docs: frame project and agent workflow exploration

Record the reported navigation and coordination friction, verified native capabilities, and provisional workflow comparison. Keep proposed delegation and tool choices distinct from accepted decisions; collect workload, persistence and trial constraints together.
This commit is contained in:
kiri committed 2026-09-09 17:50:50 +02:00
1 parent c8cbc90d87
commit 7514ceb04f
2 files changed
+49 -10

No files matched your search

+39 -7
View File
@@ -122,13 +122,45 @@ The focus is the broader environment, not further API/database polish. Candidate
areas remain places/ownership, navigation/interaction, human/agent collaboration,
and system operation/maintenance; these are discussion areas, not a work queue.
Agent-session navigation was the last open conversation: agent-deck is already
installed (previously observed as 1.15.0). Its help offers native project sessions,
search, attach and Ctrl+Q detach, with optional groups. No sessions were inspected
or migrated. Start from the user's experience before changing bindings or selecting
it over Herdr. Proposed roles: Walker enters places/apps, Niri navigates windows,
and the session manager resumes agent work. Avoid duplicating its session list or
imposing groups before use demonstrates a need.
Current discussion: consistent project/agent work and Niri workspace use. The user
values Herdr's spaces and overview of running/waiting agents, but dislikes its
keybindings and mouse orientation. Agent-deck has startup issues (not diagnosed).
Kitty/Niri, tmux and Zellij are candidates, not selected replacements.
Read-only inspection found existing Niri overview, workspace movement and tabbed
columns; Kitty has no declared project/session organization. Installed versions:
Kitty 0.48.2, Niri 26.04, tmux 3.7c, Zellij 0.45.1 and Herdr 0.9.0.
Herdr's installed default config confirms remappable navigation and agent jumps.
Current [Herdr keyboard docs](https://herdr.dev/docs/keyboard/) describe direct
chords; [Kitty sessions](https://sw.kovidgoyal.net/kitty/sessions/) supply native
session creation/switching. Session layouts and window overviews do not themselves
provide authoritative agent attention states. tmux detach keeps processes alive;
Zellij resurrection recreates layouts/commands, not process memory. None of these
capabilities was exercised interactively; no sessions were inspected or changed.
Proposed workflow, awaiting discussion: organize human-visible work by project/task;
one lead conversation owns each coherent change. Use bounded subagents when work
is independent, with explicit ownership and evidence returned to the lead. Separate
worktrees for independent writing efforts, one integration owner. Keep enduring
repository rules in AGENTS.md, current goals/acceptance in task prompts, and durable
decisions/current state in existing project docs. No fixed agent team is selected.
Proposed comparison: hold the desktop and model settings fixed; compare solo,
solo plus independent review, then bounded parallel delegation on the same starting
commit and task. Record accepted correctness, unnecessary complexity, human
interventions/rework, total completion time and aggregate usage including workers.
Use fresh isolated runs and common acceptance checks; repeat promising variants
before drawing reliability conclusions. Measure desktop navigation separately.
[Codex subagents](https://learn.chatgpt.com/docs/agent-configuration/subagents)
and [evaluation guidance](https://developers.openai.com/api/docs/guides/evaluation-best-practices)
support bounded delegation and task-specific evaluation; the proposed trial is
not a measured result. No benchmark framework or comparative runs created.
Inputs requested together: typical project/agent count and clients; required
persistence (project switch, terminal close, logout/reboot, remote connection);
a representative trial task and rough time/spend limit. These determine the first
concrete trial. No configuration changes, tests, activation or agent runs in this
discussion; documentation whitespace checked.
Mail search remains unresolved. The workbench's custom mail reader, inventory
pilot, justfile and old README are archived; native Garden/kubectl/K9s now own the