forgev0.19

One command. Many models. Every diff reviewed.

forge is a skill for your coding agent. It plans your goal into tasks, sends each one to the model and CLI you choose, has a second model read every real diff, and merges only what passes.

$ npm install -g @giamme/forge && forge install
Your first run How it works Runs inside Claude Code, Codex, OpenCode, OpenClaude, Antigravity and Grok.
Session waitingdemo run

forge reads the diff, not the claim

Every builder model says it's done. forge doesn't take its word. A different model, the reviewer, is handed the real diff and has to break it. Here is a real kind of catch, on demo code.

What the dwarf saidattempt 1

sol xhigh codex token-model

Done. Added the token model with issue, verify and revoke, plus expiry handling.

All 14 tests pass. Ready to merge.

attempt 2 gets both findings, in this same worktree, and fixes its own code. The next review reads the whole task again.

changed src/auth/tokens.py +36 -0
ran pytest -q tests/test_tokens.py
14 passed in 0.41s
said nothing about timing or revocation

forge: fail

Decided from the findings, not the reviewer's word: a confirmed P1 is at or above the P2 threshold.

A review with no verdict at all would fail too.

Retry

The findings go back to the dwarf. Nothing merges until a review passes.

What qa readopus high claude
src/auth/tokens.py +36 -0 2 findings
29+def verify(token: Token, presented: str) -> bool:
30+ if token.expires_at < datetime.now(timezone.utc):
31+ return False
32+ return token.digest == _digest(presented)
P1Confirmedline 32 reach=realistic origin=new evidence=traced
The digest comparison is not constant-time, so response timing leaks how much of a token matched. Use hmac.compare_digest.
33+
34+def revoke(store, token: Token) -> None:
P2Confirmedline 34
revoke deletes the stored digest, but verify never checks the store, so a revoked token keeps working until it expires.
35+ store.delete(token.digest)
# read in a throwaway copy of the repo; editing it would void the review FORGE_TESTS: RAN FORGE_VERDICT: FAIL

Any model, from any CLI

One name picks a model everywhere: alias:effort:cli. Mix providers in one command, keep the reviewer on a different family, and move a model to another CLI when an account runs dry. Try it: this reads forge's own registry.

Role
Modelthe alias you type
CLIthe first listed is its default
Effortstruck through: forge would clamp it
model
runs on
effort

Nothing runs until you approve the table

Ask for a parallel run and forge splits the goal, rates each task by the judgment it needs, and routes it by your rules. Then it stops and shows you this. Retier a row and its models change. Approve, and the waves run.

Approval table/forge "add auth: tokens, middleware, login, rate limit, docs" --decompose-level medium --profile balanced
StateTaskDifficultyBuildsReviewsWaveApproach
token-model1Token dataclass, issue and verify, digest at rest; owns src/auth/tokens.py
middleware1Bearer check before every /api route; owns src/auth/middleware.py
login-route2POST /login issues a token through the token model; waits on both
rate-limit2Token bucket per IP on /login, in the middleware's hook
docs3docs/auth.md: issuing, verifying and revoking tokens

Own worktreeEach row runs alone in its own git worktree on forge/a1c4/<task>, with a fresh session and a brief of what already merged.

WavesWave 2 waits for wave 1 to merge; tasks run together only when they own different files, and forge says why one waits.

ApproachOne planner wrote every row's approach, so the seams between tasks fit. The reviewer gets the same plan.

Steer it while it runs

A run can take an hour. You stay in charge the whole time: answer the questions only you can answer, resize it, pin a model, retry, or accept a task the reviewer won't let go, from the terminal or from forge-ui.

Brieflogin-route must answer

Should a failed login say whether the email exists?

Why it matters: telling them apart helps users, and helps attackers list accounts. The task can't pick for you.

default if unanswered: one message for both cases

  • Brief modeBefore a task starts, a cheap read-only model writes the questions only you can answer, with options and a default. gated holds a task until its must-answer questions are answered.
  • Task summariesA few plain sentences per task, rewritten only when its attempts or review change, so you know where it stands without reading logs.
  • Retro notesBuilders and reviewers note what got in their way: a missing tool, a flaky test, a wrong instruction. You decide what becomes an issue.
forge controlrun a1c4 live
# press a control. each one is a forge command,
# logged with who asked and applied at the scheduler's next pass.
  • Infra isn't failureA rate limit, quota, sign-in or network stop spends no attempt. forge pauses only the model or CLI that's stuck, moves sideways to a healthy one in the same tier, or waits it out.
  • EscalationA task that keeps failing moves to a stronger model from your own pools, and the table shows who decided.

