dsh-context-imports

---

Why

	Without this plugin	With it
AGENTS.md itself	Injected by the stock agent-instructions plugin	Still injected by it, no duplication
@docs/architecture.md references	Plain text, ignored	Full file content in context
Nested imports	Nobody handles them	Expanded recursively, depth limited
After compaction	Referenced files are gone with the history	Re-injected automatically
Resumed sessions	May stack a second copy	Skipped when one is already present

How it works

agent/session-start (startup / resume / clear / compact) records the scene
agent/pre-step (every step) decides whether to inject
  ├─ scan the instruction files (default AGENTS.md, CLAUDE.md) for @path imports
  │    ├─ fenced code blocks are skipped; the path must have a file extension
  │    ├─ relative paths resolve against the importing file's directory
  │    └─ duplicates collapse by resolved path, so import cycles are harmless
  ├─ merge the explicit files list (injected in full; instruction files are
  │    scanned but never re-injected)
  ├─ budgets: 64 KB per file (truncated with a note), 128 KB total (skipped
  │    with a note); a missing file becomes a one-line note instead of an error
  ├─ at most one injection stays in the active history: the full event log is
  │    fingerprint-scanned, and compaction that removes the block re-enables it
  └─ the bundle is appended to the end of the step messages, after AGENTS.md
       and every other context injection

The injected message is a <system-reminder> containing <file path="..."> blocks and nothing else. Set template if you want your own framing around it.

Features

• Expands @path imports recursively. Default depth is 3; import cycles and duplicates are collapsed.
• Follows the imports of every instruction file the harness loaded — including a project's AGENTS.md it picks up mid-session once the agent touches files under it.
• Triggers on new sessions, resume, clear, and after compaction. Each scene can be turned off.
• Never sends the same file twice: a file already in the model's context is left alone, and a newly loaded instruction file adds exactly its own imported files.
• Six-language settings card (zh, en, ja, fr, ru, ko) under Settings, Plugins, Plugin configuration. Every option is editable there and applies without a reload.
• Byte budgets per file and in total. Missing files are reported, never fatal.
• Uses the official API surface only: cordis events, createUserMessage from dsh-llm, a schemastery config, and the optional dsh-settings namespace. No feature-plugin dependencies.

Configuration

Field	Default	Description
files	[]	Extra files to inject, relative to the session working directory or absolute
scanImports	true	Scan AGENTS.md and the other instruction files for @imports
instructionFiles	["AGENTS.md", "CLAUDE.md"]	Files to scan. They are never re-injected themselves
expandLoadedInstructions	true	Also follow the imports of the instruction files the harness itself loaded, including a project AGENTS.md it brought in mid-session
maxDepth	3	How many levels of nested imports to follow
maxFileBytes	65536	Per-file cap in bytes; larger files are truncated
maxTotalBytes	131072	Total budget; files beyond it are skipped with a note
injectOn	all four	Which session-start scenes trigger injection
template	""	Custom wrapper text. {{content}} marks where the files go; empty uses the plain wrapper

Install

dsh plugin --profile web add dsh-context-imports

It works with no configuration: AGENTS.md and CLAUDE.md are scanned and their imports injected. To build from source instead:

git clone https://github.com/ciskonc/dsh-context-imports.git
cd dsh-context-imports
npm install
DSH_CHECKOUT=<path-to-your-dsh-installation> npm run build
npm run build:client

Compatibility

Plugin version	DeepSeek Harness	Notes
0.3.1+	≥ 0.2.0-rc.1（0.3.1 tested on 0.2.1-rc.2 and 0.2.1-rc.3）	Current line. Adds expandLoadedInstructions: the imports of every instruction file the harness loaded are expanded, including a project AGENTS.md it brings in mid-session. Injection is incremental — a file already in the context is never sent twice. Needs the host's published instruction scopes (agent-instructions message source + agent.inbox.nextStep), so it targets DSH ≥ 0.2.0
0.3.0	≥ 0.2.0-rc.1（tested on 0.2.1-rc.2）	Same feature, but the inbox lookup read session.inbox instead of agent.inbox and therefore never fired — the step-batch path carried the feature on its own. Fixed in 0.3.1
0.2.3	≥ 0.1.7-rc.1（tested on 0.2.0-rc.1, 0.2.0-rc.2 and 0.2.1-rc.2）	Config access hardened: on DSH ≥ 0.2.0-rc.2 the volatile config projection no longer fills schema defaults, so every field is normalized at entry (0.2.2 could crash session creation or silently skip injection there). Note: on DSH ≥ 0.2.0-rc.2 install via dsh plugin add only — the loader verifies lockfile consistency, hand-copied node_modules are silently dropped
0.2.0 – 0.2.2	≥ 0.1.7-rc.1（0.2.1 tested on 0.1.7-rc.2 and 0.2.0-rc.1; 0.2.2 tested on 0.2.0-rc.1）	0.2.2 fixed re-injection after compaction and settings-page mounting, but on DSH ≥ 0.2.0-rc.2 it can crash session creation (injectOn.includes is not a function) or silently skip injection (cfg.files is not iterable) — upgrade to 0.2.3
0.1.x	0.1.5-rc.1 – 0.1.5-rc.3（tested on 0.1.5-rc.2/rc.3）	Legacy line. Does not work on 0.1.7 — it waits on the removed settingsScope service and the removed agent/session-start event

Each release note names the DSH version it was tested against.

Development

File	Role
src/index.ts	Host side: session-start ledger, import expansion, budgets, pre-step injection, settings namespace
src/client/index.ts	Browser side: six-language dictionaries and the settings card slot
src/client/form.ts	Staged form model behind the card
src/client/ContextImportsCard.tsx	The card component

License

MIT