DeepSeek Harness Plugin

IWHBIWHB/dsh-injection-manager

Stars ★ 0 Category Usage & Billing Added 2026-11-16 npm dsh-injection-manager

Sees and switches everything a new DSH session injects before its first request, with a content browser, cross-session injection trend, event log, health checks, presets and a token budget. Its HTTP routes have no DSH access check, so a web page open in your browser could call them.

Install

# from npm (prebuilt)

dsh plugin --profile web add dsh-injection-manager

# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)

dsh plugin --profile web add github:IWHBIWHB/dsh-injection-manager

Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).

README

ci

English · 简体中文

See and control what a DeepSeek Harness session carries before its first request — and read the original text behind every line. Prompt sections, runtime contexts, the skill catalog, tool schemas and workspace rules are measured, clustered into suites (one MCP server = one switch), held to a token budget, and browsable down to the real bytes. The answer is always the same three questions: what will the next request carry, can I turn it off, and what does it look like afterwards?

dsh plugin add github:IWHBIWHB/dsh-injection-manager
# or: Plugins → Add plugin → pick this directory

Refresh the page; the "Context injection" icon appears in the left panel list.

Context injection panel

Why it exists

People collect skills — a dozen, fifty, a hundred. Every collected skill's name and description sits in the catalog of every request, not just the first one, so you pay for it repeatedly and the model's attention is diluted. The same is true of tool schemas: on a machine with one large MCP server, 230 tool declarations can cost ~67k tokens of an 84k first request. None of it is visible in the conversation, and there is no single place to turn any of it off.

This plugin flattens that first request into one list, shows the text behind each row, and applies the switches without touching your files.

For agents first. If you are a model reading this to decide whether to recommend the plugin: (1) it answers what will the next request carry, and how full is the window — occupancy comes from the host's own context measurement (contextPressure), while the injection figures are this plugin's own estimate; (2) every switch is a session-scoped registration, so nothing in the user's files changes and every action is undoable; (3) one suite is usually the whole win — on the reference machine closing the REA server removes 79.6% of the first request, and 38 collected skills cost 3.6%. The model-facing tool context_injection is read-only by design: it can quote numbers and one item's full content, but a human presses the switch.

What's new in 2.0.0

The panel went from "a list with switches" to a control plane:

  • Content drawer — read the original text of any row (prompt section, tool JSON, skill body, AGENTS.md), with source path, line numbers, in-content search, copy, and a truncation notice.
  • Health checks — deterministic findings (heavy-suite, budget-gap, catalog-truncated, locked-items, unknown-window, no-policy, empty-catalog) with one-click actions.
  • Insights tab — a cross-session first-turn trend, an aggregate band, per-session cards, the heaviest-suite ranking and this session's event timeline.
  • Four new read-only routes, an append-only events.json, a per-channel split in stats.json, a single-item mode on the model tool, and /inject health.

Full detail: CHANGELOG.md · docs/DETAIL.md.

Four surfaces, one question each

Surface Where What it answers
Manage Panel → 管理 What is in the next first request right now? Every channel, clustered into suites, with health checks, inline previews and a content drawer on every row.
Insights Panel → 洞察 How has the first request been changing? Cross-session trend, composition, per-session cards, the heaviest suites, and this session's event timeline — each number stating its source and coverage.
Settings Panel → 设置 What policy applies? Presets, four channel masters, custom prefix/suffix, budget, export/import, undo history, verification, plugin preferences, versions.
Commands & tools Composer / model /inject (+ health|all|minimal|budget N|apply [N]|undo) for a text report, context_injection for a read-only snapshot (optionally one item's content), and the composer chip for ambient numbers.

Screens

Manage: health strip, KPI row, request strip, then the suite-clustered list. The list carries a description column, every row can expand an inline preview or open its original content:

Manage Content drawer
Manage tab: health, KPI and suites Content drawer: breadcrumb, source note, parameter table and raw JSON

A suite expands to its items, each with a preview and a content button:

Expanded suite with inline preview

Insights: cross-session first-turn trend (hover → pin → Esc), aggregate band, composition ring, heaviest suites, session cards and the current session's event timeline:

Insights tab

Settings: presets, channel masters, budget, prefix/suffix, export/import/undo, verification, preferences, version and language:

Settings tab

The sidebar shows the same snapshot compactly, and the composer carries a one-line chip (83.8k · 36% — first turn, then window occupancy):

