agnes-the-ai-analyst/app/web/setup_instructions.py
Vojtech 41829e8a45
Setup-prompt + bootstrap fixes from 2026-05-10 init report (#240)
* Setup-prompt + bootstrap fixes from David's 2026-05-10 init report

Three issues from clean-machine bootstrap evidence:

1. `agnes refresh-marketplace --bootstrap` failed to recover when the
   local clone existed but Claude Code's marketplace registry had lost
   the `agnes` entry. Bootstrap path now parses
   `claude plugin marketplace list`, re-runs
   `claude plugin marketplace add ~/.agnes/marketplace` when missing,
   and treats `add` failures as fatal (was warn-and-continue, root cause
   of the cascade into "Marketplace 'agnes' not found" plugin install
   errors).

2. Setup prompt now always emits the marketplace-registration block,
   even when the operator has zero plugin grants. Pre-wires the
   SessionStart hook so future admin grants land automatically without
   re-running setup. Block copy adapts: empty list shows
   "no plugins granted yet", populated list shows "install plugins".

3. Setup prompt registers the Atlassian Remote MCP server unattended
   (`claude mcp add --transport sse atlassian
   https://mcp.atlassian.com/v1/sse`). Hosted Remote MCP, OAuth handled
   automatically by Claude Code on first use. Asana / GWS stay on the
   /home connector cards (PAT/keychain flows don't fit unattended
   bootstrap).

Confirm step nudges the user toward the /home connector cards for the
PAT-flow services. CLAUDE.md template renames the marketplace section
to "Agnes Marketplace" and documents that all plugins are addressed as
`<plugin>@agnes` regardless of upstream slug.

Layout: Confirm shifts from step 6/8 to step 9 across all variants
(preflight, marketplace, MCP all unconditional). Tests updated.

* Link Claude license options from /home install pane

Step-1 Claude install on /home pointed users to  OAuth without
explaining what to do if they don't have a Pro/Max subscription. Add
a one-line follow-up link to the plan-tier section on /setup-advanced
(new `#claude-plan` anchor) so first-time users discover the
subscription tiers rather than bouncing on the OAuth screen.

* Add idempotent + no-TLS-bypass guardrails to /home connector prompts

The Asana / Google Workspace / Atlassian connector prompts on /home
already shipped a precheck step that short-circuits when the service
is already wired, but they didn't carry the same idempotency +
surface-errors-verbatim + don't-disable-TLS-verification guardrails
the bash bootstrap prompt has. Add a one-paragraph 'Ground rules'
block at the top of each prompt so a connector failure doesn't
tempt the model into bypass workarounds, matching the same posture
David's 2026-05-10 init report flagged for the bash flow.

* skip Source: lines in marketplace registry detector

`claude plugin marketplace list` prints a `Source: <local path>` line
under each registered marketplace; the local clone almost always lives
under a path containing the marketplace name itself
(`~/.agnes/marketplace`). A naive \\bagnes\\b match over the full
stdout therefore false-positives whenever ANY unrelated marketplace
sits under `~/.agnes-…/` or similar. Filter Source: lines out before
matching so the recovery path actually re-adds when needed instead of
silently falling through to a broken `marketplace update agnes`.
Adds regression test covering the substring-only case.

* drop customer-specific tokens from CHANGELOG entries

Per CLAUDE.md vendor-agnostic OSS rule ("nothing customer-specific
... in changelogs"):
- "agnes-vrysanek.groupondev.com" -> "a private-CA Agnes deployment"
- "Groupon Marketplace / groupon-marketplace" -> "<Org> Marketplace /
  <org>-marketplace" (placeholder example)
- Removed "David flagged" attribution language; init-report context
  stays intact, just stripped of the named host + brand

---------

Co-authored-by: ZdenekSrotyr <zdenek.srotyr@keboola.com>
2026-05-10 20:24:00 +02:00

