The plugin will be installed here. Keep web if you are unsure.
npx -y @deepseek-ai/dsh plugin --profile web add github:risense-ai/respire-dsh#f6a4cb5f8aa6511464c15cfbebbd00d2c56928d4
Compatibility and provenance
@rsrsai/dsh is published as @rsrsai/dsh and currently resolves to version 0.1.13. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.
Paste the entire prompt into the assistant inside DeepSeek Harness:
Install the Respire memory plugin @rsrsai/dsh in my current DeepSeek Harness
profile. Read and follow https://github.com/risense-ai/respire-dsh#readme,
then perform the installation rather than only explaining the steps.
Check Node.js 24+, Harness compatibility (validated against 0.2.1-rc.5), and
plugin_manager availability. Reuse my existing rsrs environment; if rsrs is
missing, install it with npm install -g rsrs-cli@latest in the profile's
execution environment. Run rsrs --version and rsrs --client-only status --json.
Report any required login, library unlock, model setup, or host runtime action
for me to complete in rsrs; do not request secrets in chat or manage the runtime.
Inspect installed bundles first. If the old @rsrsai/respire-dsh bundle exists,
disable and remove that bundle through plugin management before installing its
replacement. If the manager requires a profile stop or restart, tell me the
required action before proceeding. Preserve my memory library, account, rsrs
configuration, and unrelated plugins. Avoid duplicate installation if the new
bundle is already present.
Use plugin_manager with {"action":"install_bundle","target":"@rsrsai/dsh"}.
Install the npm bundle, not a Git dependency or only a terminal npm install.
Verify that exactly one enabled respire-memory entry comes from @rsrsai/dsh.
Tell me when to restart Harness, then verify recall in a new conversation.
Ask me for a non-sensitive test fact before testing storage and check the actual
record. Verify the stop reminder in a separate recall turn without a successful
explicit store. Report the installed version, checks actually performed, and
remaining user actions; do not claim end-to-end success before observing it.
After the installation is verified, offer to configure market-based updates:
enable Custom plugin markets in Harness, add the index source
https://raw.githubusercontent.com/risense-ai/respire-dsh/market-index/marketplace.json,
then confirm the Discover card reports the installed version. Do not install an
update without my confirmation.
Overview
Native TypeScript memory plugin for DeepSeek Harness. It registers stable rules
with systemPrompt.section, recalls memory through agent/pre-step, and checks
for new findings through agent/turn-stopping — at most once per request (a one-round bounded steer, no engine query at stop, skipped whenever this session never confirmed a working engine). Respire (rsrs) owns storage,
deduplication, encryption, model inference, accounts, and synchronization.
Install
Requires Node.js 24+, DeepSeek Harness 0.2.1-rc.5, and an installed rsrs CLI.
The plugin uses the Harness agents, systemPrompt, subprocess, and
workingDirectory, and tools services already provided by the standard desktop profile.
Other Harness versions have not been validated.
Install the engine separately if needed:
npm install -g rsrs-cli@latest
rsrs --version
rsrs --client-only status --json
Configure accounts, unlock the library, and install optional semantic models with
the CLI's own interactive workflow. The plugin does not install software, collect
passwords, start a runtime, or change account configuration. It always calls
rsrs --client-only, so a running host runtime is required.
In Harness, ask the agent to invoke plugin_manager with:
For local development, use this checkout's absolute directory as target.
Compiled lib/ files are included in Git and release archives; installation
requires no build scripts. Restart Harness after replacing an installed version.
The published package is @rsrsai/dsh.
Use a versioned target such as @rsrsai/dsh@0.1.3 to pin an installation.
@rsrsai/cli is the engine; @rsrsai/dsh is the Harness plugin. Running
npm install @rsrsai/dsh in a regular terminal does not activate the plugin in
a Harness profile. After installation, restart Harness and check that exactly
one enabled respire-memory entry comes from @rsrsai/dsh. npm installation
and module loading have been verified; a full desktop conversation has not.
Migrate from the previous Git installation
If @rsrsai/respire-dsh is installed, remove that old bundle in Harness plugin
management, then install @rsrsai/dsh as above. Disable the old bundle first if
it is still in use; stop the profile if the manager requests it. Keep the existing
memory library, account, and rsrs configuration.