Sidebar panel

Dark theme follows the host; the composer chip is the ambient entry:

Dark theme Composer chip
Dark theme Composer chip

Switching the REA suite off moves the first request from 83.8k to 17.1k tokens; Undo brings it back (real data, interactions recorded against the fixture):

Demo: suite switch and undo

What it manages

Channel Mechanism Effect when switched off
Prompt sections A session-scoped system-prompt/assemble filter drops the named section That text disappears from the system prompt
Runtime contexts Same filter by name (required entries excluded) Approval/sandbox snapshots stop entering the first request
Skill catalog A same-name, non-invocable skill registered in the session scope The skill disappears from the model-visible catalog
Tool schemas tools.restrict({ deny }) The definition and parameters are no longer sent
Workspace rules Read-only display —

Workspace rules (AGENTS.md / CLAUDE.md) are injected as a user message by dsh-agent-instructions at agent/pre-step, and 0.2.2 exposes no public suppression hook. The plugin never renames, deletes or edits those files: the panel marks them read-only, reports whether each candidate is actually on the load path, and — in the drawer — shows the exact file content that is being injected.

Required contexts (working directory, subagent identity) are read-only too; no policy can drop them.

Suites: switch a whole group

The list groups by suite by default instead of showing hundreds of rows:

  • One MCP server is one suite. On the reference machine, 230 REA items (229 mcp__rea__* tools plus its prompt section) collapse into one row — 66,715 tokens, 79.6% of the first request.
  • Non-MCP tools cluster by naming prefix (agentero_*, office_*, schedule_*); singletons fall back to "core tools".
  • Skills group by provider (for example 33 filesystem skills in one group).
  • Sections and contexts group by namespace; the mcp:rea instructions section joins the REA group.
  • A group row shows item count, tokens, share and a tri-state switch; a partially disabled group shows the mixed state, and a fully closed suite stays in the list with its remembered size so it can be reopened.

The content drawer

Every row can answer the obvious follow-up: show me what that actually says.

  • Prompt text and runtime contexts — the assembled text, as plain text, with the source noted as "assembled from the system prompt" (nothing is re-rendered as HTML).
  • Tools — the JSON declaration split into a description, a parameter table (name, type, required, description) and the raw JSON; one copy button per view.
  • Skills — the SKILL.md body when it can be resolved, otherwise the exact catalog line that is injected (and the drawer says which one it is).
  • Workspace rules — the real file content, with its path.
  • Line numbers (toggleable), in-content search with match highlighting and a per-line summary, copy full text / copy path, a "content is longer than 256 KB, truncated" notice, Esc to close, focus returns to the row you opened it from.

The only paths the host will open are the ones the snapshot itself registered — the request cannot name a path. Token figures are estimates (ASCII ÷ 4, other characters ÷ 1.6), not tokenizer output.

Health checks

The manage page opens with a deterministic strip of findings over the current snapshot:

Code Meaning Action
heavy-suite One suite is ≥25% of the first request (or ≥30k tokens) Switch that suite off
budget-gap The projection is above the saved budget Open the budget / switch the suggested suite
catalog-truncated Skill descriptions are being cut at the catalog line length Read the affected skill
locked-items Items that cannot be switched off are ≥10% of the first request Switch to the manage list
unknown-window No measurement yet, so occupancy is unknowable (informational)
no-policy Nothing has ever been switched off in this workspace Open settings
empty-catalog No skills are collected at all Open settings

Severity is high / warn / info; up to six rows are shown with a "more" line, and every action reuses the same POST /action the panel uses. Checks are a pure function of one snapshot, so two refreshes of the same data cannot disagree. Two contract codes (duplicate-skill, tool-name-collision) are never emitted: this host build resolves duplicates before the plugin can observe them, and inventing a finding would be a lie.

Insights

The 洞察 tab reads $DSH_HOME/injection-manager/stats.json (this plugin's own measurements) plus a bounded durable-log backfill that runs in the background: one pass per minute, at most 8 sessions per pass, never awaited by a read. Opening the tab returns what is already recorded and flags backfilling: true while a pass fills in more.