809 lines
40 KiB
Python

"""Single source of truth for the "Setup a new Claude Code" clipboard payload.
Both the JS-embedded clipboard renderer (`_claude_setup_instructions.jinja`)
and the read-only HTML preview on the dashboard and /install pages consume
these lines. Keep it in Python so there is exactly ONE place that edits.
Placeholders `{server_url}`, `{token}`, `{wheel_filename}`, and `{server_host}`
are substituted at render time. `{wheel_filename}` and `{server_host}` are
pre-substituted server-side via `resolve_lines()`; `{server_url}` and
`{token}` survive into the JS template and are filled in at click time.
`{wheel_filename}` is server-pre-substituted because `uv tool install`
validates the PEP 427 filename *in the URL path* before fetching, so a
stable alias like `agnes.whl` fails with "Must have a version" — we need
the real versioned filename inlined.
`{server_host}` is server-pre-substituted because the `git config` and
`claude plugin marketplace add` lines need the bare host (no scheme), and
the click-time JS only knows the full origin (`{server_url}`).
## Cross-platform trust strategy (when `ca_pem` is supplied)
The trust block (step 0) is the load-bearing piece. Three things bit us in
practice and the design here exists to dodge each one:
1. **rustls rejects the Agnes leaf cert as `CaUsedAsEndEntity`.** The Agnes
server's self-signed cert is simultaneously its own CA (basicConstraints
`CA:TRUE`) AND the leaf served on the wire — a setup OpenSSL tolerates
but webpki/rustls strictly refuses. So `uv tool install <https-url>`
never works against the Agnes wheel endpoint. We download the wheel via
curl first (curl uses OpenSSL, accepts the cert), then `uv tool install
--native-tls --force <local-file>` lets rustls reuse the OS trust store
for PyPI dependency resolution. No HTTPS hop through rustls touches the
Agnes host.
2. **`SSL_CERT_FILE` REPLACES the trust store, it doesn't append.** Pointing
it at `~/.agnes/ca.pem` alone breaks every Python tool that needs to
reach a public host (PyPI, GitHub) — `da` works fine because it only
talks to Agnes, but `uv run --with <pkg>` immediately fails with
`UnknownIssuer`. We materialize a combined bundle at
`~/.agnes/ca-bundle.pem` (system roots + Agnes CA) and point all
`SSL_CERT_FILE` / `REQUESTS_CA_BUNDLE` / `GIT_SSL_CAINFO` at it.
`NODE_EXTRA_CA_CERTS` keeps pointing at just `ca.pem` because Node's
semantics is *additive* (appends to bundled roots), so a single-cert
file is correct there.
3. **Bun-compiled `claude` (Windows + macOS distributions) ignores every
CA env var AND the OS trust store for marketplace HTTPS.** On macOS
arm64 the binary at `~/.local/bin/claude` is a Mach-O with a `__BUN`
segment (single-file `bun build --compile`); on Windows claude.exe is
the same shape. `strings` shows the binary recognizes
`NODE_EXTRA_CA_CERTS`, `SSL_CERT_FILE`, `REQUESTS_CA_BUNDLE`,
`CURL_CA_BUNDLE` (including a "NODE_EXTRA_CA_CERTS detected" log
string), but in practice the values never reach the TLS context — a
known limitation of Bun's compiled-binary HTTPS path. Registering the
cert in the OS trust store (Windows: `certutil -user -addstore Root`;
macOS: `security add-trusted-cert`; Linux: `update-ca-certificates` /
`update-ca-trust`) doesn't fix it on Windows or macOS either — the
binary's bundled CA list isn't refreshable from the OS store.
So the marketplace step always uses system `git clone` regardless of
platform — system git honors `GIT_SSL_CAINFO` from the combined bundle
in step 0(d). We tried having Linux attempt direct HTTPS first (where
node-based claude DOES respect `NODE_EXTRA_CA_CERTS`), but `claude
plugin marketplace add <https-url>` is broken end-to-end on every
distribution: it does succeed at downloading the marketplace.json, but
stores it as a single file. The plugin entries' `source: "./plugins/<name>"`
paths are then resolved as local filesystem paths against that file's
parent dir — and the plugin tree obviously isn't there. Only the clone
path produces a real directory tree that `plugin install` can read.
The OS trust-store registration in (c) is still done on all three
platforms because it's needed for *non-claude* native tools — e.g.
the system git fetch path itself (Schannel on Windows, Security
framework on macOS) trusts via the OS store, not via env vars.
Marketplace refresh: after the initial clone, `agnes refresh-marketplace`
incrementally `git pull`s against the same clone and runs `claude plugin
marketplace update agnes`. Credentials are injected per-pull via a
one-shot git credential helper (PAT from `~/.config/agnes/token.json`)
so the cloned repo's `origin` URL stays PAT-free at rest. The
SessionStart hook (installed by `agnes init`) calls refresh-marketplace
on every Claude Code session so changes server-side propagate
automatically.
## Step ordering
The numbered steps are arranged so that:
- All installation work (CLI, plugins) happens first, in one go.
- `agnes init` is mandatory — it bundles auth, workspace bootstrap,
CLAUDE.md fetch, and Claude Code SessionStart/End hooks into one
