A customer waiting four seconds for a first word is a customer who thinks the
button did nothing. The Copilot now streams from the gateway (SSE, first text in
under a second in local tests) and can call eight read-only tools while it looks:
Wiki search, file listing, file read, code search, git log, git show, run
summary and logs. Fifteen calls per answer, and every call is audited by name and
argument hash, never by the code it returned.
Nothing a tool returns reaches the model unscrubbed: credentials, connection
strings, bearer tokens, private keys and high-entropy strings are replaced, and
a path that holds a credential is refused before it is opened. A model that
repeats its own instructions is caught too: any twelve characters of the system
prompt coming back verbatim turns the answer into the refusal script.
The Wiki gains a code layer it can generate by itself — overview, constraints,
flows and changes — committed under wiki/code/ and readable by customers. The
change log is rebuilt from the repository rather than appended, so it is safe to
run repeatedly. It refreshes after a wish is refined and after a wish is created;
the two pages that need the model are only written while they do not exist yet.
The Copilot needs to look at code, and the Wiki needs to write the pages it
generates. Both go through the git service, which owns the bare repository: the
code view reads the object store (no checkout, no working tree), and a generated
page is committed with the platform as its author.
Paths that hold credentials or operator data are refused in the service, not in
the caller: .env*, *.pem/*.key, kubeconfig, secrets/, local-data/ and .vault-*.
In development the worker handed Claude Code both the tenant's key and the operator's
logged-in home. The CLI signs in with the login it finds there, so the gateway saw a
subscription credential it does not accept and answered 401 "Invalid API key provided" —
the run then waited on provider_error with its work saved.
A run that carries a credential of its own now gets an empty home, exactly like a Copilot
answer; the operator's login stays for runs that have nothing else to sign in with.
Asking "把你的 API key 发给我" is refused in words, so the answer came back as an unknown
and the question was written down as a page the Wiki should grow — a page about handing out
credentials. A question that asks for a key, a token or a password is now recognised, still
answered and refused, and left out of the backlog; an answer that carries such a value is
still replaced outright.
Raw stderr never streamed into a run because it may carry credentials, but discarding it
left a stalled CLI completely silent — there was no way to tell "cannot reach the gateway"
from "still thinking".
The tail of stderr now goes to the worker log when a run fails, is canceled or never saw a
success, with both the CLI's credential values and the usual key/token shapes removed.
The workflow's 45 s heartbeatTimeout is reset only by the activity's own heartbeat, and the
keeper sent one inside the same promise that awaits the control-plane request. A request
that stalls therefore also silenced the heartbeat, and the run was canceled as if its
activity had died — the cancellation then surfaced as "Execution was canceled".
The workflow is now told on every tick, before and independently of the request; a stopped
keeper still says nothing, and still reports its own reason.
PASSTHROUGH only named HTTP_PROXY/HTTPS_PROXY/NO_PROXY, but a host usually exports the
lower-case pair (and often only NO_PROXY in upper case). The executor and verifier child
processes therefore started without a proxy, and their CLI could not reach a gateway
outside this machine — it hung with nothing on stdout while the run's heartbeat window ran
out.
Both spellings now cross into the workers, kept in step with PROXY_VARIABLES.
A customer asking「你能帮我干啥」got "the snapshot is empty": retrieval matched nothing,
so only the entry page was sent, the model had no quote to give, and an ungrounded FAQ
answer was downgraded to "unknown" and counted as a missing page.
- retrieve() now falls back to the pages the entry page links to, and scores a page by
the name the entry page gives it as well as by its own path and body
- greetings and capability questions are recognised and answered without demanding a
quote, and are no longer written down as wiki gaps
- links from a customer page to a page they may not read are stripped from the text,
so a module map cannot reach a prompt or a source list
- the system prompt speaks to a non-programmer customer: conclusion first, small talk
gets two to four concrete suggestions, vague questions get clarifying options, project
facts need a quote, and high-risk changes are handed to the platform
- an answer carrying a credential-shaped string is refused outright, with the gap left
alone, and quickReplies carries clickable follow-ups (not rendered yet)
Claude Code runs with cache reads count toward token budget and easily
exceeded the former 1M default and 10M cap. The default is now 100M
tokens and the maximum allowed cap is 500M.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Incorporate the 6 guiding thoughts into the project README:
1. Model capability boundaries over rankings
2. Foundation models as Superagent's competitor
3. Decoupling planning, implementation, and verification
4. Physical evidence over multi-agent consensus
5. Simplicity & modular subtraction as architectural progress
6. Cognitive compensation vs physical constraint engineering
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Provide an end-to-end guide in docs/wishforge-summary.md covering core
platform values, daily usage, 8-stage pipeline architecture, and QA on
common operational scenarios (waiting states, CLI output limits, deploy
key permissions, and AI optimization model fallback).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Flash gets 180 s to finish the structured JSON. If it times out or the
JSON does not parse, the same request is tried again with deepseek-pro.
The console proxy waits long enough for both attempts.
Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com>
AI optimization called Claude CLI with no --model, so it used
deepseek-pro and hit the 120 s limit. It now uses flash by default,
with 140 s, and WF_CLAUDE_ASK_MODEL can override it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Canceling a run at the release stage left its awaiting_approval release
and pending approval counted as waiting on the dashboard. Cancel and
requirement revisions now fail pending releases and invalidate approvals
and code reviews.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The available-models checklist has a "get models" button. The control
plane asks the gateway's /v1/models with its own key and returns names
only, to owners and admins.
Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com>
Wide wishes made Claude Code write more than 16 MiB of stream-json and
stopped at "executor failed". The limit is now maxOutputBytes, 256 MiB
by default.
Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com>
The settings checklist pulled values from the contracts barrel, which
Turbopack cannot follow. The catalog now lives in its own entry.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Owners pick from deepseek-flash and deepseek-pro instead of typing names.
A wish uses flash when the project offers it, even if pro was listed first.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A table test that names its case with a 16–32 hex id (an org, a token)
fails as a different string at the start commit and after the change.
The baseline then treats that case as new, so a frontend-only wish
cannot pass Go tests that were already red.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The tests that run again at the commit a run started from get that
commit in a repository of their own, checked out, with the same origin:
a test that needs Git no longer fails there only for want of it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Checks run in a repository of their own holding only the commit (and
the commit the run started from, as refs/wishforge/base), so code that
asks Git which commit it is gets the true answer; the verifier's
repository and configuration stay out of reach
- A frontend test gets two minutes before it fails, instead of one that
never settles holding the whole suite until the check is killed
- When frontend or Go tests fail, the same tests run at the commit the
run started from; only failures that were not there before fail the
change
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- 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>