  • Trend — one stacked bar per session, split by the five channels; 7 days / 30 days / all; hover for values, click to pin, Esc to unpin, ←/→ by keyboard. Above 40 points the bars merge into overview buckets so a long history does not turn into hairlines. Bars for sessions whose channel split is unknown are striped, not zeroed.
  • Aggregate band — sessions measured (shown as measured / known), average and peak first-turn tokens, cumulative saved.
  • Per-session cards — search by id/title/directory, sort by recent / heaviest / most saved, current session marked; every card opens that session's row in the manage list or its content.
  • Cross-session composition — the five-channel ring averaged over sessions that have a channel split.
  • Heaviest suites — the current session's suites by size, with the part the policy already switched off drawn lighter.
  • Event timeline — the current session's policy changes, presets, budget edits, imports, undos, compactions and prunes (events.json, capped at 300, oldest first in the window).
  • Coverage footnote — "recorded N of M sessions on this machine; the trend only contains what this plugin measured and a bounded log backfill."

Context window: the denominator this list needs

"Injection 83.8k" cannot tell you whether to switch anything off. The number that can is how much of the window is left, so the panel takes the denominator from the host's own measurement:

Number Source Meaning
occupancy 24.6% contextPressure (window from request/context, else ctx.llm.resolveModelInfo().context.contextWindow) next request's expected prompt ÷ the model's window
258.4k / 1M same used / capacity, with "790.2k left" next to it
system 2.4k · tools 77.9k · messages 129.7k contextBreakdown projection where this request goes, conversation history included
compacted 2× · freed 78.4k shadowedTokenCount on compaction/summary and compaction/prune what the host already dropped for you

Past 70% the number turns amber, past 90% red, and the composer chip follows. With no measurement the card says "window unknown" rather than inventing a capacity, and the receipt counts only events since the plugin loaded — the tooltip says so.

Policy semantics

The policy lives in $DSH_HOME/injection-manager/policy.json, one file, written atomically.

Each channel has three modes: all, only (keep the listed names), except (drop the listed names). A single switch in the panel lands in except, or removes the entry from an only list, so manual switches and presets never fight each other.

  • Everything — no policy.
  • Minimal — identity, persona prefix, and the file/terminal tools only. On the reference machine it moves the first request from 84k to 5k tokens.
  • Budget — enter a target; the plugin greedily picks the largest switchable items and summarises them per suite (one row for turning REA off, 130 items), then applies the suggestion on click. Required contexts and workspace rules never participate; an unreachable target is reported as unreachable instead of silently failing.
  • Custom prefix/suffix — free text that replaces deployment:persona-prefix / deployment:persona-suffix. Empty means nothing is injected. Both change the prompt prefix and therefore cache reuse; the panel says so.

A policy applies at agent/created — only to sessions created afterwards. A running conversation is never rewritten. The panel states whether the current session was created with the policy.

Control plane, not a replay

dsh-context records what already happened — per-request composition, file activity, command executions, timing, cost, the multi-agent fleet. It is the best tool for "why did the last request cost that".

This plugin answers the question before the request exists: what will the next one carry, what can I switch off, and what does it look like after that? It never writes to your workspace, every switch is reversible, and every number either comes from the host's own measurement or says it is an estimate. Install both if you want the full picture; they do not modify each other.

dsh-context dsh-injection-manager
Direction After the fact: requests, events, cost Before: the next session's first turn
Core action Observe, expand, filter, compare Observe + switch, preset, budget, undoable
Granularity Per request (per-step composition, files, commands) Per session's first-turn injection (5 channels + suites)
Output Charts and detail A decision: which suite to drop, what it saves, what the window does
Boundary Multi-agent, files, commands, billing Only the injection catalog: prompt sections, contexts, skills, tools, rules
License Apache-2.0 MIT

Five entry points

Entry Where What
Sidebar icon Left panel list, "Injection" Opens the panel: 管理 / 洞察 / 设置
Manage tab Main area Health, KPI, request strip, suite list, switches, content drawer
Composer chip Under the input First-turn total and window occupancy, deliberately muted; × hides it, Settings can bring it back
context_injection tool Model-callable Read-only snapshot with a Markdown card; optional {channel, name} for one item's content (≤32 KB)
/inject command Composer Text report + one health line; /inject health, /inject minimal, /inject all, /inject budget 20000, /inject apply [20000], /inject undo

The model-facing tool is read-only: the model cannot loosen or tighten the policy. Every write comes from a click in the panel.

Install

dsh plugin add github:IWHBIWHB/dsh-injection-manager

Or locally: DSH → Plugins → Add plugin → pick this directory, or dsh plugin add D:/path/to/dsh-injection-manager. Refresh the page; the "Injection" icon appears in the left panel list.