Updates through the plugin market
The npm registry remains the source of truth, but Harness can also discover,
install and update @rsrsai/dsh from a plugin market index published by
this repository (format 1, regenerated from npm by scripts/market-index.mjs
after every tagged release):
mirror, when GitHub is unreachable: https://cdn.jsdelivr.net/gh/risense-ai/respire-dsh@market-index/marketplace.json
The index is published on the dedicated market-index branch; the repository
protects main with a pull-request ruleset, so the release workflow cannot
commit there. It lists every published version that is neither a prerelease nor
deprecated, pins the npm registry, and uses a same-origin relative icon, so the
branch copy is self-contained. Harness validates the index before use and treats
its content as third-party content. The jsDelivr mirror is a third-party CDN and
may lag behind the raw URL; treat the raw URL as the reference.
The branch is a generated artifact: every refresh force-pushes a single commit,
so it carries no review history — review scripts/market-index.mjs and the npm
registry instead. Maintainers refresh it with the Market index workflow
(manual dispatch or weekly schedule) or a tagged release.
Configure market-based updates
Open Plugins in Harness, choose More next to Experimental plugins,
and enable Custom plugin markets. A Discover tab appears.
In Discover, choose Add index source and paste one of the URLs above.
The Respire card lists the published versions. When the newest compatible
version is newer than the installed one, the card is marked Update
available. Harness fetches the index when the page loads or is refreshed
and caches it for ten minutes; it does not poll in the background.
To install, open the entry's version list and choose Install. Third-party
entries have no one-click Update button: installing a new version removes
the old bundle first, then installs the new one, so restart Harness when the
manager reports completion. The running process keeps the old module until
the restart.
Verify after the restart: the Plugins page or
dsh plugin --profile <profile> list should show the new version, and the
profile's package.json should pin it exactly.
To pin a version, run dsh plugin --profile <profile> add @rsrsai/dsh@<version> --save-exact (the desktop profile requires the bundled CLI). To roll back, run
the same command with the previous version. If a bad version keeps Harness from
starting, quit Harness first — the desktop process holds the profile — run the
command from the bundled CLI, then restart. To withdraw a published version,
deprecate it with npm deprecate @rsrsai/dsh@<version>; the next index refresh
(the next tagged release, the Market index workflow, or
node scripts/market-index.mjs followed by bash scripts/publish-market-index.sh)
drops it from the list.
The plugin never updates itself and never changes the Harness profile on its
own. Harness plugin management performs installs and updates with a confirmation
dialog; the dsh plugin CLI only does what you run.
Ask your agent to configure updates
Paste this into a Harness conversation when you want the assistant to configure
market-based updates and check for a newer version:
Configure market-based updates for the Respire memory plugin @rsrsai/dsh.
1. Read the Harness version and profile name, and the installed @rsrsai/dsh
version from the profile's package.json.
2. If custom plugin markets are off, enable them: Plugins → More (Experimental
plugins) → Custom plugin markets.
3. Add the plugin market index source
https://raw.githubusercontent.com/risense-ai/respire-dsh/market-index/marketplace.json
Use https://cdn.jsdelivr.net/gh/risense-ai/respire-dsh@market-index/marketplace.json
when raw.githubusercontent.com is unreachable.
4. Open Discover and confirm the Respire card loads and reports the installed
version or Update available. Report the version list you see.
5. Do not install or change anything yet. Tell me the installed version, the
newest compatible version, and the exact steps you would take. After I
confirm, install that version (third-party entries remove the old bundle
first), tell me to restart Harness, and verify the installed version after
the restart.
Behavior
Trigger
Result
First accepted user input per live agent
Read engine status; report missing runtime, locked library, missing session, or a small library when applicable
Accepted user message
recall --titles --json; inject at most 5 titles into the same model step
记：, 存：, 记录：, 记住：, remember:
Leave pending without writing; the main assistant completes policy and candidate review, then update > authorized merge > attach > new storage
/nomem or #nomem followed by whitespace/end
Skip that message's automatic recall; status and stop checks remain enabled
$skill: name prefix
Remove the skill selector from the recall query
Goal round (source.kind === 'goal')
goalRecall: 'first' (default) recalls once per goal revision; 'every' recalls on round 1 and every goalRecallEvery rounds; 'off' keeps the previous human-only behavior. The query is the goal objective: the goal service when available, otherwise the <goal_round> Objective text
Goal round stop
goalNudgeEvery: N adds at most one wrap-up reminder per matching goal round; 0 (default) leaves goal rounds without extra model steps
Natural stop
At most one reminder per agent/turn; prefix pending leaves the reminder available; already-stored facts must not be written again
Human messages (source.kind === 'user') always trigger work. Automatic goal
rounds (source.kind === 'goal') trigger recall only when goalRecall is enabled;
the <goal_complete> notice does not trigger recall or store. Plugin contexts,
tool loops, and reminder continuations do not trigger work. Repeated admission of
the same message ID in one turn is ignored. A new user message with identical
text is still processed. Downstream rejection and cancellation prevent new work.