non-interactive call. Replaces the old `agnes auth import-token` +
`agnes auth whoami` pair.
- The interactive question (skills copy vs on-demand) is the LAST step
before Confirm — by that point everything else is done, the user only
needs to decide one thing, and the assistant blocks on their answer.
- `agnes diagnose` runs late so it doubles as a final smoke test after
plugins are in place, instead of gating them.
Layout (with marketplace plugins to install):
0 TLS trust block (only when ca_pem is supplied)
1 Install CLI
2 agnes init (auth + workspace bootstrap)
3 agnes catalog (smoke verify)
4 Pre-flight: git + claude
5 Marketplace + plugins
6 Diagnose
7 Skills (interactive — assistant waits for user)
8 Confirm
Layout (no plugins): steps 4-5 collapse out, diagnose/skills/confirm
renumber to 4-5-6.
The combined-bundle source uses a fallback chain so the prompt still works
on machines without the system Python `certifi`: we try (a) `python3 -c
'import certifi'`, (b) the platform's curl/openssl bundle path, (c)
`uv run --with certifi` as a network last-resort. The user explicitly
permitted that fallback chain — it's not improvising-around-a-TLS-error.
"""
from __future__ import annotations
# Marketplace name as published by app.marketplace_server.packager.
# Hard-coded here (rather than imported) to keep this module dependency-free
# and trivially testable. If the value ever drifts, the regression test
# below catches it.
_MARKETPLACE_NAME = "agnes"
def _tls_trust_block(ca_pem: str) -> list[str]:
"""Step 0 — cross-platform TLS trust bootstrap for the Agnes server.
Emitted only when the server has a non-publicly-trusted cert. Does four
things in a single numbered block (see module docstring for the full
rationale):
(a) Detect platform (Windows Git Bash / macOS / Linux) and pick the
shell rc file that the user's login shell actually reads.
`$SHELL`-driven, NOT existence-of-rc-driven — old setups put a
legacy `.bashrc` next to a default zsh shell on macOS, and the
`[ -f .bashrc ]` heuristic silently writes to the wrong file.
(b) Write the cert PEM to `~/.agnes/ca.pem` via single-quoted heredoc
(so `$` / backtick chars in real-world certs never shell-expand).
(c) Register the cert in the OS trust store (so native binaries that
bypass our env vars — claude.exe, system git's Schannel backend,
Python apps using `truststore` — still trust the host).
Idempotent: re-running just re-affirms the entry.
(d) Build a *combined* CA bundle (system roots + Agnes CA) at
`~/.agnes/ca-bundle.pem`, with a fallback chain for the system
roots source. Persist `SSL_CERT_FILE` / `REQUESTS_CA_BUNDLE` /
`GIT_SSL_CAINFO` pointing at the bundle, plus
`NODE_EXTRA_CA_CERTS` pointing at just `ca.pem` (Node
appends-not-replaces). Persistence is idempotent via a grep
guard for the `AGNES_CA_PEM_TRUST` marker.
"""
pem = ca_pem.strip()
lines: list[str] = [
"0) Trust the Agnes TLS certificate — cross-platform setup for a self-signed / private-CA host.",
"",
" (a) Detect platform + pick the shell rc file your login shell actually reads.",
" Driven by $SHELL + uname (NOT by which rc files happen to exist on disk).",
"",
" case \"$(uname -s)\" in",
" Darwin) PLATFORM=macos ;;",
" Linux) PLATFORM=linux ;;",
" MINGW*|MSYS*|CYGWIN*) PLATFORM=windows ;;",
" *) echo \"Unsupported OS: $(uname -s)\" >&2; exit 1 ;;",
" esac",
" SHELL_NAME=\"$(basename \"${SHELL:-bash}\")\"",
" case \"${SHELL_NAME}:${PLATFORM}\" in",
" zsh:*) RC=\"$HOME/.zshrc\" ;;",
" bash:macos) RC=\"$HOME/.bash_profile\" ;;",
" bash:windows|bash:linux) RC=\"$HOME/.bashrc\" ;;",
" *) RC=\"$HOME/.profile\" ;;",
" esac",
" echo \"Platform: $PLATFORM, shell: $SHELL_NAME, rc: $RC\"",
"",
" (b) Write the cert (single-quoted heredoc so $/backticks in the body don't expand):",
"",
" mkdir -p ~/.agnes",
" cat > ~/.agnes/ca.pem <<'AGNES_CA_PEM'",
]
# PEM body is flush-left: `<<'DELIM'` heredocs preserve leading whitespace,
# and any indent inside the cert breaks `openssl x509` / Python ssl parsers.
lines.extend(pem.splitlines())
lines.extend([
"AGNES_CA_PEM",
"",
" (c) Register the cert in the OS trust store. Native binaries (claude.exe,",
" system git's Schannel/Security.framework backends) read the OS store",
" and ignore our env vars — without this, the later marketplace `git",
" clone` (when plugins are configured) and any user-side git/native",
" tooling against the Agnes host will fail.",
" No admin rights needed (user-store only). Idempotent.",
"",
" case \"$PLATFORM\" in",
" windows)",
" WIN_CA=\"$(cygpath -w ~/.agnes/ca.pem)\"",
" certutil.exe -user -addstore \"Root\" \"$WIN_CA\"",
" ;;",
" macos)",
" # Will prompt once for the keychain password.",
" security add-trusted-cert -r trustRoot \\",
" -k \"$HOME/Library/Keychains/login.keychain-db\" \\",
" ~/.agnes/ca.pem",
" ;;",
" linux)",
" if command -v update-ca-certificates >/dev/null 2>&1; then",
" sudo cp ~/.agnes/ca.pem /usr/local/share/ca-certificates/agnes.crt",
" sudo update-ca-certificates",
" elif command -v update-ca-trust >/dev/null 2>&1; then",