Every run, live, in 3D

forge-ui is the optional local viewer forge offers at the start of a run. This is its run cloud, drawn here live: every run as a block, grouped by repository, lamps lit by state. Drag it, hover a run, change the layout.

The plateA run's tasks in their waves, then on the clock attempt by attempt, with the critical path lit.
Findings on linesExactly the diff each reviewer read, every P0 to P3 finding under the line it names.
Each conversationWhat every dwarf and reviewer said, ran and changed, read from its CLI's own log.
Projects, worktrees, diskEach repository's runs, findings and memory; every worktree forge made; what forge takes on disk.

Measured on your code, not on benchmarks

forge remembers what each run taught it about your repository, and records every run's inputs and outcomes outside your repos. Over time you know which models pass first time on your code, and what a reviewer's PASS was worth.

# .forge/memory.md  (small, capped, given to every run)
## Verify
- pytest -q runs the suite; make test also lints and is 4x slower

## Known traps
- src/proto/*_pb2.py is generated, never hand-edit  [src/proto]

## Recurring QA findings
- 3x: route added without updating openapi.yaml  [src/routes]

The model that hit the landmine writes it down. A review finding has to show up in two runs before it counts, a fact about a deleted file is dropped, and the file never grows past 40 lines.

Modelsdemo data first attempts
ImplementerTasksFirst-attempt passRetriesMedian time
solxhigh4779% 1319m18s
opushigh2878% 721m28s
sonnetmedium4261% 1711m08s
haikuxhigh4268% 1113m27s
grokhigh4564% 1413m49s

Observational, not causal: forge sends hard tasks to strong models, so a model that gets the hard ones looks worse than it is. forge-ui says so too.

What forge never does

forge works next to your code, not on top of it. Every run keeps these.

  • It never pushes.There is no push step. Your remote changes only when you push.
  • It never commits to your branch.Parallel work lives on branches forge makes for itself until you run integrate --approved. A single task's change stays uncommitted.
  • It never picks a model for you.Every dispatch spends your quota. Leave out the builder and forge asks.
  • It never merges unreviewed code.Only the exact commit a reviewer passed can merge. Overriding a reviewer takes a written reason and --approved.
  • It never deletes a failed attempt.The branch and worktree stay: the record of what went wrong and the place to retry from.
  • It never drops a sandbox on its own.Only --yolo-dwarf or --yolo-qa lets a role out, and only that role.

Built with forge

forge develops itself. Its own history is full of runs: tasks attempted, reviewed, retried and merged by forge, then checked by a person.

  1. 0.2
  2. 0.3
  3. 0.4
  4. 0.5
  5. 0.6
  6. 0.7
  7. 0.7.1
  8. 0.8
  9. 0.9
  10. 0.10
  11. 0.11
  12. 0.11.1
  13. 0.12
  14. 0.12.1
  15. 0.13
  16. 0.13.1
  17. 0.14
  18. 0.14.1
  19. 0.15
  20. 0.16
  21. 0.17
  22. 0.18
  23. 0.19

23 releases since 6 October 2026.

# git log on forge's dev branch
d8a6eed forge: merge sweep-1
68114ae forge(sweep-1) attempt 3: Fix 4 review findings in scripts/forge-parallel.sh
5cfa8df forge(sweep-1) attempt 2: Fix 4 review findings in scripts/forge-parallel.sh
06fcecd forge(sweep-1) attempt 1: Fix 4 review findings in scripts/forge-parallel.sh
b262e21 forge: merge sweep-2
595f6ae forge(sweep-2) attempt 1: Fix 4 review findings in scripts/forge_brief/proposals.py
61ef643 forge(sweep-5) attempt 1: Fix 1 review finding in scripts/forge-lib-plans.sh
4d6349d forge(sweep-3) attempt 1: Fix 1 review finding in scripts/forge-groups.py

What's coming

forge moves fast and in the open. Here is what 0.19 brought and what's planned next, straight from the issue tracker.

Hand forge one goal

Install it, open your agent CLI in a repo, and give it something small you'd normally do yourself. Pick the builder; the reviewer defaults to opus.

$ npm install -g @giamme/forge && forge install

# then, inside Claude Code, Codex or any CLI forge found
/forge "make the cache key include the tenant id" --dwarf sol:high