- Skills: per-project know-how the platform proposes from each succeeded
run (in the background, or on request from a run), people confirm,
edit, disable, delete or write; active ones enter later prompts
- Export one skill as SKILL.md, or all as a zip with a directory per
skill, as .claude/skills/ expects
- A wish's report: every run, its stage timeline, the agent's plan,
replies, commands, changed files and model usage, the change, the
checks, failures, reviews and releases, and the full event log, as
Markdown, or laid out for printing to PDF
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- countedTokens: input, output and cache creation in full, cache reads a
tenth, as they cost; the worker meters Claude CLI usage with it and the
usage event says how much counted
- A run's token budget may be raised to 10,000,000
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Projects list the models their wishes may run with, the first by
default; owners and admins set them in the project settings
- A wish may choose one of its project's models; the choice is checked
and pinned to the run, which shows it, and the executor starts the
agent with it (--model), else with the agent configuration's
- Model names may be spelled as the model gateway does
(deepseek-pro@latest[1m]), in projects and agent configurations alike
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- cluster-poc.mjs: image (in-cluster Kaniko), bootstrap, secrets, apply,
migrate --rehearsal|--final, public HTTPS entry, backup and restore
- One restricted Pod (platform, executor, Runner, PostgreSQL, Temporal,
MinIO, Caddy, backup) listening only on loopback except Caddy
- Protected checks run in per-workspace check Pods; cancellation during
start-up now always removes them
- Daily snapshots on the node (consistent dump with row counts, Temporal
online backup), copied to the Mac by launchd with SHA-256 checks and a
notification on failure; restore verified on the rehearsal copy
- Downloads and uploads survive the Mac's lossy link to the cluster
- Local public entry, evidence isolation labels, acceptance records
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Each wish says whether it may change backend code. A frontend-only wish is told
which paths it must leave alone; a change to them goes back before any check runs,
the platform puts those files back before the repair round, and the control plane
never integrates them. A wish that may change backend code integrates directly for
people with backend permission and waits for a code review otherwise.
- Owners and admins grant backend permission per person on the members page as well
as per project in the project settings.
- AI optimization has its own button for the wish and for the criteria. It writes a
wish of at most three lines (goal, scope, limits), keeps the details in checkable
criteria, and asks only what the person must decide, as choices whose answers join
the wish; suggested criteria are taken one by one.
- Lists, run titles, Feishu cards and pull request titles show a wish's first line;
the run details show the rest.
- Changed paths no longer follow renames, so moving backend code into a frontend
directory counts at both places for code review and for the frontend-only scope.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What the pilot asked for, from its meeting:
- Code review: a project names the paths changes to which a person reviews
before they reach the target branch (globs, with exceptions). After the
checks pass, the verifier asks the control plane, and the run waits at
integration until an owner or admin approves the exact commit; approving
integrates it without repeating the checks, rejecting hands the comment to
the agent as a new requirement. Continuing cannot get around it: the
verifier asks again before every integration.
- Undo: "take back this change" starts a run that reverses what a run
integrated on the current target branch, three-way, keeping later commits;
only lines later commits changed too go to the agent. It then verifies,
integrates and delivers like any run. Releases get a rollback button for the
version live in their environment.
- Wish optimization: the local Claude account rewrites a draft wish and its
acceptance criteria by built-in and project rules, with no tools at all;
the person compares and decides, nothing is submitted for them.
- Release confirmation: an environment can make every release wait for a
person, not only those the risk rules flag.
- Feishu notifications: a group bot is told of releases awaiting approval and
changes awaiting review (mentioning the decision makers), runs that need a
person, finished runs and failed releases; each once, signed when a secret
is set, with text from runs never acting as card markup.
- Reviewers see the agent's change summary and who submitted the wish, and
whether the checks ran in the macOS sandbox; the wish form offers only the
executors this deployment can run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In the development host mode, protected checks and dependency
installations ran directly as the operator. They run the agent's code:
a test it wrote could read the home directory's secrets, reach the
network, or plant Git configuration that the trusted Git commands after
it would execute.