" sudo cp ~/.agnes/ca.pem /etc/pki/ca-trust/source/anchors/agnes.crt",
" sudo update-ca-trust",
" else",
" echo \"WARN: install ~/.agnes/ca.pem into your distro's trust store manually\" >&2",
" fi",
" ;;",
" esac",
"",
" (d) Build a COMBINED CA bundle (system roots + Agnes CA) for Python tools",
" and curl. SSL_CERT_FILE *replaces* the trust store, so pointing it at",
" the Agnes CA alone would break public hosts (PyPI etc.). Source the",
" system roots from a fallback chain — the first source that produces",
" a non-empty, existing path wins. Don't abort on the first miss; that's",
" what the chain is for.",
"",
" CERTIFI_PATH=\"$(python3 -c 'import certifi; print(certifi.where())' 2>/dev/null || true)\"",
" [ -z \"$CERTIFI_PATH\" ] && CERTIFI_PATH=\"$(python -c 'import certifi; print(certifi.where())' 2>/dev/null || true)\"",
" if [ -z \"$CERTIFI_PATH\" ]; then",
" for p in /mingw64/ssl/certs/ca-bundle.crt /usr/ssl/certs/ca-bundle.crt \\",
" /etc/ssl/certs/ca-certificates.crt /etc/pki/tls/certs/ca-bundle.crt \\",
" /etc/ssl/cert.pem; do",
" [ -f \"$p\" ] && CERTIFI_PATH=\"$p\" && break",
" done",
" fi",
" if [ -z \"$CERTIFI_PATH\" ]; then",
" CERTIFI_PATH=\"$(uv run --native-tls --with certifi --no-project python -c 'import certifi; print(certifi.where())' 2>/dev/null || true)\"",
" fi",
" if [ -z \"$CERTIFI_PATH\" ] || [ ! -f \"$CERTIFI_PATH\" ]; then",
" echo \"ERROR: locate a system CA bundle. Install Python 3 + certifi and re-run.\" >&2",
" exit 1",
" fi",
" echo \"Base CA bundle: $CERTIFI_PATH\"",
" cat \"$CERTIFI_PATH\" ~/.agnes/ca.pem > ~/.agnes/ca-bundle.pem",
"",
" (e) Persist env vars in the rc file picked in (a). Idempotent — won't",
" duplicate on re-run thanks to the AGNES_CA_PEM_TRUST grep guard.",
" Note the asymmetry: SSL_CERT_FILE (and REQUESTS_CA_BUNDLE, GIT_SSL_CAINFO)",
" point at the COMBINED bundle because those tools REPLACE trust.",
" NODE_EXTRA_CA_CERTS points at just ca.pem because Node APPENDS to its",
" bundled roots.",
"",
" if ! grep -q 'AGNES_CA_PEM_TRUST' \"$RC\" 2>/dev/null; then",
" cat >> \"$RC\" <<'AGNES_RC_BLOCK'",
"# AGNES_CA_PEM_TRUST — added by Agnes setup",
"# Combined bundle (system roots + Agnes CA) for tools that REPLACE trust:",
"export SSL_CERT_FILE=\"$HOME/.agnes/ca-bundle.pem\"",
"export REQUESTS_CA_BUNDLE=\"$HOME/.agnes/ca-bundle.pem\"",
"export GIT_SSL_CAINFO=\"$HOME/.agnes/ca-bundle.pem\"",
"# Single-cert file for Node (APPENDS to bundled roots):",
"export NODE_EXTRA_CA_CERTS=\"$HOME/.agnes/ca.pem\"",
"export PATH=\"$HOME/.local/bin:$PATH\"",
"AGNES_RC_BLOCK",
" fi",
" # Apply for THIS shell too:",
" export SSL_CERT_FILE=\"$HOME/.agnes/ca-bundle.pem\"",
" export REQUESTS_CA_BUNDLE=\"$HOME/.agnes/ca-bundle.pem\"",
" export GIT_SSL_CAINFO=\"$HOME/.agnes/ca-bundle.pem\"",
" export NODE_EXTRA_CA_CERTS=\"$HOME/.agnes/ca.pem\"",
" export PATH=\"$HOME/.local/bin:$PATH\"",
"",
" IMPORTANT for the Bash tool: env vars do NOT persist between separate",
" Bash invocations. Re-export the four lines above (SSL_CERT_FILE,",
" REQUESTS_CA_BUNDLE, GIT_SSL_CAINFO, NODE_EXTRA_CA_CERTS) plus PATH at",
" the top of every later step's bash block that talks to Agnes.",
"",
])
return lines
def _install_cli_lines(*, has_ca: bool, server_url_placeholder: str = "{server_url}") -> list[str]:
"""Step 1 — install the `agnes` CLI.
When the trust block was emitted (`has_ca=True`), we MUST avoid
`uv tool install <https-url>` against the Agnes wheel endpoint:
rustls rejects the Agnes leaf cert with `CaUsedAsEndEntity`, regardless
of `--native-tls` (the rejection is at chain validation, not at trust
lookup — putting the cert in the OS store doesn't fix it). Solution:
download the wheel with `curl --cacert` (curl uses OpenSSL, no rustls),
then `uv tool install --native-tls` from the local file. PyPI deps
still resolve over HTTPS, but `--native-tls` makes uv use the OS trust
store for that path, which is fine because PyPI's CA chain is public.
When `has_ca=False`, we trust the server's cert is publicly valid, so
the simple direct install works.
"""
if has_ca:
return [
"1) Install the CLI.",
" The Agnes server's self-signed cert trips rustls' CaUsedAsEndEntity check,",
" so direct `uv tool install <https-url>` against the wheel endpoint fails",
" (even with --native-tls). Workaround: curl-then-local-install.",
"",
" If uv is missing first:",
" curl -LsSf https://astral.sh/uv/install.sh | sh",
" export PATH=\"$HOME/.local/bin:$PATH\"",
"",
" WHEEL=/tmp/{wheel_filename}",
f" curl -fsSL --cacert ~/.agnes/ca.pem -o \"$WHEEL\" {server_url_placeholder}/cli/wheel/{{wheel_filename}}",
" uv tool install --native-tls --force \"$WHEEL\"",
"",
" If `agnes --version` fails after install because ~/.local/bin is not on PATH:",
" export PATH=\"$HOME/.local/bin:$PATH\"",
" # persist: append the same line to your ~/.zshrc or ~/.bashrc",
" # (the trust block in step 0 already does this for you on first run).",
]
return [
"1) Install the CLI:",
f" uv tool install --force {server_url_placeholder}/cli/wheel/{{wheel_filename}}",
"",
" If uv is not installed yet:",