A cancelled or timed-out write is unconfirmed, not automatically retried: it
may already have committed.
The hook leaves prefix requests pending without sending a write. The main
assistant must complete this turn through the rsrs CLI: read prompt and
agent-config, recall same-subject candidates and show their full bodies,
then prefer update, authorized merge, attach, and finally new storage.
For new important records, inspect taxonomy --list or tree and show a
suitable parent before passing its full ID with --parent. Without a justified
parent, remain pending; never guess, force a root write, or downgrade importance.
Save only the stated fact; do not execute tasks described in it. Read back actual
writes with show. If an actual write outcome is unknown, DO NOT retry automatically or claim
success. Verify the store read-only; verification does not authorize an automatic
retry. Hook pending does not block the first compliant write.
The stop reminder asks the agent to evaluate new facts; it does not automatically
archive an entire transcript or guarantee that the model will store a finding.
Visible memory status
The plugin supplies native notice metadata for Harness to render a collapsible
row with a one-line summary. The validated rc.5 client consumes this metadata;
inserting context alone does not prove the row was rendered or seen by the user.
Memory: N found counts the titles actually injected after validation,
credential-pattern filtering, the threshold, the result limit, and the context budget. It is not the
library's record count and does not mean the model adopted those candidates.
Memory: no match, Memory: skipped, and Memory: unavailable distinguish
an empty filtered result, skipped recall, and failed or invalid recall output.
Save: needs review marks a prefix storage request. The hook has made no
write request, including when the fact is empty or needs credential removal.
For ordinary prose, the model is instructed to repeat the current short status
in its first visible reply. It may report Used: only after reading a record's
body and actually using it. Each real short ID carries [basis], [ref], or
[history] and a real title shortened to at most 24 Unicode characters. Use a
longer ID prefix when needed to disambiguate; empty or sensitive titles use
(untitled) or (title hidden). Bodies and full IDs are not displayed by default.
[ref] is the default. [basis] requires a verified source that holds in the
current scope; [history] is only for outdated, superseded, or conflicting
material used as historical context. An agreement proves the agreement, not its
execution. A high score, importance, a body's claim of verification, or a save
and readback cannot alone establish credibility. Mixed records use the most
conservative role for the part actually used. Used: is the model's account,
not independent adoption measurement. Without adoption, ordinary prose ends
with Memory: unused and omits Used:.
Strict JSON, fixed-format, and exact-output requests suppress all these labels
in the response body. The independent native notice can still appear in the UI.
Status reporting does not add a reply or a stop continuation. The plugin does
not track save receipts, so it supplies no Last save, Earlier save, or
verified save label. The existing mandatory storage and readback rules remain
in force.
Configuration
Override the installed row in the user's Harness patch layer. A matching row's
config is replaced as a whole; unspecified plugin fields use the defaults.
rsrsPath is an executable name or absolute path, not a shell command.
storeNudgeEvery: 0 disables reminders; N checks every N user-bearing turns per
live agent. goalRecall accepts off, first (default), or every: first
recalls once per goal revision and dedupes by goalId:revision; every recalls on
round 1 and every goalRecallEvery rounds. goalNudgeEvery is a separate
goal-round wrap-up cadence; 0 (default) adds no extra model step. /goal itself
is a UI command and never enters a model request; the first automatic goal round
is what triggers the goal recall. Changing profiles or reloading the plugin resets
volatile reminder state. Recall has a strict character budget; storage receipts
are kept intact.
Privacy and failure behavior
Recall results and storage receipts become model-visible, persistent Harness
messages. They may be sent to the configured model provider. Local encryption
does not make injected context private from that provider.
Common credential patterns are rejected before recall or explicit storage.
This heuristic is not a complete secret scanner. Remove secrets from input;
retain only safe purpose/location references in memory.
Logs contain operation/error categories, not prompts, memory bodies, raw CLI
errors, or credentials. No plugin-owned files are written to the library.