With WF_HOST_CHECK_SANDBOX=true they run under Seatbelt instead. File
contents in the home directory are unreadable except the command's own
directory and the paths the operator lists; it may write only there,
in the caches the operator lists and in a temporary directory of its
own, never in a .git directory. Only installations reach the network;
checks may use loopback ports by opt-in, for test suites that start
local servers. The platform's own ports (control plane, Temporal and
its web UI, database, object storage, provisioner, Git service, IDE,
console) stay blocked in any case: the local start lists them from
their configured addresses. Checks get no answer from the keychain,
and a command whose home is the executor's (holding its credentials)
gets a private one. A sandbox that cannot start leaves the run waiting
on infrastructure; verification evidence records checkSandbox.
scripts/check-host-check-sandbox.ts checks all of this on the target
Mac, from outside the sandbox, then runs the POC's installation and
checks sandboxed on a recent droboticflow source. Enabled on the local
POC, a real wish passed through both, including a repair round for the
agent's own failing tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An attempt that stopped before its agent started (a failed
installation, an unavailable sandbox, a pause) kept its whole clone and
a recovery note on disk, a new one on every continue, although nothing
in it was worth recovering. Only an attempt whose agent ran and whose
checkpoint was not saved now keeps its workspace.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A check stopped at its time or output limit lost everything it had
printed: the repair prompt only said that WishForge stopped it, not
where it hung. The stop now carries the output so far, and the
verification evidence and failed installations keep it, with the
reason appended.
- A command that left a process running (a test's server, say) held
its pipes open until the time limit and counted as hanging, and in
host mode such processes stayed on the machine. A command now ends
when its own process exits: the rest of its process group gets two
seconds to drain the pipes and is then stopped, as is anything else
it left running.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pausing revoked the executor at once: the agent was killed and the work
of the current attempt stayed only in the executor's local workspace,
so continuing started that attempt over.
A pause of a run whose executor holds the lease now drains it first:
- The run keeps `running` with pauseRequestedAt (the console shows
pausing). Its agent gets no model credential and no new model
request; continuing is refused until the pause completes; cancelling
still stops it at once.
- The executor learns of it from its heartbeat, stops the agent, saves
a checkpoint with its live lease and completes the pause through
/internal/runs/:id/pause. Continuing restores that checkpoint.
- The workflow is stopped only after the pause completes; stopping it
earlier would abort the save. A model request interrupted by the
stop no longer expires the pausing run's lease.
- If the executor does not complete it within two minutes, or stopped,
the heartbeat, the next lease, the coordinator's wait or the
background sweep pause the run anyway.
Runs no executor is working on still pause at once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every failed installation waited as an infrastructure fault. When the
agent's own change broke it (a misspelled package, say), continuing
installed the same manifests before the agent ran, failed again, and
the agent never got a turn: the run could not move without a new
requirement revision.
- The verifier attributes a failed installation to the checkpoint when
it changed dependency manifests (package.json, go.mod, lockfiles and
the like) and the output is not a network or registry failure. The
evidence lists those manifests, and the control plane records an
`install` failure that the next repair round fixes.
- A repair round for an `install` failure starts the agent even when
the installation fails again, with that output in the prompt.
- Prompts, lessons and the run page name the failed installation.
Any other failed installation still waits for the environment, without
spending a repair round.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With WF_CLAUDE_COMMAND_SANDBOX=true, Claude Code also gets Bash, run in
its built-in sandbox through --settings:
- The sandbox must start (failIfUnavailable), and no command may run
outside it (allowUnsandboxedCommands: false).
- No network: an empty allowedDomains with strictAllowlist, since an
absent allowlist leaves the network open. Local binding stays off,
because it would also open every local port, Temporal's included.
- Commands cannot read the home directories, the CLI's home or the
worker's own directory, except the workspace, a per-run scratch
directory and WF_CLAUDE_COMMAND_READ_PATHS; they write only to the
workspace and the scratch directory; credential variables are unset.
- WF_CLAUDE_COMMAND_ENV gives commands toolchain variables, such as the
POC's read-only Go module cache.
The executor probes the sandbox first (Seatbelt on macOS; bubblewrap
and socat on Linux) and keeps file tools only where it cannot start,
saying so on the timeline. The adapter now writes the execution
contract into the prompt, for the tools the agent actually has.
scripts/check-claude-command-sandbox.ts checks a machine with one short
real session and judges each probe by its effect, not the agent's
account. On the local Mac every probe was blocked; the login keychain
remains reachable, because Claude Code's Seatbelt profile always allows
the security server. That limit is documented.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude Code with a tenant key started with bypassPermissions and every
tool, Bash included, and loaded the repository's settings, hooks and MCP
servers. It now runs like the native development login already did:
restricted mode with dontAsk, only Read/Edit/Write/Glob/Grep, confined to
the workspace, and no repository or user settings. Every Claude Code run
is told the verifier runs installs and tests.
CodeBuddy's agent runs commands with no sandbox of its own. Like Codex
without its sandbox, it now starts only in a development (host) sandbox,
says so on the timeline, and stops with PROVIDER_SANDBOX_UNAVAILABLE on
Kata.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Detect actual service and Runner readiness, restore LAN access after network changes, protect production web builds with rollback, honor cancellation during executor startup, and show truthful cached-data and credential states.
Add authenticated LAN access, native Claude execution, checkpoint and budget recovery, and signed POC build/deploy/rollback contracts. Include local service management, regression coverage, and evidence from the completed POC delivery.