" curl -LsSf https://astral.sh/uv/install.sh | sh",
"",
" If `agnes --version` fails after install because ~/.local/bin is not on PATH:",
" export PATH=\"$HOME/.local/bin:$PATH\"",
" # persist: append the same line to your ~/.zshrc or ~/.bashrc",
]
def _init_lines(server_url_placeholder: str = "{server_url}") -> list[str]:
"""Steps 2-3 — `agnes init` (auth + workspace bootstrap) + smoke verify.
`agnes init` is the workspace-rails delivery mechanism for everyone:
it authenticates with the PAT, fetches CLAUDE.md (RBAC-filtered),
writes AGNES_WORKSPACE.md (human-facing docs), installs Claude Code
SessionStart/End hooks (auto-refresh), and runs an initial `agnes pull`
so DuckDB views are ready. Subsumes the legacy `agnes auth import-token`
+ `agnes auth whoami` pair — `init` already verifies the PAT against
`/api/catalog/tables` internally, and `agnes catalog` then doubles as
a smoke verify of the data plane.
The PAT minted by `/setup` is `general` scope with a 90 d TTL, so the
init call will succeed for the operator's whole 90 d window without
re-clicking "Generate prompt".
"""
return [
"",
"2) Bootstrap your Agnes workspace in this directory:",
f" agnes init --server-url \"{server_url_placeholder}\" --token \"{{token}}\" --workspace .",
"",
" This authenticates with the PAT, fetches your CLAUDE.md (RBAC-filtered),",
" writes AGNES_WORKSPACE.md (human-facing docs), installs Claude Code",
" SessionStart/End hooks (auto-refresh), and runs an initial `agnes pull`",
" so your DuckDB views are ready.",
"",
"3) Verify the data is queryable:",
" agnes catalog",
"",
" This should list the tables your account has grants for. Empty list",
" means your admin hasn't granted you access yet — contact them.",
]
def _diagnose_skills_lines(*, diagnose_num: str, skills_num: str) -> list[str]:
"""Diagnose + skills steps — moved AFTER the marketplace block.
Putting these last (instead of right after `whoami`) means: by the time
we ask the user the skills question, all installation work is finished —
the only thing the prompt is still waiting on is one human-loop answer.
`agnes diagnose` then doubles as a server-health smoke test that runs after
plugins are in place, not as a gate before them. With the new ordering
skills is the LAST step before Confirm, so the assistant must wait for
the user's answer before finalizing — there's no "run other steps in
parallel" affordance any more (and it isn't needed).
Step numbers are filled in by the caller because they shift between
the no-marketplace layout (4, 5) and the marketplace layout (6, 7).
"""
return [
"",
f"{diagnose_num}) Run diagnostics:",
" agnes diagnose",
"",
" This should print \"Overall: healthy\". `db_schema: unknown` and",
" `data: 0 tables` are NORMAL in two cases:",
" - fresh install (no tables registered yet), and",
" - non-admin roles (e.g. `analyst`) that don't have grants to read",
" the system schema even on populated instances.",
" Only flag actual yellow/red checks (api / duckdb_state / users).",
"",
f"{skills_num}) Skills (ask the user — this is the last interactive step before Confirm):",
" The CLI ships with reusable markdown skills (setup, connectors,",
" corporate-memory, deploy, notifications, security, troubleshoot),",
" listable via `agnes skills list` and readable via `agnes skills show <name>`.",
"",
" Ask the user verbatim: \"Do you want me to copy the Agnes skills into",
" ~/.claude/skills/agnes/ so they are always loaded in Claude Code,",
" or should I pull them on-demand via `agnes skills show <name>` when",
" needed?\"",
"",
" Wait for the user's answer before moving to Confirm.",
"",
" If they say copy:",
" mkdir -p ~/.claude/skills/agnes",
" for s in $(agnes skills list | awk '{print $1}'); do",
" agnes skills show \"$s\" > ~/.claude/skills/agnes/\"$s\".md",
" done",
" echo \"Copied skills to ~/.claude/skills/agnes/\"",
]
def _finale_lines(*, confirm_step_num: str, has_ca: bool) -> list[str]:
"""Final Confirm step. Bullets it asks the assistant to report on must
only reference earlier steps that were actually emitted, otherwise the
assistant either hallucinates an answer or asks the user about a
non-existent step. The CA-bundle-source bullet only makes sense when
the trust block ran (`has_ca`). The marketplace clone bullet is
unconditional now — preflight + marketplace are always emitted (Fix B
in the 2026-05-10 init-report response). Init + catalog + diagnose +
skills + version always render, so their bullets are unconditional."""
bullets = [
" - `agnes --version` output",
" - First few lines of `agnes catalog` (tables you can see)",
" - Confirmation that `./CLAUDE.md` and `./AGNES_WORKSPACE.md` exist",
" - Confirmation that `./.claude/settings.json` contains SessionStart/End hooks",
" - The `agnes diagnose` overall status",
" - Whether skills were copied or left on-demand",
" - Confirmation that `~/.agnes/marketplace/.git/` exists "
"(the marketplace clone) and that any granted plugins installed",
" - Reminder to scroll to the connector cards on /home and connect "
