - End report.md with a References list. --- agents/skills/agent-team/SKILL.md | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/agents/skills/agent-team/SKILL.md b/agents/skills/agent-team/SKILL.md index 478b3321..5a9112b9 100644 --- a/agents/skills/agent-team/SKILL.md +++ b/agents/skills/agent-team/SKILL.md @@ -53,6 +53,14 @@ Workers never need approval: Herdr starts claude and agy with `--dangerously-ski For bulk and web research, agy writes `report.md` and the orchestrator reads only that file. +Web research is vacuum work, so put these rules in its unit prompt: + +- Give the URL of each key reference next to the claim it supports. +- Fetch every cited URL in this session (`defuddle parse --md`). Never cite a URL from memory. +- Drop or replace a URL that fails, redirects to an unrelated page, or does not support the claim. +- Record the publish date, update date, or version of each source. Flag a source that a newer release or doc replaces. +- End `report.md` with a References list: URL, date or version, and one line on what it supports. + Responsibilities do not cross unless assigned: design does not ship code, implement does not redefine the outcome, write does not touch implementation or architecture, verify does not expand scope, bulk does not touch architecture. Write output is short sentences in the repo `AGENTS.md` style. ## Flow From 7252e83c0576504c3cda33ee59d8eec7c8c9bbfd Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Oscar=20S=C3=B6derlund?= Date: Thu, 19 Nov 2026 10:38:57 +0100 Subject: [PATCH 2/2] feat(herdr): pick Linear projects in start-workspace - The Linear picker listed only issues, so a project with no issues could not be picked. - List open projects of the default team next to issues; mark one with Tab to track it by slug ID. - A bare Enter now picks nothing; Enter accepts only marked lines and otherwise returns the typed query. --- herdr/start-workspace | 25 +++++++++++++++++++------ 1 file changed, 19 insertions(+), 6 deletions(-) diff --git a/herdr/start-workspace b/herdr/start-workspace index ab16b185..f5a09097 100755 --- a/herdr/start-workspace +++ b/herdr/start-workspace @@ -346,23 +346,36 @@ function _pick_linear_context --argument-names dir test -e "$file"; and return 0 command -q linear; or return 0 - _status_show "Loading Linear issues..." + # Lines are kindidlabel; fzf shows only the label. + _status_show "Loading Linear issues and projects..." set -l lines (timeout 10 linear issue list -s started -s unstarted --json --no-input 2>/dev/null \ - | jq -r '.[] | "\(.identifier)\t\(.title)"' 2>/dev/null) + | jq -r '.[] | "issue\t\(.identifier)\t\(.identifier) \(.title)"' 2>/dev/null) + set -a lines (timeout 10 linear project list --json --no-input 2>/dev/null \ + | jq -r '.[] | select(.status.type | IN("backlog", "planned", "started", "paused")) + | "project\t\(.slugId)\tproject: \(.name) (\(.status.name))"' 2>/dev/null) _status_clear + # Enter accepts only tab-marked lines; with none marked it returns just the + # query, so a bare enter picks nothing. set -l picked (printf '%s\n' $lines | fzf --prompt="Linear> " --height=100% --reverse \ - --multi --print-query --delimiter=\t --bind='enter:accept-or-print-query' \ - --header='tab: select · type issue IDs or a project slug · esc: skip') + --multi --print-query --delimiter=\t --with-nth=3 --with-shell='sh -c' \ + --bind='enter:transform:[ "$FZF_SELECT_COUNT" -gt 0 ] && echo accept || echo print-query' \ + --header='tab: select issues and one project · type issue IDs or a project slug · esc: skip') test $status -le 1; or return 0 - # Selected list lines win; otherwise read issue IDs and one project from the query. + # Marked list lines win; otherwise read issue IDs and one project from the query. set -l issues set -l project set -l selected (string match -- '*'\t'*' $picked[2..-1]) if test (count $selected) -gt 0 for line in $selected - set -a issues (string split -f1 \t -- "$line") + set -l fields (string split \t -- "$line") + switch $fields[1] + case issue + set -a issues $fields[2] + case project + test -z "$project"; and set project $fields[2] + end end else for token in (string split -n ' ' -- "$picked[1]")