coder

opensource/coder

Fork 0

mirror of https://github.com/coder/coder.git synced 2026-10-16 14:48:56 +08:00

Commit Graph

Author	SHA1	Message	Date
Thomas Kosiewski	9aa7130e01	feat: evaluate experiment rules per user in discovery and chatd (#29964 ) Slice 3 of 4 for runtime experiment rules: evaluate stored rules per user in discovery, the frontend experiments query, and the pilot `mcp-tool-search` gate in chatd. There is still no way to edit rules through the product; that comes in slice 4. With no rules stored and a healthy database, backend decisions match today. This PR deliberately changes three things: 1. Fail closed on rule reads. If the rules read fails, `mcp-tool-search` is off, even when it is enabled at startup. 2. Frontend freshness. The experiments query now refetches. 3. Caching. `/api/v2/experiments` and rendered HTML are sent with `Cache-Control: no-store`. Static assets keep their long-lived caching. ## What's in it - coderd builds `api.ExperimentEvaluator` from `experiments.NewDBStore` plus the startup `--experiments` list. - Discovery: `GET /api/v2/experiments` and the `<meta property="experiments">` tag both use `EnabledExperiments(ctx, caller)`. - chatd: `mcp-tool-search` is now user-scoped. `chatd.Config.ExperimentEvaluator` is required, so `chatd.New` returns an error without it. `prepareGeneration` decides once per turn for the chat owner (`chat.OwnerID`); child chats inherit the owner. - Ruleguard (`userScopedExperimentStaticCheck` in `scripts/rules.go`): rejects static `Experiments.Enabled(...)` and `httpmw.RequireExperiment(...)` checks of user-scoped experiments in server packages. A codersdk test keeps its list equal to `codersdk.ExperimentsUserScoped`. - Frontend:* the experiments query is keyed by user ID and seeded from the meta tag only when the embedded user matches. It is stale 60 s after page load and refetches on mount, window focus and reconnect. Logout also clears the experiments metadata. The dashboard keeps the last loaded list when a background refetch fails, instead of replacing the page with an error. - Once per turn: chatd decides `mcp-tool-search` on a turn's first step and reuses that decision for the turn's later steps, so a rule change applies from the next turn and never withdraws a `find_tools` call the model already made. The decision lives on the chat runner, keyed by the turn's prompt row, like the stop-nudge tracker. A runner handoff in the middle of a turn evaluates again. - Reference audit: the only non-test server check of `mcp-tool-search` is in `generation_preparer.go`, and it now goes through the evaluator. ## Testing - Go: per-user results and `no-store` on `/api/v2/experiments`; the per-user meta tag and the HTML cache headers in `site`; group and org conditions end to end in enterprise; chatd owner eligibility, on → off → on across turns including a child chat, and a failed rules read. - Vitest: user-keyed seeding, staleness from page load after a remount, refetch on focus, logout clearing, and the dashboard surviving a failed refetch. Remote dogfood UAT through Coder Agents passed on a licensed test deployment: - per-user discovery in both the API and the meta tag; - widening a rule, and turning it off, with no restart; - `off` beating a startup default, and `inherit` restoring it; - fail closed on malformed JSON and on an unknown mode, with no condition text in the logs; - the `no-store` headers; - real-model chatd enforcement: `find_tools` offered on turn 1, gone after `off`, back on turn 3. That round tested `b57207d4ee`. The later commits change only tests, and a follow-up round will cover the final head. For the author: `coderd/x/chatd/ARCHITECTURE.md`'s list of the runner's local state doesn't mention the per-turn trackers (debug turn, session start, stop nudges, and now this experiment decision). I did not edit that document. Known behavior: a broken rule logs one error line per evaluation (per request or turn), with only the experiment, revision, category and position. ## Stack 1. #29869: evaluator library 2. #29898: storage and audit types 3. This PR: enforcement and discovery 4. Control plane (experimental API, `coder exp experiment-rules` CLI, docs) > [!NOTE] > This PR was written by an AI agent (Xum). --- _Generated with [`xum`](https://github.com/coder/xum) • Model: `anthropic:claude-opus-5-5` • Thinking: `high`_ --------- Signed-off-by: Thomas Kosiewski <tk@coder.com>	2026-09-29 12:23:00 +00:00

Author

SHA1

Message

Date

Thomas Kosiewski

9aa7130e01

feat: evaluate experiment rules per user in discovery and chatd (#29964 )

Slice 3 of 4 for runtime experiment rules: evaluate stored rules per
user in discovery, the frontend experiments query, and the pilot
`mcp-tool-search` gate in chatd. There is still no way to edit rules
through the product; that comes in slice 4.