"Asana / Google Workspace / Atlassian (those run separately from this script)",
]
if has_ca:
bullets.append(
" - Which CA bundle source got picked in step 0(d) "
"(system Python certifi / system curl bundle / uv-fetched)"
)
return [
f"{confirm_step_num}) Confirm:",
" Tell me \"Agnes workspace is ready\" and summarize:",
*bullets,
]
def _preflight_block(step_num: str) -> list[str]:
"""Pre-flight check — runs before the marketplace clone.
`claude plugin marketplace add` (and our git-clone fallback) shells out
to `git`, AND the marketplace step calls `claude` itself, so a missing
binary on either side fails the step with a confusing error. We check
both here so the user gets a single clear "install X" message instead
of debugging a downstream error.
Cross-platform install commands cover the three supported workstation
OSes:
- macOS: Homebrew (`brew install git`). The Xcode CLT bundle also
ships git; we prefer brew because it's non-interactive.
- Windows: winget (`winget install --id Git.Git -e ...`). Bundled
with Windows 10 1809+ and Windows 11; non-interactive with --silent.
- Linux: apt or dnf, depending on distro family.
For `claude` we point at the official platform installer docs rather
than vendoring an install one-liner — Anthropic ships per-platform
installers (npm on Linux, native binary on macOS/Windows) and the
canonical instructions live at https://docs.claude.com/claude-code.
`step_num` is parameterized because step ordering shifted between
layouts (the marketplace block now runs before diagnose/skills, so
preflight + marketplace are steps 4-5 instead of 6-7).
"""
return [
"",
f"{step_num}) Make sure git and claude are installed (required for the marketplace clone):",
" git --version",
" claude --version",
"",
" If `git --version` fails (\"command not found\" or similar), install git:",
" - macOS: brew install git",
" - Windows: winget install --id Git.Git -e --source winget --silent",
" - Linux: sudo apt-get install git OR sudo dnf install git",
"",
" If `claude --version` fails, install Claude Code:",
" - npm (Linux / WSL): npm i -g @anthropic-ai/claude-code",
" - macOS / Windows native installer: see https://docs.claude.com/claude-code",
"",
" Then re-run both `--version` checks to confirm before continuing.",
]
def _marketplace_block(
plugin_install_names: list[str],
step_num: str,
) -> list[str]:
"""Build the marketplace + plugin-install block.
`plugin_install_names` may be empty: registering the per-user
marketplace clone with Claude Code is useful even when the operator
has zero plugin grants, because it pre-wires the SessionStart hook
and the grant flow — admin grants land on the next Claude Code
session without re-running setup. The block copy adapts for the
empty case so the comment-bullet doesn't promise plugin installs
that won't happen.
`step_num` is parameterized because step ordering shifted between
layouts (this block now runs before diagnose/skills, so it's step 5
instead of the old step 7).
The whole block is one CLI invocation: ``agnes refresh-marketplace
--bootstrap``. The CLI handles clone + PAT-strip + chmod + register-
with-Claude + auto-install-from-manifest internally. This is what
used to be a 15-line shell sequence inline; pulling it into the CLI
bought:
1. **Claude Code permission gate friendliness.** The agent-driven
onboarding flow inside Claude Code denies ``rm -rf`` by default;
the inline script tripped on it. Wrapping the destructive prep
inside agnes lets the CLI's already-trusted permission grant
cover it (Python ``shutil.rmtree`` doesn't pattern-match the
shell ``rm -rf`` block).
2. **Idempotence without inline ``rm``.** Re-running the install
prompt over an existing clone now does fetch+reset under the
hood (no destructive cleanup needed). The prompt's "safe to
re-run" promise holds without forcing the operator to delete
anything by hand.
3. **One source of truth.** ``agnes refresh-marketplace`` is also
the SessionStart hook command, so install + refresh share the
same code path — version-aware reconcile, hook JSON output,
credential helper PAT injection, all consistent.
Why always clone (with the CLI doing it) instead of trying direct
HTTPS marketplace add first? ``claude plugin marketplace add
<https-url>`` does succeed against our ``/marketplace.git/`` endpoint
(returns 200 + JSON), but Claude Code stores the response as a
single-file marketplace and resolves plugin ``source:
"./plugins/<name>"`` paths as local filesystem refs — so the
subsequent ``claude plugin install`` looks for plugin trees at
``<marketplace-dir>/plugins/<name>/`` and 404s because the dir is a
file. Only the git-clone path produces a real directory tree with
plugin contents in place. Broken end-to-end on every Claude Code
distribution; cloning is the only reliable install path.
TLS handling for the in-binary ``git clone`` is fully covered by the
cross-platform trust block (step 0) when the server's cert needs
bootstrapping (`ca_pem` non-empty), and by the OS trust store when
the cert is publicly-trusted. There used to be a legacy fallback
