Files
dsh-facet/README.md
T
NotExpectedYet 5d06829118
Release / release (push) Successful in 6s
feat: pick the focus and the allowed paths from a directory listing
Both scope controls were free-text boxes, and the vocabulary they asked the user
for is paths relative to the session folder. "Which folder is `../shared`" is the
one question a text box cannot answer, so the panel now shows the Host's own
directory: the `uiWorkspace.listDirectory` primitive the workspace picker's
browser drives, one level of child folders, a breadcrumb trail, and a
hidden-entries toggle. The path field stays editable inside the picker, because
the Host lists folders and an extra may legitimately name a file.

Picks are stored in the form that keeps their meaning. A folder at
`/work/root/src/shared` is declared `src/shared` when the session started in
`/work/root`, and `../shared` going up; only a path with nothing shared — another
drive, another tree — stays absolute. The Focus row shows that declared form,
with the resolved path in its title, which is what the agent will see.

`Set` is gone with the box it committed: the panel writes on pick, as every other
control here already did. `Clear` stays, because empty means the session folder —
the schema's own default rather than an override naming the same place.

The card was 380px wide and grew without bound, a single column of prose taller
than the viewport. It is 760px now, with a two-column grid that `auto-fit`
collapses to one on a narrow card, the reference prose folded into `<details>`,
and a scroll-bounded body. `Modal` sizes its own card from a hashed class in the
primitives' stylesheet, so this bundle installs one small stylesheet of its own
and outranks that rule with a doubled class selector — no `!important`, and
nothing reaching into a class name it does not own.

Two fixes came with it. The footer's "Deployment mode:" line read a slot prop
that never carried the mode, so it said "ask" in every deployment; it reads the
settings document now. And `uiWorkspace` is resolved with `ctx.get` at the call
rather than injected, because the panel is still the only way to declare scope in
a deployment with no directory browser: it keeps the panel and says what it
cannot do, instead of losing the button.
2026-09-27 07:05:47 +01:00

365 lines
19 KiB
Markdown

# dsh-facet
Scope an agent session to **one facet of a system** — the folder the
conversation was started in, plus whatever you explicitly grant it.
A DeepSeek Harness bundle: one host row (the policy) and one agent preset (the
enforcement).
---
## Why this exists
A multi-faceted system is built from modules that look alike. In the project this
was written against there are fourteen controllers — the same idioms, the same
naming, the same shape — and a conversation about one of them that can see all
fourteen spends its first several steps working out which one it is looking at.
That is a quality problem before it is a cost problem, and the cost is real too:
in a long session the dominant expense is re-reading accumulated history, and
most of that history is file content the agent did not need.
So this narrows the world to one module rather than trying to steer through all
of them. It is the same idea as the `focus` argument in
[`dsh-recall`](https://github.com/green-dalii/dsh-recall) — which confines a
*search* to one module — applied to the whole agent: what it may read, where it
may search, and what its prompt says it is.
## What it does
Three things, and they are deliberately in three different places:
| | where | why there |
|---|---|---|
| **the rule** ("work inside the facet; ask to go outside") | system prompt | invariant for the session, so it does not invalidate the prompt cache |
| **the binding** (which folder, what is granted) | runtime-context snapshot | it changes when you grant something, and a changed system prompt costs a full-context cache miss |
| **the guard** | `tools/pre-execute` | the part that does not depend on the model agreeing |
**The scope is not configured — it is where you already are.** A session's
working directory is recorded immutably when the session is created, so starting
a conversation in `src/controllers/room/occupancy` makes that controller the
facet, and starting one at the project root simply leaves the session unscoped
(which is the honest answer for a conversation that is genuinely about
everything).
**Going outside is a question, not a refusal.** In that same project 41.6% of the
code — shared types, services, infrastructure — sits outside every controller. A
wall that cannot be opened would not confine the agent, it would only make it
wrong, and an agent that cannot reach a shared type does not stop needing it.
So an out-of-scope file call is put to you, and saying yes records the grant for
the rest of the conversation.
### The `facet` tool
| action | effect |
|---|---|
| `status` | the root, the mode, the standing grants and this session's grants |
| `add` | asks you to approve an extra path, then records it for this session |
| `remove` | drops a session grant |
| `reset` | returns to the session's own folder |
`add` is the only one that needs approval. It is also asked in `deny` mode,
because otherwise there would be no way to grant anything at all.