With no rules stored and a healthy database, backend decisions match
today. This PR deliberately changes three things:

1. **Fail closed on rule reads.** If the rules read fails,
`mcp-tool-search` is off, even when it is enabled at startup.
2. **Frontend freshness.** The experiments query now refetches.
3. **Caching.** `/api/v2/experiments` and rendered HTML are sent with
`Cache-Control: no-store`. Static assets keep their long-lived caching.

## What's in it

- **coderd** builds `api.ExperimentEvaluator` from
`experiments.NewDBStore` plus the startup `--experiments` list.
- **Discovery:** `GET /api/v2/experiments` and the `<meta
property="experiments">` tag both use `EnabledExperiments(ctx, caller)`.
- **chatd:** `mcp-tool-search` is now user-scoped.
`chatd.Config.ExperimentEvaluator` is required, so `chatd.New` returns
an error without it. `prepareGeneration` decides once per turn for the
chat owner (`chat.OwnerID`); child chats inherit the owner.
- **Ruleguard** (`userScopedExperimentStaticCheck` in
`scripts/rules.go`): rejects static `Experiments.Enabled(...)` and
`httpmw.RequireExperiment*(...)` checks of user-scoped experiments in
server packages. A codersdk test keeps its list equal to
`codersdk.ExperimentsUserScoped`.
- **Frontend:** the experiments query is keyed by user ID and seeded
from the meta tag only when the embedded user matches. It is stale 60 s
after page load and refetches on mount, window focus and reconnect.
Logout also clears the experiments metadata. The dashboard keeps the
last loaded list when a background refetch fails, instead of replacing
the page with an error.
- **Once per turn:** chatd decides `mcp-tool-search` on a turn's first
step and reuses that decision for the turn's later steps, so a rule
change applies from the next turn and never withdraws a `find_tools`
call the model already made. The decision lives on the chat runner,
keyed by the turn's prompt row, like the stop-nudge tracker. A runner
handoff in the middle of a turn evaluates again.
- **Reference audit:** the only non-test server check of
`mcp-tool-search` is in `generation_preparer.go`, and it now goes
through the evaluator.

## Testing

- **Go:** per-user results and `no-store` on `/api/v2/experiments`; the
per-user meta tag and the HTML cache headers in `site`; group and org
conditions end to end in enterprise; chatd owner eligibility, on → off →
on across turns including a child chat, and a failed rules read.
- **Vitest:** user-keyed seeding, staleness from page load after a
remount, refetch on focus, logout clearing, and the dashboard surviving
a failed refetch.

Remote dogfood UAT through Coder Agents passed on a licensed test
deployment:
- per-user discovery in both the API and the meta tag;
- widening a rule, and turning it off, with no restart;
- `off` beating a startup default, and `inherit` restoring it;
- fail closed on malformed JSON and on an unknown mode, with no
condition text in the logs;
- the `no-store` headers;
- real-model chatd enforcement: `find_tools` offered on turn 1, gone
after `off`, back on turn 3.

That round tested `b57207d4ee`. The later commits change only tests, and
a follow-up round will cover the final head.

**For the author:** `coderd/x/chatd/ARCHITECTURE.md`'s list of the
runner's local state doesn't mention the per-turn trackers (debug turn,
session start, stop nudges, and now this experiment decision). I did not
edit that document.

**Known behavior:** a broken rule logs one error line per evaluation
(per request or turn), with only the experiment, revision, category and
position.

## Stack

1. #29869: evaluator library
2. #29898: storage and audit types
3. **This PR:** enforcement and discovery
4. Control plane (experimental API, `coder exp experiment-rules` CLI,
docs)

> [!NOTE]
> This PR was written by an AI agent (Xum).

---
_Generated with [`xum`](https://github.com/coder/xum) • Model:
`anthropic:claude-opus-5-5` • Thinking: `high`_

---------

Signed-off-by: Thomas Kosiewski <tk@coder.com>

2026-09-29 12:23:00 +00:00