here that emitted a host-scoped ``git config http.<host>.sslVerify
false`` line for the ``AGNES_DEBUG_AUTH`` path; that's gone — it
masked operator misconfigurations (a ``self_signed_tls=True``
instance without ``/data/state/certs/fullchain.pem`` on disk) and
its ``sslVerify=false`` shell command tripped Claude Code auto-mode
classifiers. Operators serving a self-signed or private-CA cert
must place the fullchain at ``AGNES_TLS_FULLCHAIN_PATH`` (default
``/data/state/certs/fullchain.pem``) so step 0 can read it via
``_read_agnes_ca_pem``.
"""
has_plugins = bool(plugin_install_names)
header = (
"Register the Agnes Claude Code marketplace and install plugins:"
if has_plugins
else "Register the Agnes Claude Code marketplace (no plugins granted yet):"
)
bullet_5 = (
" # 5. install every plugin listed in the served manifest"
if has_plugins
else " # 5. (no plugins to install — your account has zero grants)"
)
if has_plugins:
trailer = [
" These run non-interactively. After they finish, tell the user to /exit",
" and run `claude` again so the new plugins load. From then on, the",
" SessionStart hook keeps the marketplace clone in sync via",
" `agnes refresh-marketplace --quiet` on every Claude Code session.",
]
else:
trailer = [
" Your account has no plugin grants right now, but registering the",
" marketplace anyway pre-wires the SessionStart hook. When an admin",
" grants you a plugin later, `agnes refresh-marketplace` (run by the",
" hook on every Claude Code session) will install it automatically —",
" no need to re-run this setup script.",
]
return [
"",
f"{step_num}) {header}",
" # `agnes refresh-marketplace --bootstrap` does:",
" # 1. clone the per-user marketplace bare repo to ~/.agnes/marketplace",
" # 2. strip the PAT from the cloned origin URL (refreshes use a",
" # per-invocation git credential helper, not the URL)",
" # 3. best-effort chmod 700/600 on POSIX (no-op on Windows NTFS)",
" # 4. `claude plugin marketplace add ~/.agnes/marketplace`",
bullet_5,
" # Idempotent — re-runs over an existing clone do fetch+reset+reconcile",
" # via the same path the SessionStart hook uses.",
" agnes refresh-marketplace --bootstrap || {",
" echo \"ERROR: agnes refresh-marketplace --bootstrap failed\" >&2",
" exit 1",
" }",
"",
*trailer,
]
def _mcp_servers_block(step_num: str) -> list[str]:
"""Register the Atlassian Remote MCP unattended.
Why only Atlassian here:
- Atlassian publishes a hosted SSE MCP at
https://mcp.atlassian.com/v1/sse with OAuth handled by Claude Code
automatically on first tool call. No PAT/keychain dance, no
per-user setup beyond clicking through OAuth once when an
operator first asks Claude to read a Jira ticket. Safe to
register unattended in the bootstrap script.
- Asana and Google Workspace need PAT/keychain flows that don't
survive non-interactive bootstrap, so those stay on the /home
connector cards (operator-driven).
Idempotent across re-runs: `claude mcp add` returns non-zero when the
server name already exists, so we soft-fail with `|| true` and a
one-line note rather than tripping the `set -e` style operators
sometimes wrap the prompt in. Subsequent runs of the prompt are
no-ops.
Reference: 2026-05-10 init-report — David's `claude mcp list` showed
only the pre-existing claude.ai Drive connector; Atlassian/Asana/GWS
weren't registered because the prompt had zero `claude mcp add`
lines. Fix C in the response plan.
"""
return [
"",
f"{step_num}) Register the Atlassian MCP server (Jira + Confluence on demand):",
" # Hosted Remote MCP — Claude Code handles OAuth automatically the",
" # first time you ask it to read a Jira ticket or Confluence page.",
" # Idempotent: re-runs are a no-op (the `|| true` swallows the",
" # \"server already exists\" error from `claude mcp add`).",
" claude mcp add --transport sse atlassian https://mcp.atlassian.com/v1/sse || true",
"",
" Asana and Google Workspace use per-user PAT / CLI flows that don't",
" fit an unattended bootstrap — connect those from the /home connector",
" cards after this script finishes.",
]
def _preamble_lines(*, has_ca: bool) -> list[str]:
"""Header that opens the prompt before the numbered steps. The
`step 0(d) fallback chain` reference is only emitted when the trust
block actually exists (`has_ca`); without it the line points at a
non-existent step. The "don't disable TLS verification" advice itself
stays unconditional — it's good guidance regardless of whether the
server runs with a private CA."""
lines = [
"Set up the Agnes CLI on this machine.",
"",
"Server: {server_url}",
"Personal access token: {token}",
"(Just generated; treat it as a secret.)",
"",
"Run these, in order. The script is idempotent — safe to re-run if a step",
"fails partway through. If a step fails with an unfamiliar error, paste the",
"exact error back and stop. Do NOT improvise around TLS errors by disabling",
"verification (`-k`, `NODE_TLS_REJECT_UNAUTHORIZED=0`,",
"`git -c http.sslVerify=false`, etc.) — those are dead ends that hide the",
"real problem.",
]
if has_ca:
lines.append(
