The Last Stack is a small, curated set of agent skills for driving a LastDB-backed workflow with an AI coding agent — the agent layer that sits on top of LastDB and its two companion tools:
- Brain —
brain: long-lived notes over LastDB (the why: decisions, designs, milestones). - Kanban —
kanban: a task board over LastDB (the what's in flight: cards moving through columns).
The skills are written for agent harnesses that load
Agent Skills
(a SKILL.md per skill, discovered by the agent and invoked by name) — e.g.
Claude Code, Codex, Factory, OpenCode. They give an agent a consistent playbook
for filing tasks, driving a single task to a merged pull request, waiting on PRs
robustly, and closing out finished work.
Compatibility skill names from the old fkanban naming remain installed where
they are needed for scheduled prompts: fkanban-card-authoring points agents at
kanban, and fkanban-grooming points agents at kanban-grooming. New prompts
should use the kanban names directly.
The companion routines (routines/) are the engine that runs the playbook
on a schedule: small, parameterized prompts you register as scheduled (cron)
agents. The operator model is Generate -> Claim -> Reconcile -> Intake
friction: generate PR-sized work, claim and ship it through pickup, reconcile
merge/proof/board state, and turn fleet friction into Brain records before it
becomes board work. The skills assume this pipeline exists; the routines
provide it. See routines/README.md and the target fleet
cheat sheet in docs/routines-target-fleet.html.
One line — clone the repo and run setup:
git clone https://github.com/EdgeVector/last-stack ~/.last-stack && ~/.last-stack/setupsetup auto-detects which agent harnesses you have (Claude Code, Codex,
Factory, OpenCode) and registers every skill into each one. The skills stay in
the cloned repo; each harness gets a directory with a symlinked SKILL.md, so a
later git pull updates every installed skill at once. It also installs the
companion routines CLI automatically. Starting scheduled agent work remains
explicit: run routines install-daemon when you are ready to enable the fleet.
For Claude Code, setup also installs the bundled safety hooks and allowlists
the brain MCP read tools plus brain_put in ~/.claude/settings.json, so
scheduled routines can use the brain tools without unattended permission
declines.
Options:
~/.last-stack/setup --host claude # install for one harness only
~/.last-stack/setup --local # vendor into ./.claude/skills (this project only)
~/.last-stack/setup --uninstall # remove the registered skillsTo download the usable LastDB app stack in one pass, run:
~/.last-stack/bin/last-stack-install-appsThat installs the LastDB daemon and downloads Brain, Kanban, Situations,
Routines, Dogfood Graph, Org, LastSecrets, Search, and LastDB Browser. The
routines CLI is linked alongside the other commands; installing its daemon is
an explicit follow-up because it starts scheduled agent work. The browser
launcher is linked as lastdb-browser and reads the machine's own LastDB
socket. LastGit is intentionally excluded until it is stable enough for the
public bundle. See docs/lastdb-apps.md for the full
guide.
Prefer to copy skills by hand? Each skill is a self-contained directory under
skills/—cp -R skills/<name> ~/.claude/skills/.setupjust automates that across every harness and keeps them updatable.
Prefer the clean-only helper (refuses a dirty install tree; never force-resets):
~/.last-stack/bin/last-stack-self-upgradeManual equivalent when the tree is clean:
cd ~/.last-stack && git pull --ff-only && ./setupOr tell your agent "upgrade the last stack" — the included
last-stack-upgrade skill runs the helper and shows what changed.
Skills can cheaply check for a new version via bin/last-stack-update-check
(cached VERSION lookup; git HEAD checks are uncached; prints UP_TO_DATE /
UPGRADE_AVAILABLE / GIT_UPDATE_AVAILABLE / UNKNOWN).
