spike(ai): a durable Symfony AI agent, driven from workflow code

[gplanchat]

What this is

A spike, confined to symfony/ (the sample application): Symfony AI's agent loop run from
workflow code, so that a conversation is a workflow execution, a message is a signal, and the
agent survives a restart between two tool calls. No package, no bundle, no ADR.

The agent itself is not modified. A regular Provider is composed with two implementations of
ours: ModelClientInterface becomes an await on an activity (the only HTTP call in the whole
agent), ToolExecutorInterface becomes an await per tool call, preceded by a guard. Runner,
which is @internal, is never touched. symfony/ai is pinned to v0.13.0, the latest release,
because 0.x makes no compatibility promise between minors.

What the demo shows

• A guard on tool calls. Each tool carries an effect (read, write, external); the mode
  (standard, edition, auto) decides what passes without asking. Anything else suspends the
  workflow until a tool_decision signal, with a deadline.
• Questions to the human, as a questionnaire the workflow waits on.
• Watch and alert. The agent can go to sleep on a business fact and wake up knowing what it was
  doing; app:agent:evenement raises such a fact from the CLI.
• A context budget on the model call, and a cold resume that starts a new execution from a
  summary of the previous one. ?contexte=N makes compaction observable by hand.
• Delegation to a sub-agent, which never gets more authority than its parent.
• Reasoning blocks cross the boundary and show in the thread, folded.
• The Mistral bridge, taking only its pure parts (normalizers, result converter) — the HTTP
  client is replaced. Without MISTRAL_API_KEY, a scripted client answers: the demo runs with no
  key and no network.
• Mercure for what arrives from outside the page; the page also polls, so the hub is optional.

How to run it

documentation/user/use-cases/durable-agent.md (and its French mirror) is back — 761ea76 had
removed it for lack of a launch path — with a "What the demo shows" section and a launch path
that needs PHP 8.2 with ext-grpc, Composer and Docker, nothing else:

cd symfony
composer install
docker compose up -d --wait
php bin/console messenger:consume durable_temporal_journal    # terminal 1
php bin/console messenger:consume durable_temporal_activity   # terminal 2
php -S localhost:8012 -t public                               # terminal 3

Then http://localhost:8012/durable/chat. This path was run end to end, including a from-scratch
composer install. For it to work on any machine, the last commits make the per-facet hosts
(samples.durable.localhost, agent.durable.localhost) an opt-in through .env.local; empty by
default, they put no host constraint on the routes. Mercure's CORS is opened to any origin, since
the page's port is whoever launches it's choice.

The use-cases section itself is not on this branch (it is 373 commits behind main); the table
row in _index.md / _index.fr.md goes back at merge time, and 761ea76 shows which lines.

To be aware of

• symfony/composer.json and composer.lock change: symfony/ai-agent,
  symfony/ai-mistral-platform (both v0.13.0), symfony/mercure-bundle, symfony/lock.
• The branch carries fix(temporal): do not cancel an already-cancelled timer, with
  TimerCancelNotRepeatedTest, in src/Bridge/Temporal. It was found by the demo.
• git merge-tree against main reports no conflicts.
• With symfony serve, the Docker stack is one of the workers in .symfony.local.yaml: it lives
  and dies with the server, and a worker that loses Temporal exits rather than retrying.
• Not proven, as the page says: nothing in the suite exercises the Mistral path; no scripted
  cross-process crash; the reasoning signature is not carried by any ai-platform normalizer.