## Install
From the repository, following releases:
```sh
dsh plugin --profile web add "git+https://git.notexpectedyet.com/Not-Expected-Yet/dsh-facet.git#semver:^0.1.0"
dsh --profile web --dump-config | grep preset-facet # expect the row
```
Or from a checkout, while working on the plugin itself:
```sh
dsh plugin --profile web add "/path/to/dsh-facet"
```
The `#semver:^0.1.0` fragment is the difference between an install that can be
updated and one that cannot. Leave it off and pnpm resolves the default branch to
a **commit** and records that commit in `pnpm-lock.yaml` — `dsh plugin update`
then has no version to move to, because there is nothing to move between. With
the fragment, pnpm resolves against the repository's tags, which the release
workflow creates on every push to `master`. The fragment needs a tag to resolve
against, so it only works once the first release has been published.
Either way the command is `pnpm add` run inside the profile, and it registers the
bundle in `dsh.profile.bundles` for you. A git-hosted plugin that builds on
install can be blocked by pnpm's build policy; add the key pnpm asks for under
`allowBuilds` in the profile's `pnpm-workspace.yaml`. This bundle ships sources
with no build step, so it does not trip that.
The bundle contributes two layers:
- **`cordis.patch.yml`** — the host row `facet-policy`, which owns the mode and
the standing grants and publishes them as `ctx.facetPolicy`.
- **`presets/facet.patch.yml`** — the agent preset `facet`, which is the shipped
`standard` composition plus the guard and a configured compaction row.
Then pick **Facet** as the preset for a session — in the Web UI that is the
preset picker in the new-session surface, or *Settings → General → Agent preset*
to make it the default for new sessions. The preset is fixed once a session's
first turn has run, so choose it when you start the conversation.
### Configure it in your profile, not in the package
The host row's config is where this *deployment's* facts belong — the profile's
own `cordis.patch.yml`:
```yaml
- id: facet-policy
name: dsh-facet
config:
mode: ask # ask | deny | off
extraRoots: [] # optional: a seed of paths always in scope, every session
```
Both are also on the plugin's card under **Plugins** in the Web UI, which is the
same layer reached a different way: a settings write rewrites this entry in the
profile patch rather than storing anything new. An edit there is live — the guard
reads `mode` and `extraRoots` on every call, so a change decides the next file
call without a restart.
The one exception is `off`. Choosing it while a session is running stops the guard
immediately and empties the binding on the next request, because the binding is a
callback. The rule is not: it was registered at load as a plain string, and
nothing unregisters it, so the prompt would go on stating a rule that nothing
enforces. An `off` *deployment* is therefore settled at load — set it in the
patch. `ask` and `deny` swap freely in both directions.
`sectionOrder` and `contextOrder` are deliberately absent from the card. They
place the rule and the binding among the other prompt sections, which is decided
once at load, and no section is re-ordered without a remount.
### Per-session scope, and who gets to decide it
Scope has two layers, and the split is the point:
| layer | what it is | who can write it |
|---|---|---|
| **the user's declaration** | `scopes[<session id>]` in this row's config — a `focus` and a list of `extras` | only the context panel, through the settings service |
| **the agent's** | a narrowing `focus` and any session grants, held in memory | the agent, and only ever to *reduce* |
The effective root is the agent's narrowing while it still sits inside your root,
and your root otherwise. So an agent that goes wrong cannot widen its own scope,
and `facet action=reset` returns it to what you declared rather than to the
session directory — landing on the directory would have made `reset` a widening.
You set the declaration from the **crosshair button on the composer tool row**,
which appears only in sessions running the `facet` preset. It shows the focus and
the allowed paths, adds and removes them, and carries the mode details and usage
notes that the preset picker's guide dialog cannot show for a preset no core
package ships. Writes go to the profile patch, so reopening a chat gives back the
scope it had.
Both controls are a **directory listing**, not a text box: the panel's vocabulary
is paths relative to the session folder, and "which folder is `../shared`" is the
one question free text cannot answer. Pressing it opens a second dialog over the
panel, shaped like the host's own workspace browser — one level of child folders,
a breadcrumb trail, a hidden-entries toggle — driven through
`uiWorkspace.listDirectory`, the same Host primitive that browser uses. A pick is
stored in the form that keeps its meaning: `/work/root/src/shared` becomes
`src/shared` when the session started in `/work/root`. The picker's own path field
stays editable, because the Host lists folders and an extra may legitimately name
a file.
Because the panel owns widening, `facet action=add` is the *other* path and the
narrower one: it is gated on an approval channel, and a deployment with none — the
`danger-full-access` permission preset sets `approval: never` — refuses it before
anyone is asked. In that deployment the panel is the only way to widen, which is
exactly what the binding tells the agent.
