Guild Stack v3.0.0 released · Releases

Turn coding requests into reviewed team runs

Guild Stack turns one request into a planned, reviewed, evidence-gated team run inside Claude Code and Codex.

more… 
curl -fsSL https://guildstack.dev/install.sh | bash -s -- --host claude-code-cli

Live host · conformance-verified (31/31).

Follow the install guide →

Not ready to install? See a worked example first.

31/31 conformance captures on each live host, signed and hash-bound for v3.0.0.

The run route A task moves from plan to build, through an open review gate to qa, then ops, and ends at a verified receipt. plan build gate · open qa ops receipt ✓

receipt ✓ · v3.0.0 promotion

Claude Code
31/31
Codex
31/31
refusals held
18/18
attestors signed
3/3
release evidence · .guild/artifacts/release/v3.0.0/conformance.json how it’s verified →
Proof

From request to receipt, on the record

guild · run-20261013-152923claude-code
  1. > lead run "v3 release prep" (3 of its lanes shown)
  2. plan 3 lanes: qa · developer · design none waits on another
  3. dispatch qa test the new v3 site Σ 3 lanes
  4. dispatch developer fix bugs #217 and #219
  5. dispatch design pick the visual direction
  6. qa site build: 37 pages + 57 redirects PASS
  7. qa Lighthouse audit, home page: 100/100/100 PASS
  8. qa 2 checks FAIL-open: failed, logged, sent to the lead
  9. gate Codex review of qa: REQUEST_CHANGES → fixed → APPROVE ⛩ gate: approved
  10. dev 2 new tests: fail first, then pass · all 1624 hook tests pass
  11. dev 2 known-bug tests: PASS → XPASS (bug gone, test marks to update)
  12. gate Codex review of developer: SATISFIED in 1 round
  13. design 8 reference sites rendered and measured
  14. gate Codex review of design: APPROVED after 3 rounds
  15. receipt qa lane: done, receipt filed
  16. receipt developer lane: done, receipt filed
  17. receipt design lane: done, receipt filed ✓ receipt
  18. verify 3/3 receipts present · 3/3 review gates passed
run-20261013-152923 · 13h 20m ✓ verified
Real run, scrubbed · reconstructed from its lane receipts · run-id in footer
lane
One specialist's piece of the work.
gate
A review the lane must pass. Codex reads the work and gives a verdict, copied here as recorded: changes requested, or approved.
#217, #219
The two bugs that lane fixed: a status command that wrote files on a fresh repo and an init command you could not safely run twice.
XPASS
A test written to catch a known bug now passes. The bug is fixed; the test's label still says "known bug".
FAIL-open
A check failed and the lane wrote it down for the lead instead of hiding it.
receipt
The written record a lane hands back: what it did and the proof.
verify
The last check. Every lane needs a receipt and a passed gate.

3 lanes · 3 gates · receipt ✓

Hosts

Same lifecycle, honest about each host

Guild Stack reads each host's capability row and degrades in a fixed order. Every loss is labelled.

16 hosts

Live host, conformance-verified at 31/31. Every Guild Stack capability runs on a native Claude Code tool.

curl -fsSL https://guildstack.dev/install.sh | bash -s -- --host claude-code-cli
dispatch ● native Agent tool / agent-team Task
skills ● native skill autoload (.claude/skills)
shell ● native Bash
edit ● native Edit / Write
questions ● native AskUserQuestion
Product

Six things a run does that a chat does not

  1. 01 Plan

    Scope before code

    A clarified spec, then a team and a lane plan you approve before anything is dispatched.

  2. 02 Dispatch

    Lanes, not one agent

    Each specialist works one lane with bounded context and hands back a receipt.

  3. 03 Review

    Review that pushes back

    Receipts are checked for spec first, then quality. Findings need evidence.

  4. 04 Verify

    No claim without a diff

    The done-gate checks tests, scope, and success criteria against the change on disk.

After the receipt · human gate

  1. 05 Learn

    Memory you can read

    Ingested sources reach the local wiki only through a human gate. Harvest can be turned off and reverted.

  2. 06 Evolve

    Skills improve, gated

    Reflections propose skill changes. Nothing is promoted without your approval.

Docs at your depth

Pick your depth. The docs remember it.

  1. Start Install → first task → read the receipts Run one bounded task safely. Start with install →
  2. Control Commands → configuration → cost and tiering Steer, tune, and resume ongoing work. Open the command reference →
  3. Extend Dispatch → host adapters → conformance Render the lifecycle on a new host. Open hosts and adapters →
Knowledge

What one run learns, the next run can recall

run-20261013 · lane receipt

decision: keep install docs as the
  single source for hero commands
rejected: hand-written per-host strings
status:   candidate
  1. capture
  2. promote · human gate
  3. recall

next session · recall

> where do hero install commands come from?
.guild/wiki/decisions/hero-install-source.md
  the install doc; edit it, the hero follows
Illustrative example. The gate and paths are real; the content is synthetic.
Evidence

The numbers, with their sources

Run your first gated task today

Copy the install command →

DocsUse casesResourcesGitHubEvidence