Adding an executor touched the contracts, the API's availability check, the
worker and the console separately. Now it is an adapter plus one entry in
@wishforge/contracts' providerCatalog (name, executable variable, whether the
model proxy isolates its credentials) and one in the worker's
providerRegistry; the API, the console's options and MCP's submit_wish read
the catalog, and a test checks every entry is registered with the worker.
The CodeBuddy adapter already parsed Claude-Code-style stream-json, so it is
now a StreamJsonCliAdapter with a dialect per CLI. Claude Code runs headless
with --verbose and --permission-mode bypassPermissions inside the per-run
sandbox and receives only ANTHROPIC_* credentials (WF_CLAUDE_BIN). Like
CodeBuddy it is development-only until the model proxy supports it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In Kubernetes mode the per-run network policy allowed only DNS, the API,
Temporal and Git, so dependency installs could not reach a registry and image
builds had no builder.
- WF_RUN_INSTALL_PROXY_URL points both sandboxes at an operator's egress proxy
(pods labelled egress-proxy); the per-run policy admits it and installation
commands get HTTP(S)_PROXY, also inside bubblewrap, while checks still have
no network.
- WF_VERIFIER_DOCKER_HOST with a TLS client certificate points the verifier at
a remote builder daemon (pods labelled image-builder); a separate policy
lets only the verifier pod reach it, and release deletes that policy too.
Builds pass DOCKER_HOST, DOCKER_TLS_VERIFY and DOCKER_CERT_PATH, and refuse
a Docker context and a remote builder together.
- The control namespace's baseline admits sandbox traffic to both components.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A first implementation round may end, without changing any files, with a
QUESTIONS: section of up to five questions. The executor then puts the run in
the clarification wait with the questions in its detail, where the console and
MCP show them; answering already creates a new requirement revision that
starts again from planning. Repair rounds cannot ask: they must fix what the
verifier found.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
People's changes now reach the target branch the same way the agent's do. The
IDE credential carries a ref prefix and the Git service runs a pre-receive hook
on every push that refuses updates outside it, so the IDE can push only
`ide/` branches; the platform's own integration is unaffected.
- A wish may name a source branch. The run starts from that branch's head, the
agent builds on it, and verification, repair rounds, merging onto a moved
target and conflict repair work as before. A missing branch makes the run
wait (git_storage) until it is pushed.
- Release risk is now computed from the target branch before integration, not
from where the run started: a migration committed on the source branch would
otherwise have been missed and could have skipped approval.
- Clones work on `ide/work`; a plain `git push` publishes it.
- The console lists a project's IDE branches with a button to submit a wish
that merges one, and the new-wish form has a starting-point field; MCP
`submit_wish` accepts `sourceBranch`. The Git service lists branches under
a prefix for this (`POST /refs`).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Verification ran git inside the tree under test, kept no command output and
treated every infrastructure fault as a test failure; repairs went through an
executor lease the coordinator should not hold; and a project without delivery
could never succeed. This reworks the worker execution path and the control-plane
contracts it depends on.
- Verification exports the exact commit to a directory without .git, installs
dependencies (network allowed) and runs the protected commands under
bubblewrap in kata mode or directly in host mode; evidence records exit code,
duration, digest, a redacted output tail and the isolation used. Production
rejects host evidence. A failed installation waits as infrastructure.
- Waiting reasons are an enum with a redacted detail; dispatcher, GitHub and
model-proxy states map onto it, and the console explains each one with a way out.
- Repairs are requested by the coordinator through /internal/runs/:id/repair;
the control plane derives the failure summary from the verifier's evidence and
enforces the repair budget. Continuing after the budget is spent grants one more
round; a new requirement revision starts a fresh budget.
- Execution identities may only move a run within their own stages and only to
running or waiting; success comes from integration evidence (no delivery
configured) or deployment acceptance. The unused preview stage is gone.
- Integration fast-forwards when the target has not moved, merges and re-verifies
otherwise, retries when a push finds the target moved, and turns merge conflicts
into a repair round that merges the current target for the agent to resolve.
- Successful runs that needed a repair leave a project lesson; active lessons are
injected into later prompts, and admins can list, add, disable and delete them.
- Executor progress (messages, commands, file changes, plan updates) is stored as
normalized events; previously agent messages were recorded as the literal word
"output". Provider failures keep the work as a checkpoint and wait as
provider_error instead of failing the run.
- Workers declare WF_SANDBOX_MODE=host|kata (host refused in production),
heartbeat with retries until the lease could have expired, and retry safe
control-plane reads with backoff.
- tests/local-loop.test.ts runs the loop end to end with real Git, the in-process
control plane behind its internal routes and a scripted agent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>