Skip to content

ai-workflow.md

Latest commit

 

History

History
110 lines (82 loc) · 4.3 KB

ai-workflow.md

File metadata and controls

110 lines (82 loc) · 4.3 KB
Raw

AI Workflow Reference

Daily Modes

Interactive (default)

cd ~/myproject && kiro-cli chat

Model: auto (1.00x). Hands-on work -- design, debugging, exploration.

Parallel dispatch (active, you're present)

cd ~/firstmate && kiro-cli chat

Talk to firstmate, it spawns crewmates in zellij tabs via treehouse worktrees. You answer escalations, approve merges.

Background unattended (you're away)

cd ~/myproject && gnhf --agent "acp:kiro-cli acp --trust-all-tools" "objective"

Long-running loop. Each iteration commits. Set --max-iterations or --max-tokens to cap spend.

Critical Incantations

Firstmate crewmate raw launch command

kiro-cli chat --trust-all-tools --no-interactive "$(cat __BRIEF__)"

Pass this as the harness arg to bin/fm-spawn.sh. The __BRIEF__ placeholder gets substituted.

gnhf with kiro-cli

gnhf --agent "acp:kiro-cli acp --trust-all-tools" "reduce test flakiness"

Treehouse (instant worktree)

cd ~/myproject && treehouse       # drops into isolated worktree subshell
exit                              # returns worktree to pool
treehouse status                  # see all worktrees

Firstmate Config (~/firstmate/config/)

File Value Purpose
backend zellij Session provider for crewmate tabs
crew-harness claude Default adapter (override with raw cmd at spawn)

To spawn with kiro-cli explicitly, firstmate passes the raw command as the third arg to fm-spawn.sh.

Model Routing (from 2026-07-04 benchmark)

Task type Model Credits
Creative Blender / spatial reasoning auto (minimum), opus for hero 1.00-2.20x
Code implementation, bug fixes auto 1.00x
Mechanical refactors, simple Q&A haiku-4.5 0.40x
Long background loops (gnhf) auto 1.00x
Avoid for creative/spatial haiku, minimax, qwen3 fail the task

Auto (1.00x) beat sonnet (1.30x) on cost AND quality for Blender work. Everything below auto fails at creative tasks.

Tool upkeep

Chezmoi installs missing workflow and AXI tools through run_after_install-agentic-tools.sh and run_after_install-axi-clis.sh. Existing tools are upgraded explicitly with agent-tools-update --apply; run it without flags first to inspect the plan. The updater follows each active binary's installation manager and requests the current release rather than storing a version table here. It covers installed agent clients, AXI packages, workflow helpers, and a clean tracked firstmate checkout. A failed update remains visible and exits nonzero; an installed binary alone is not evidence that it is current. Inspect native updater output for declined or skipped updates, since some clients exit successfully even when an update was not applied. Run updates from a host terminal with package-install and network access.

Bundled AXI skills follow their package updates. Cursor discovers those skills through the symlinks maintained by run_after_install-axi-clis.sh. Claude Code and Codex also receive AXI discovery through their existing startup hooks. Shared personal skills remain version controlled in chezmoi.

The delegation hint is rendered from one chezmoi template into each client's native global instructions. It is placed near the start so Claude, Codex, Cursor, and the other clients receive the same standing preference. Capacity checks validate observed behavior and response data, never an exact CLI version. Unknown quota allows a bounded manual worker choice when appropriate.

Zellij Notes

  • Firstmate creates session named firstmate with one tab per crewmate (fm-<task-id>)
  • Attach to watch: zellij attach firstmate
  • Dead EXITED sessions block new ones: zellij delete-session firstmate --force
  • Peek at crewmate: zellij -s firstmate action dump-screen --pane-id <id> --full --path /tmp/peek.txt

Preventing Rot

  • Regularly: agent-tools-update --apply, then inspect any failed or manual update steps.
  • After updates: run agent-capacity --json --task routine and verify the native instruction context in new client sessions.
  • Quarterly: Re-run model benchmark (~/blender-ab-test/run-experiment.sh) -- model pricing and quality shift
  • When things break: Check bin/fm-bootstrap.sh output for missing tools