Scheduled fleets: last-stack-routine-read auto-runs
last-stack-self-upgrade when the install is behind and clean, so routines do
not stay stuck on LAST_STACK_ROUTINE_STALE. Register the
self-upgrade routine as a 1–2h backstop. Keep
~/.last-stack free of local edits — develop in a portal worktree
(./bin/wt start …), never the install tree. Where things live:
instructions/run-dev-state-board.md
(RUN / DEV / STATE / BOARD).
Run
setupAFTER any gstack setup //gstack-upgrade. gstack./setupre-symlinks its own skills into~/.claude/skills/<name>; when a gstack skill shares a name with a Last Stack one (e.g. gstack's mermaiddiagramvs. the hand-drawn architectural/diagram) it silently replaces ours. Last Stacksetupre-points our links and finishes by runningbin/last-stack-verify-skill-links, which verifies every Last Stack skill still resolves into the Last Stack tree and repairs any that a foreign installer stomped. Run that guard standalone any time to check (--check) or repair.
Privacy-safe install health for the Exemem admin SPA (Kanban deliver path —
delivery_slice / lastdb.slice.v1). Not an admin SPA tab (that is a
separate exemem-infra card); this is the Mini publisher + dogfood deliver.
Payload (all non-secret):
| Field | Source |
|---|---|
version |
VERSION |
install_head_short |
install checkout git rev-parse |
self_upgrade_result |
last-stack-self-upgrade --check-only |
skill_link_status |
last-stack-verify-skill-links --check (ok / drift / error) |
# Write slim LastStackHealthSnapshot (key health-latest) on Mini:
~/.last-stack/bin/last-stack-publish-status
~/.last-stack/bin/last-stack-publish-status --json
~/.last-stack/bin/last-stack-publish-status --dry-run --json
# Stage (and optionally approve) a deliver to the admin kanban-consumer.
# Recipient keys are operational — reuse the enroll-kanban-consumer bundle;
# never commit them. Env names:
# LAST_STACK_ADMIN_RECIPIENT_PUBKEY
# LAST_STACK_ADMIN_MESSAGING_PUBLIC_KEY
# LAST_STACK_ADMIN_MESSAGING_PSEUDONYM
~/.last-stack/bin/last-stack-deliver-status --dry-run --json
~/.last-stack/bin/last-stack-deliver-status # stage only
~/.last-stack/bin/last-stack-deliver-status --approve # stage + sendv1 reuses the existing kanban-consumer identity (schema-agnostic deliver).
Mailbox poll + openDelivery stay on the admin consumer side (exemem-infra);
this repo only owns publish + stage/approve on Mini.
Dogfood evidence (2026-07-15, non-secret): staged + approved
message_type=delivery_slice, shared=1, schema
last-stack/LastStackHealthSnapshot, single record health-latest.
The Last Stack is homed in LastGit at lastdb:///last-stack; agent-authored
changes go through LastGit change requests with the required ci-required gate
from .lastgit/ci.sh. The GitHub repository remains the public read-only clone
and browse mirror for installers and documentation links.
The committed .last-stack/pr-venue marker is what makes the shared
last-stack-pr-venue helper route this repo to LastGit. Mirror synchronization
is an operational concern of the LastGit multi-repo mirror supervisor; do not
open ordinary development PRs against the GitHub mirror.
Treat reusable agent improvements as upstream candidates by default. When a
session produces a new skill, routine, permission pattern, or process rule,
first decide whether it is workspace-specific or generally useful. If it is
portable, make the change safe for this repo and upstream it here so every
installed harness can pick it up on the next git pull && ./setup.
The expected path is:
- Capture the rationale in the brain (
brain) when the change should survive the current chat. - File or update a board card (
kanban) for the delivery/audit trail. - Patch the shared skill or routine in
skills/orroutines/using placeholder-based, product-neutral wording. - Verify the changed prompt still works cold, without chat memory or local-only assumptions.
- Leave workspace-only details in the workspace's own agent docs, not in this pack.