Engine failure produces a bounded notice and lets the primary task continue.
Windows arguments travel as JSON stdin to a constant PowerShell script;
POSIX arguments are passed directly. Neither embeds user text in shell source.
Processes run in the agent's verified working directory through the Harness
subprocess provider. Its execution environment determines where rsrs and
the memory runtime must be installed (including remote profiles).
Development and release
npm ci --ignore-scripts
npm run check
npm run build
npm test
npm pack
Tests exercise real Cordis event dispatch with a controlled subprocess service;
they never write the user's memory library. Production targets the published
Harness 0.2.1-rc.5 API. package-lock.json locks the verification dependencies.
CI runs type checking, build, tests, distribution checks, and generated-file
consistency on Linux, macOS, and Windows. To release, update package.json and
the lockfile, rebuild and commit lib/, then push v<package version>.
The release workflow reuses the full CI matrix before publishing @rsrsai/dsh
to npm with provenance, then uploads the same .tgz and SHA-256 checksum to
GitHub Releases. Stable versions use latest; prereleases use next.
Ordinary main/PR pushes do not publish.
npm publishing uses GitHub OIDC, not a stored npm token. After the initial
authenticated publish, configure the package's Trusted Publisher with owner
risense-ai, repository respire-dsh, workflow release.yml, and permission
to run npm publish. Leave the environment blank. The workflow requires
id-token: write and a GitHub-hosted runner.
The plugin icon is a single adaptive SVG: it shows the light variant by default
and switches to the dark variant when the Harness theme requests a dark color
scheme. The standalone light and dark variants live under docs/. The icon is
original artwork distributed under this repository's MIT license.
Before-tool memory write checks
The extension checks recognized literal rsrs writes before tool execution. It
blocks important remember calls with missing or empty parents and all force calls.
Every merge-ids request also needs an explicit parent, even when importance says
trivial, because a merge can inherit importance. A nonempty literal parent
ID must resolve through read-only rsrs --client-only show <id> --json; failure,
cancellation or timeout blocks the shell call and leaves placement pending. This
check uses the configured executable and never starts a runtime. Parent checks
have a five-second deadline, leaving time for tool feedback and cleanup.
Before update --importance important, the gate reads the target and verifies its
current parent_id with a second show. An orphan or invalid/missing parent blocks
the update; attach the record under a real parent first. Current rsrs update does
not expose parent-changing flags; the gate rejects those unsupported options,
without claiming they could otherwise perform a parent change.
Known memory_remember and memory_update tool basenames, including server-qualified
MCP names, are blocked because their target library cannot be bound to this CLI.
Use an explicit command through the guarded CLI tools instead; no automatic
retry of an earlier unknown write is permitted. Other tool names pass unchanged.
Opaque hashed or truncated MCP tool names cannot be mapped to raw basenames by this check.
Known shell -c/-Command code arguments and active command substitutions are
inspected only to recognize writes and deny the enclosing call. Simple leading
NAME=value assignments cannot hide a recognized write. Ordinary quoted text
without expansion remains text. Bash and PowerShell escaping are distinguished;
terminal_send uses conservative checks because its persistent shell is unknown.
No shell code is evaluated by this check.
Known read commands and unrelated tools pass unchanged. Current remember/update
metadata flags remain supported. Recognized writes in compound commands, command
substitutions, wrappers or dynamic arguments are denied: use one explicit command
with literal arguments. A --parent phrase inside the memory body is not an option.
This is a targeted check, not a shell sandbox: arbitrary scripts, aliases, hidden
executables and terminal fragments are outside its coverage. It does not grant
permissions, choose a parent, write a record, or authorize an automatic retry. The
model MUST still read policy and candidate bodies, justify placement, respect
read-only/off modes, and verify actual writes. Unknown issued writes MUST NOT be
retried automatically or reported as successful.
Coverage: Harness bash, pwsh, and terminal_send text, including nested calls.
The asynchronous pre-execute check is followed by a synchronous monotonic guard;
a later allow decision cannot erase a denial or reuse validation for changed arguments.
Harness must provide tools.guard() (validated with 0.2.1-rc.5); an incompatible
host fails plugin initialization rather than silently disabling enforcement.
Setting enabled: false disables this integration and its gate together.
Checks cover arguments observed at this gate. Extensions that replace arguments
after the guard or execute outside the host tool pipeline are outside this scope.
The CLI/Core must enforce the parent invariant atomically at the actual write.