"The fallback chain inside step 0(d) is documented and OK to "
"use; that's what fallback chains are for."
)
lines.append("")
return lines
def _step_numbers(*, has_skills: bool = True) -> dict[str, str]:
"""Compute the step numbers for the unified layout.
Returns a dict keyed by logical step name; values are stringified
1-based step numbers (preserving the existing string-based helper API
so call sites stay diff-minimal).
Mandatory steps (always emitted): install (1), init (2), catalog (3),
preflight (4), marketplace (5), mcp_servers (6), diagnose (7), skills
(8), confirm (9). Preflight + marketplace + mcp_servers are all
always-on:
- Marketplace registration is useful even when the operator has
zero plugin grants (SessionStart hook reconciles future grants
automatically).
- Atlassian MCP registration is unattended-safe (hosted Remote MCP
with Claude Code-managed OAuth) and applies to every analyst
whose work touches Jira/Confluence — high enough hit rate to
justify default-on.
`has_skills` is kept as a parameter for future flexibility; default
True (the Resolved-Question section in the original plan settled on
always-on).
Step-0 (TLS trust block) sits outside this numbering — it is gated by
has_ca and has its own "0)" header rendered inside the trust block
helper.
"""
n = 4
preflight = str(n); n += 1
marketplace = str(n); n += 1
mcp_servers = str(n); n += 1
diagnose = str(n); n += 1
skills = str(n) if has_skills else ""
if has_skills:
n += 1
confirm = str(n)
return {
"preflight": preflight,
"marketplace": marketplace,
"mcp_servers": mcp_servers,
"diagnose": diagnose,
"skills": skills,
"confirm": confirm,
}
def resolve_lines(
wheel_filename: str,
*,
plugin_install_names: list[str] | None = None,
server_host: str = "",
ca_pem: str | None = None,
) -> list[str]:
"""Return the template lines with server-side placeholders substituted.
Pre-substitutes `{wheel_filename}` and `{server_host}`. Leaves
`{server_url}` and `{token}` as placeholders for click-time JS
substitution (or for `render_setup_instructions()` below).
When `plugin_install_names` is empty/None, the output matches the
six-step no-marketplace layout (Confirm = step 6). When non-empty, a
step-4 pre-flight + step-5 marketplace block are inserted and Confirm
becomes step 8.
`ca_pem` (PEM-encoded fullchain of the Agnes server's TLS cert) gates
the cross-platform step-0 trust-bootstrap block AND switches step 1 to
the curl-then-local-install pattern AND switches step 5 to the
platform-aware marketplace strategy. Caller decides whether the cert
needs the bootstrap (typically: skip for publicly-trusted certs like
Let's Encrypt, emit for self-signed or private corp CA).
Fallback: callers pass `"agnes.whl"` when no wheel is present on disk.
The resulting URL (`/cli/wheel/agnes.whl`) will 404 at download time, but
the instruction text still renders so operators can see the snippet shape
and diagnose the missing wheel on the server.
"""
names = list(plugin_install_names or [])
has_ca = bool(ca_pem and ca_pem.strip())
# Step layout. Preflight + marketplace go BEFORE diagnose/skills so the
# human-loop skills question is the last step before Confirm. Both are
# always emitted: the marketplace registration is useful even with zero
# plugin grants (the SessionStart hook reconciles future admin grants
# automatically without re-running setup). `_step_numbers` returns the
# renumbered step labels in one place — no branch on every helper — so
# the layout is unambiguous and trivially extendable.
steps = _step_numbers(has_skills=True)
lines: list[str] = []
if has_ca:
lines.extend(_tls_trust_block(ca_pem)) # type: ignore[arg-type]
lines.extend(_preamble_lines(has_ca=has_ca))
lines.extend(_install_cli_lines(has_ca=has_ca)) # 1
lines.extend(_init_lines()) # 2, 3
lines.extend(_preflight_block(steps["preflight"])) # 4
lines.extend(_marketplace_block(names, step_num=steps["marketplace"])) # 5
lines.extend(_mcp_servers_block(steps["mcp_servers"])) # 6
# Diagnose + skills come AFTER marketplace + MCP wiring.
lines.extend(_diagnose_skills_lines(
diagnose_num=steps["diagnose"], skills_num=steps["skills"],
))
lines.append("")
lines.extend(_finale_lines(
confirm_step_num=steps["confirm"],
has_ca=has_ca,
))
return [
line.replace("{wheel_filename}", wheel_filename).replace("{server_host}", server_host)
for line in lines
]
def render_setup_instructions(
server_url: str,
token: str,
wheel_filename: str = "agnes.whl",
*,
plugin_install_names: list[str] | None = None,
server_host: str = "",
ca_pem: str | None = None,
) -> str:
"""Render the setup instructions as a single string.
Used server-side for tests and any non-JS rendering path. The browser
clipboard flow uses the JS renderer embedded in the Jinja partial; both
must produce byte-identical output for a given (server_url, token,
wheel, plugins, host, ca_pem) tuple.
"""
lines = resolve_lines(
wheel_filename,
plugin_install_names=plugin_install_names,
server_host=server_host,
ca_pem=ca_pem,
)
text = "\n".join(lines)
return text.replace("{server_url}", server_url).replace("{token}", token)