This keeps one-off local improvements from quietly forking the fleet while still preserving project-specific rules where they belong.
| Skill | What it does |
|---|---|
| kanban | Board CRUD over LastDB — file/list/show/move/groom cards. |
| kanban-agent | Drive a card to merged, reconcile in-flight PRs, or validate post-merge END STATE checks. |
| fkanban-card-authoring | Compatibility shim for old prompts; use kanban for current card authoring rules. |
| fkanban-grooming | Compatibility shim for old prompts; use kanban-grooming for current board grooming rules. |
| kanban-setup | Bootstrap kanban on a fresh machine — install, init (resolve published schemas), doctor, optional MCP registration. |
| onecontext | Search prior Codex sessions, with guarded Aline usage and a JSONL fallback when Aline is unavailable. |
| registry-rotator | Generic engine for registry-backed scheduled routines: pick the most-overdue eligible entry, run its recipe, file cards, and stamp the registry log. |
| wait-merge | Robustly wait for a GitHub PR to merge by interpreting PR state, not a watcher's exit code. |
| close-out | The post-change loop: open a PR from a worktree, drive it to merged, file session papercuts, write a full brain report of what was done, file a follow-up card. |
| last-stack-upgrade | Update the stack in place (clean-only self-upgrade) and re-register the skills. |
| session-miner | Generic engine for mining recent agent session transcripts with profiles for papercuts, incidents, owner-stated knowledge, tooling friction, and revenant-watch (settled-dead product truth reanimated; Brain-only). |
And the routines (routines/) — parameterized scheduled-agent templates that
run the skills on a cadence. Operator-facing ownership is intentionally folded:
pipeline-health owns merge-babysit and drain-style pipeline unblock work, while
board-reconcile is the single board closeout/proof surface made from
kanban-watch, the always-on zero-LLM closeout, kanban-validate, and the
reaper.
| Routine | What it does |
|---|---|
| self-improvement-loop | Mine recent sessions for friction; upgrade the agent's own skills/routines/permissions. |
| revenant-watch | Daily session-miner profile: flag agents reanimating settled-dead product truth (Brain-only ledger; open-work exemption). |
| papercut-reconciler | The ONLY papercut→card path: harvest session papercuts into Brain records, cluster ALL open Brain papercuts into patterns, file pattern-level cards. |
| devops-continuous-improvement | Inspect CI, merge flow, deployment, testing, and release gates; ship one small DevOps fix or file precise cards. |
| worktree-cleanup / disk-reclaim | Prune stale worktrees/branches; reclaim disk; keep the machine healthy. |
| drain-open-prs | Drive every open PR across all repos toward zero (merge or close). |
| kanban-pickup / kanban-watch / kanban-validate | Drain ready PR queue; reconcile PRs; proof lane (DONE-WHEN + backlog validation proofs + post-merge END STATE). |
| groom-board / north-star-driver / milestone-driver | Promote ready work and turn North Star intent into milestones and PR-sized cards. |
| program-rollup / consolidate-brain / morning-sync | Mirror the board into the brain; keep statuses honest; deliver the daily decision briefing. |
program-driver is DROP/superseded: do not use it as a milestone driver and do
not add new feature-owner cards. New feature or North Star work flows through a
Brain North Star, MILESTONE_REQUEST, north-star-driver, and
milestone-driver.
See routines/README.md for how routines + skills compose,
and fill in the <PLACEHOLDERS> before scheduling any of them.
For a new project, start with the routine fleet bootstrap guide and record
templates in docs/routine-fleet-portability.md
and templates/routine-fleet/.
The registry-rotator skill expects each project registry to be a Markdown
record with one rotatable entry per heading, entry fields for track,
cadence, recipe, pass =, and isolation, and a single
rotation-log:start/end table with feature, last_run, result, and
cards filed columns. The engine reads project paths and venues from the
project's config source (workspace-config / repo-venue-map while those are
still interim brain shims), selects the eligible entry with the largest
age / cadence overdue ratio, dispatches that entry's recipe, files kanban
cards per sop-routine-shared-contract, and rewrites only the rotation-log row
for the selected entry. Scheduled tasks should be thin triggers that pass a
registry slug such as registry=dogfood-registry.