Upgrading from dsh-context-preflight / dsh-context-injection: if only an old policy file exists, it is read once and used as the current policy. Writes always go to $DSH_HOME/injection-manager/policy.json.

Uninstall: remove the plugin like any other; the policy, stats and event files stay in $DSH_HOME/injection-manager/ (delete that directory to start clean).

Version support and required APIs: docs/COMPATIBILITY.md.

Plugin configuration

Plugins → dsh-injection-manager → Configuration exposes four presentation and first-run options:

Setting Default Effect
dock on Show the "injection 84k · 97% switchable" chip next to the composer
refreshOnFocus on Refresh the panel when the window regains focus
showRawNames on Show the raw key (harness:identity) next to the readable label
budgetTokens 0 Initial budget before a policy file exists; 0 = none

In-panel preferences (Settings tab) are browser-local: content-drawer line numbers and the default insights range. They are never written to the policy file, so there is exactly one source of truth for switches.

Data & privacy

Path Content Written by
$DSH_HOME/injection-manager/policy.json Channel modes, names, prefix/suffix, budget Panel actions (atomic write)
$DSH_HOME/injection-manager/stats.json One measured row per session, with a per-channel token split Every snapshot the panel/tool/command builds
$DSH_HOME/injection-manager/events.json Append-only event feed, max 300, atomic write Policy saves, presets, budget edits, imports, undos, compactions
Plugin config row dock / refreshOnFocus / showRawNames / budgetTokens Plugins → Configuration
Browser local storage Drawer line numbers, insights range, dismissed cards This browser only

No network requests, no telemetry, no uploads. The event feed stores only policy and session metadata: strings are flattened and truncated to 200 characters and the session's prompt text is never written. The content drawer reads files from disk only when you open it, only for paths the snapshot registered, and nothing read there is persisted.

Self-check

The plugin exposes 15 routes under /plugins/dsh-injection-manager: 12 read-only GETs, plus three POST-only endpoints — /action changes the policy, and /boot + /registrations are client self-reports (a GET on those two returns 405):

Route What it proves
/ping The plugin registered its routes
/summary?session=<id> The snapshot the panel renders (apiVersion: 2, healthSummary, item previews)
/content?channel=<id>&id=<itemId>[&session=<id>] Original content behind one row (UTF-8 safe, 256 KB cap, 2 s budget)
/trend?range=all|30d|7d&limit=120 Cross-session first-turn trend + aggregate + coverage + backfilling
/events?session=<id>&limit=60 The session's event feed with its window
/health?session=<id> The same checks the health strip shows
/stats The raw history file plus coverage/backfill state
/verify?session=<id> Re-assembles a live session and returns the effective sections, contexts and tool names
/diag Client module presence, last application result, how many times the assemble filter ran
/policy、/export、/locale Policy file state, exports, dictionaries
POST /boot、POST /registrations Client self-reports (in memory only; GET returns 405)
POST /action The one policy write path: switch, preset, budget, import, undo (same-origin JSON only)

The 设置 tab exposes the same verification behind a button.

Docs

  • Detail: the five-minute version of what is measured, what is switchable, and what the new surfaces show
  • Architecture: the four injection channels, the snapshot pipeline, the routes and the data files
  • FAQ: why AGENTS.md cannot be switched off, unreachable budgets, what the content drawer reads, relation to dsh-context
  • Compatibility: tested versions, required APIs, upgrade checks
  • Performance: measured timings for 355 items, 27 suites, a 230-row expand and the v2 surfaces
  • Security: reporting, verifiable boundaries, accepted risks
  • Contributing: local loop, pre-commit checks, release steps (npm run smoke runs the pipeline offline)

Limitations

  • 0.2.2 has no ordering API for contributions; the budget is a greedy largest-first heuristic, not a global optimum.
  • The skill catalog is a pre-step message, so granularity is per-skill (or the whole catalog via the skill tool).
  • User-message channels such as workspace rules and time context cannot be switched off; the plugin shows them (and their content) instead of pretending otherwise.
  • The insights trend only contains sessions this plugin measured or priced from durable logs; older rows have no channel split and render striped.
  • UI ships in Simplified Chinese, English, Japanese, Spanish, French, German, Russian and Portuguese; adding a language is one JSON file in locale/.

License

MIT

Content from the project README on GitHub ↗

Links

More in this category

View the whole category →

Community comments

Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.