`extraRoots` is only a seed now. Before per-session scope it was the *only* way to
widen anything, which meant "this conversation needs the entity folder" had to be
written as a fact about every conversation in every workspace. It is for paths a
deployment genuinely always wants, and it is empty by default.
## The compaction row
`presets/facet.patch.yml` also configures `compaction-basic`, because a scoped
session and a budgeted context are two halves of one idea — scoping reduces
context at the source, compaction is the ceiling:
```yaml
thresholdRatio: 0.35 # a 350k ceiling, not the shipped 800k
retainTokens: 80000 # an absolute working set, not 16% of a 1M window
maxTokens: 16384 # the summarisation cap, stated rather than inherited
```
The shipped ratio is `0.80` of a 1,000,000-token window, i.e. 800k. Measured
across this deployment's 40 recorded sessions, the deepest ever reached 787,668
pressure tokens — **the trigger had never once fired**, so every token below it
was re-read on every subsequent step. `maxTokens` is worth stating explicitly:
its default is not 8192 but `headroomTokens`, 65536.
If you want that budget on *every* session rather than only scoped ones,
`presets/standard-compaction.patch.yml` is an opt-in overlay that applies the
same row to the shipped `standard` preset. It is not in the bundle's default
patch list, and the header of that file explains why: a preset declaration
replaces the whole plugin list rather than merging into it, so the file is a
frozen copy of `standard` that will not follow the harness when `standard`
changes.
## What it does not cover
Honest boundaries, because each one is a decision rather than an oversight:
- **`bash` is not guarded.** A shell command cannot be judged reliably from
argv, so the facet is enforced on the file tools and `bash` remains a
legitimate way out. The prompt tells the model this rather than pretending
otherwise.
- **Tools that name a path in a shape the table cannot read are not guessed at** —
including MCP tools. See `PATH_ARGUMENTS` in `lib/scope.js`; it is a list of
tools whose path argument is a plain top-level string.
- **This guards the tool pipeline, not the filesystem.** There is no read seam in
the harness — `dsh-fs-sandbox` states plainly that "reads pass through
untouched: every mode permits reading" — so anything that reads through
`ctx.fs` without going through tool dispatch (another plugin, the HTTP
workspace-files endpoint) is unaffected. Fencing *all* readers would mean
replacing `ctx.fs` with a stricter `FileSystem` subclass.
- **A grant is not persisted**, so it does not survive a process restart.
- **Sessions at the project root are not scoped at all**, which is deliberate.
## What has been verified, and what has not
Verified:
- 73 tests (`npm test`) over the scope algebra, the policy state, the guard, and
the browser half — including that containment is decided on canonical targets
(a symlink inside the facet pointing outside it is refused), that a prefix is
not a parent (`/a/room` does not contain `/a/roommate`), that a malformed tool
call never makes the guard throw, that a live `mode` edit decides the next call
without the guard being remounted, that an agent narrowing the user has since
narrowed past is dropped rather than honoured, that `reset` lands on the
declared scope rather than the session directory, and that every `facet` result
satisfies the tool's own declared output schema.
- The bundle's patch layers parse, merge and land correctly against the real
profile: `dsh --profile web --patch ./cordis.patch.yml --patch
./presets/facet.patch.yml --dump-config` shows `facet-policy`, `preset-facet`,
the `facet-guard` row and the compaction config, with `preset-standard`
untouched. The optional `standard-compaction` overlay was checked the same way.
**Not yet verified: a live session.** The plugin has not been loaded by a running
harness, because that needs the bundle installed. The first thing to check after
installing is that `preset-facet` is not `broken` — a preset whose child row
fails to resolve still appears on the roster, with its diagnostic, and cannot
compose a session:
```sh
dsh --profile web --dump-config | grep -A 3 preset-facet
```
## Releases
`.gitea/workflows/release.yml` runs on every push to `master`: it works out the
range since the last tag, bumps `package.json`, prepends a `CHANGELOG.md`
section, commits and tags `vX.Y.Z`, then creates the Gitea release with the
commit subjects as the body.