VERSION the installed version (update-check compares against this)
setup installer — registers skills into your agent harnesses
lib/
lastdb-http.sh shared Mini socket HTTP helpers for publish/deliver
bin/
last-stack-update-check is a newer version or default-branch HEAD available?
last-stack-self-upgrade clean-only FF pull + ./setup (used by routine-read)
last-stack-routine-read serve a routine prompt; auto-heals stale clean installs
(version checks cached; git HEAD checks uncached)
last-stack-verify-skill-links
verify (and by default repair) that every Last Stack
skill link still resolves into the Last Stack tree;
undoes a gstack same-name skill stomp. setup runs it
as its final step. --check reports only.
last-stack-publish-status write slim LastStackHealthSnapshot to Mini (VERSION,
install HEAD, self-upgrade check, skill-link verify)
last-stack-deliver-status publish + stage/approve lastdb.slice.v1 to the admin
kanban-consumer (routines deliver-status pattern)
last-stack-shell-prelude sourceable PATH prelude for scheduled routines
(also bypasses host-global sccache by default for
predictable parallel Cargo builds)
last-stack-cargo Cargo entrypoint that explicitly bypasses a
host-global rustc-wrapper unless opted back in
last-stack-sccache-health report pinned-full cache state; exit 10 when the
current shell still routes builds through it
last-stack-cli-preflight verify routine-required global CLIs are on PATH
last-stack-json-get extract one simple field path from socket/API JSON
without relying on jq or inline python/node parsing
last-stack-repo-op-guard reject workspace roots before repo-scoped git/gh
last-stack-pr-venue route a repo to github, forgejo, or explicit
LastGit-native CR handling before PR/CR operations
last-stack-gh-pr-queue-state
GraphQL PR queue-state helper without gh -R drift
last-stack-forge-ci-log print a failing forge (Forgejo) CI job's log tail —
resolves the run + attempt-scoped web log endpoint
(run with dangerouslyDisableSandbox: TCP to :3300 is
sandbox-blocked)
last-stack-forge-api call the local Forgejo API with keychain/ env token
auth and optional control-char-safe jq projection
last-stack-forge-git run git against local Forgejo remotes with the token
injected as an HTTP extraHeader when needed
last-stack-forge-json-jq control-char-safe jq wrapper for Forgejo API JSON
status polls
last-stack-forge-runner-lanes
discover/verify merge-gate vs dedicated `heavy`
release/deploy runner capacity (see
docs/forge-runner-lanes.md); proof:
`bin/last-stack-forge-runner-lanes --check`
last-stack-routine-read freshness/missing-file guarded routine prompt reader
last-stack-dogfood-target-checkout
select a current dogfood checkout without mutating
the recipe's original target checkout
last-stack-git-checkout-freshness
non-mutating tracked-remote freshness preflight
last-stack-brain-append-heartbeat
append a fleet heartbeat line to a **filesystem**
log (NOT LastDB/brain). Default:
~/.last-stack/logs/routine-heartbeats.log
last-stack-heartbeats-path
print the heartbeats log path
last-stack-board-drain-report
print board position plus 1h/6h/24h pickup drain
velocity from routine-heartbeats and run metadata
last-stack-active-programs-guard
reject active-programs rewrites that drop program
headers/slugs; split closed programs into archive
last-stack-install-apps download LastDB plus the usable app stack
last-stack-lastdb-current maintain ~/.lastdb/current plus ~/.local/bin
lastdb/lastdbd/folddb shims; optionally rewrite a
LaunchAgent plist without restarting lastdbd
last-stack-uninstall remove the registered skills
skills/<name>/SKILL.md one directory per skill
(includes lastdb-safe-upgrade — multi-harness Mini safe upgrade)
instructions/brain-kanban.md
canonical brain/kanban usage guidance; setup upserts it
as a managed block into each harness's global
instructions file (`~/.claude/CLAUDE.md`,
