Commit graph

6 commits

Author SHA1 Message Date
ff2b19456f Make the secret layout configuration, not a constant
ops-warden routes API-key needs to railiance-platform, whose convention is
platform/workloads/<domain>/<workload>/<bundle>. This repository invented
secret/fluid-telegram/<campaign>/telegram instead, which is not its call to
make -- a service that picks its own paths in someone else's store is how a
policy ends up written around a mistake.

Mount and prefix are now BAO_MOUNT and FLUID_BAO_PREFIX, with the old scheme
kept as a development fallback. The runbook points at `warden access` for the
current shape and at OIDC login rather than a plain token, and names the
check that tells whether a login actually took.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172sgCZEEDJcnQmr4SGDvKa

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1361245@bnt-lap001
Assistant-Session: b3b428ef-f3e6-4688-b091-01f71461d66a
2026-09-04 22:32:09 +02:00
de7a7bdd2a Read the standard token sink, and tell a 403 from an outage
NewFromEnv now falls back to ~/.vault-token, the file `bao login` writes and
the bao and vault CLIs already read. An operator who has logged in once
should not have to re-export a secret, and a token that never has to be
typed is a token that never lands in shell history.

Preflight distinguishes the two things a 403 means. An expired token and a
token missing a policy look identical in the error, and the check that
separates them -- whether `bao token lookup` also fails -- is worth naming
where it is read rather than left to be rediscovered.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172sgCZEEDJcnQmr4SGDvKa

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1361245@bnt-lap001
Assistant-Session: b3b428ef-f3e6-4688-b091-01f71461d66a
2026-09-04 22:27:01 +02:00
7347bd6302 Implement the avatar and a preflight dry run
The avatar is now applied rather than deferred: BotFather's /setuserpic is a
conversation in which you send a photo, so the file is uploaded and sent as
a message. It is content addressed -- replacing the file is what triggers an
update, and the digest is recorded only after BotFather confirms, so a failed
upload retries rather than being remembered as done. The image is validated
before the conversation starts, because an image rejected halfway leaves the
bot registered without a picture.

Adds `provision preflight`: spec, avatar, OpenBao reachability, credentials,
session presence, salt and the resulting plan, checked in one run that writes
nothing and never contacts Telegram. Every failure it reports is one that
would otherwise surface after a phone number had been spent.

Two bugs it found immediately. The avatar path is documented as repo-relative
but resolved against the spec's own directory, so the real campaign spec
failed to find its own asset. And the OpenBao error named both variables when
only one was missing, sending the reader to check the one already set.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172sgCZEEDJcnQmr4SGDvKa

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1361245@bnt-lap001
Assistant-Session: b3b428ef-f3e6-4688-b091-01f71461d66a
2026-09-04 22:12:56 +02:00
014ee5c746 Make the plan honest and the runbook's paths real
Two things that would have gone wrong on the first live run.

The plan promised "create bot.avatar" and apply had no avatar code at all:
/setuserpic needs a photo upload that is not implemented. A plan that
promises an action apply silently skips is worse than one that admits the
gap, because it makes every other line less trustworthy. The avatar is now
deferred, with instructions for setting it by hand.

The runbook wrote secrets to "<interface-path>/telegram/...", which is not
what the code computes. An operator following it would have put api_id and
api_hash somewhere the tool never looks, and found out at the first
connection attempt. Both documents now carry the real paths --
<mount>/fluid-telegram/<campaign>/telegram/<key> -- with a copy-pasteable
bao kv put for the one secret written by hand, and a verify step before
the session is minted.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172sgCZEEDJcnQmr4SGDvKa

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1361245@bnt-lap001
Assistant-Session: b3b428ef-f3e6-4688-b091-01f71461d66a
2026-09-04 22:02:56 +02:00
5ddfee8250 Implement the MTProto client, session bootstrap and apply
Completes the provisioner's write path. internal/tg drives BotFather as a
conversation rather than pretending it is an endpoint, creates channels,
claims usernames with fallbacks, and grants post_messages. internal/apply
sequences it: bot before administrator, test channel before public, and
state saved after every step that changed the world -- a channel that
exists but is unrecorded is worse than one that does not exist, because
the next run creates a second.

The operator session lives in OpenBao, not on disk. gotd's FileStorage
would leave a full-account credential in the working directory, where it
outlives the run and can be committed by accident.

The bot token goes straight from BotFather's reply to OpenBao and is
cleared from memory; if that write fails the error says how to recover by
hand and warns against re-running, since a retry creates a second bot.

Closes T05: the redaction salt is create-if-absent with no overwrite path,
and the test asserts it, because rotating it invalidates every longitudinal
comparison with no visible failure.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172sgCZEEDJcnQmr4SGDvKa

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1361245@bnt-lap001
Assistant-Session: b3b428ef-f3e6-4688-b091-01f71461d66a
2026-09-04 21:39:30 +02:00
10018a99b6 Implement provision plan: spec, resolved state, and the diff
Go, per the toolchain decision. cmd/provision with internal/spec,
internal/state and internal/plan; plan computation is pure and takes live
observation through an interface, so the refusal logic is testable without
a Telegram account. It runs against the real campaign spec today.

Separates two refusals the design had treated as one. A deferral is the
design working -- the public channel waiting on a checked rendering, normal
on every first run. A block is the world disagreeing with the state file:
drifted rights, a taken-over username, a bot that is no longer reachable.
Collapsed together, a first run could never apply anything, because it
always defers the public channel.

The rights clamp and the private/public username rule are enforced in Go
and asserted against the same cases the JSON schema rejects, since mirroring
the schema in code is a drift risk worth a test rather than a comment.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172sgCZEEDJcnQmr4SGDvKa

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1361245@bnt-lap001
Assistant-Session: b3b428ef-f3e6-4688-b091-01f71461d66a
2026-09-04 21:25:49 +02:00