The bug it fixes is not tidiness. Installed as a bare `git+https://…/repo.git`,
a plugin is pinned to a **commit** in the profile's lockfile, so with no tags
there is nothing for `dsh plugin update` to move to. Point the profile at
`#semver:^0.1.0` and the updates follow releases instead:
```jsonc
"dsh-facet": "git+https://git.notexpectedyet.com/Not-Expected-Yet/dsh-facet.git#semver:^0.1.0"
```
The bump is read from the commit subjects being released:
| in a commit subject | effect |
|---|---|
| `feat`, as its own word (`feat:`, `feat(x):`) | minor — `0.1.0` → `0.2.0` |
| `feat!:` / `fix(x)!:` / `BREAKING` | major — `0.1.0` → `1.0.0` |
| `fix:`, `docs:`, prose, anything else | patch — `0.1.0` → `0.1.1` |
| `[minor]` / `[major]` | force the bump when the subject cannot carry it |
| `[skip release]` | this push publishes nothing |
Only the **subject** line is read, and `feat` has to be a word — `Fix the feature
flag leak` is a patch, not a minor. With no tags yet, the version already in
`package.json` is what gets tagged, so the first release matches the code.
The logic lives in `.gitea/scripts/release.sh` so it runs without a runner:
`npm run test:release` exercises it — an empty range, a repository with no tags
yet, a `feat` that is only a word — `npm run release:plan` says whether the
current HEAD would release, and `npm run release:notes` prints the notes it
would publish.
## Development
```sh
node scripts/dev-links.mjs # link the harness's cordis/schemastery/dsh-tools
npm test
```
`dev-links.mjs` links rather than installs those three packages, and for
`@deepseek-ai/cordis` that is a correctness requirement rather than tidiness: a
second copy of Cordis means a second `Context` class, two service registries in
one process, and a plugin whose services are invisible to the harness that loaded
it.
## Layout
```
index.js host row: the policy, published as ctx.facetPolicy
client/client.js browser half: the settings card, and the context panel
lib/policy.js the stateful half — scope resolution, grants
lib/scope.js the pure half — argument extraction, text, reasons
lib/guard.js preset row: rule, binding, tool, pre-execute guard
presets/facet.patch.yml the `facet` agent preset
presets/standard-compaction.patch.yml optional: the same budget for `standard`
test/ 78 tests
```
`client/client.js` is a browser half rather than a module the host imports, and
it is not optional decoration. The Plugins page does not build a card from a
`Config` schema — it draws only what browser halves register into the slots it
declares. `volatile()` on the host schema makes the namespace *servable*
(`settings.describe()` drops an entry with none, and the card registers only
while the namespace is served), but a schema alone puts nothing on the page.
The two keys in that file are deliberately different strings, and that is the
easy mistake to make here: the settings **document** is filed under the profile
entry id (`facet-policy`, the row's `id` in `cordis.patch.yml`) while the
**card** is keyed by the package name (`dsh-facet`). Every sibling plugin
happens to name them identically, so binding the form to the wrong one produces
a card that renders, saves, and writes nowhere. `test/client.test.js` pins both.
The same file carries the context panel, registered into
`conversation.input.left` — the composer tool row — and gated on the session's own
`agentPreset` projection, so it appears only where the facet is actually in force.
That gate is why `scopes` is `volatile()` in the host schema: `settings.mutate`
refuses a path that is not under a volatile node, and refusing is exactly right —
the panel is the only writer, and a plugin that could not serve the namespace
would be a control that silently stored nothing.
The listing reaches the panel through `ctx.get('uiWorkspace')` rather than through
an injection, and the injection's `available()` is asked on every render. Both are
deliberate: `ctx.inject` would make the whole context button conditional on a
service the plugin does not own, and the panel is still the only way to declare
scope in a deployment with no directory browser — it would simply have no listing
to offer, which is what it says instead. The service is resolved at the call
because the renderer memoizes a registration's inject result, so a boolean read
there would freeze at whatever the first render saw.
The two dialogs are wider than the primitive's 380px card, which is fixed in the
primitives' own stylesheet. `client/client.js` installs one small stylesheet of
its own — the widths and a row hover — and passes each card a class through
`Modal`'s `className`. Every selector in it doubles the class it names, so it
outranks the single hashed class the primitive applies whatever order the two
stylesheets land in; there is no `!important` and nothing reaching into a class
name this bundle does not own. The panel's body is a two-column grid that
`auto-fit` collapses to one on a narrow card, and each dialog scrolls a bounded
body rather than growing past the viewport — the panel used to be a tall single
column of prose on a 380px card, which is what it looked like.
## Also in this repository
- [`facet-preset.md`](../facet-preset.md) — the design note this was built from,
including the harness mechanisms it rests on.
- [`compaction-tuning.md`](../compaction-tuning.md) — the investigation into why
compaction had never fired. Its measurements reproduce exactly; note that its
"The change" section prescribes `$DSH_HOME/.agent-presets/`, a directory
**nothing reads any more** on the installed harness. Presets are bundle patch
declarations now, which is what `presets/facet.patch.yml` is.