`~/.codex/AGENTS.md`, `~/.factory/AGENTS.md`,
`~/.config/opencode/AGENTS.md`) and registers the
brain/kanban MCP servers for Codex (with a PATH env so
GUI-spawned servers can find bun); also records the
creation-time default that new repos start in LastGit
while existing venue choices remain unchanged until
explicitly migrated
routines/<name>.md one parameterized scheduled-agent template per routine
routines/README.md how routines + skills compose; how to register them
templates/routine-fleet/
portable Brain record templates for project routine
config, probe registries, and shared SOPs
The intended loop, end to end:
- File work as cards (
kanbanskill). A card body is the spec: aRepo:/Base:header plus GOAL / STEPS / VERIFY / DONE WHEN. Cards live on your LastDB node. - Drive one card to merged (
kanban-agentskill, WORK mode). The agent claims the card, works in an isolated git worktree, opens a PR, and then drives that PR to MERGED — re-arming auto-merge, updating a behind branch, rebasing conflicts — before moving the card todone, or leaving it intodo/doingwith a clearBLOCKED:marker when the END STATE requires an async post-merge check. A card is onlydonewhen its code is actually in the repo and the outcome is proven. - Wait on PRs without false failures (
wait-mergeskill). Interprets PR state rather than trusting a watcher's exit code, so transient CI/queue churn doesn't look like a failure. - Reconcile the board (
kanban-agentskill, RECONCILE mode). A scheduled sweep moves merged-but-unadvanced cards todoneand nudges stuck PRs — leaving un-started cards alone. - Validate post-merge outcomes (
kanban-agentskill, VALIDATE mode). A scheduled pass runs one dev-only post-merge END STATE check, then moves the card todoneon pass or leaves it visibly blocked withPROOF:plus a fix card/blocker on fail. - Close out (
close-outskill). After any substantive change: PR from a worktree, drive to merged, file session papercuts (brain papercut file), write a full what was done report to the brain (brain), and file any follow-up as a card (kanban).
The two halves are deliberate: the brain records why; the board records
what's in flight. Keep decisions in brain and active work in kanban, and
the agent always has both context and a worklist.
Steps 1–6 describe what an agent does when invoked. To make the loop
self-driving — so cards get filed, promoted, picked up, and reconciled
without a human kicking it each time — register the routines as scheduled
agents: generators (self-improvement-loop, papercut-reconciler) file work,
north-star-driver, milestone-driver, and groom-board promote it, separate
kanban-pickup workers claim and ship one WORK card each, pipeline-health
and board-reconcile (kanban-watch plus closeout/reaper) reconcile the
stragglers, kanban-validate
runs post-merge END STATE checks, and
program-rollup/consolidate-brain/morning-sync keep the brain honest and
surface the short genuinely-human decision set. See
routines/README.md.
The session-miner skill is the shared engine behind transcript-mining
routines. A scheduled task can become a thin trigger that passes a profile name
such as papercuts, incidents, owner-statements, friction-patterns, or
revenant-watch plus a time window; the skill handles transcript parsing,
dedupe, report-only dry runs, and profile-specific writes. revenant-watch
writes Brain records only (never kanban cards from the miner itself).
You'll also want the underlying tools installed and a LastDB node running — see
the kanban-setup skill and the kanban / brain repos.
Every skill talks to a LastDB node over HTTP. The node URL is configurable —
kanban init defaults to a node running locally on your machine, and you
override it with --node-url (and --schema-service-url for a different schema
service). Point the skills at whichever node hosts your board and brain.
MIT — see LICENSE.
See also docs/lastdb-no-product-scan.md.