dsh-project-portfolio

🚨 Unofficial community plugin. Not affiliated with, endorsed, or reviewed by DeepSeek. "DSH" appears in the name only as the community-recommended abbreviation for the harness this plugs into. 🚨 非官方社区插件：与 DeepSeek 无关联，未经其审核或背书。名称中的「DSH」仅为社区建议使用的缩写。

Long-term, cross-project memory for coding agents: one plain markdown file per project plus a global watchlist, owned by the user, maintained by the agent on request. Migrated from my skill repo mc856/project-portfolio onto DSH's plugin seams.

Where this sits (read before installing)

DSH already ships the general capability this builds on. The official skill subsystem — dsh-skill (registry), dsh-skill-filesystem (local skill directories), dsh-tool-skill (model-facing loader) — means that if all you want is these instructions, you can drop the original skill file into a skills directory today and stop there. This plugin is the packaged step past that, and everything it does goes through documented seams:

• Embedded skill via ctx.skills.register(...) — the official mechanism for plugins to contribute an in-memory skill. One install brings the instructions and the four project templates (exposed as the skill's resource base directory), and they update with the package.
• Two registered tools via ctx.tools — portfolio_status and portfolio_log, which turn the ledger's bookkeeping discipline (dated history lines, last_checked refreshes, path-safe file addressing) into validated tool calls instead of hoped-for freeform edits.

Neighbors — same neighborhood, different squares. dsh-memoir is automatic cross-session conversation memory (BM25 recall over what was said). dsh-projects organizes sessions and folders inside the DSH UI. This plugin is neither: it is a user-owned ledger of goals, next steps, follow-up windows and history in markdown you can grep, edit, version, and take with you — nothing is recorded automatically; the agent maintains it deliberately, through tools, on your instruction. If your need is "remember what we talked about", use dsh-memoir; if it is "don't lose the thread of my job hunt / side projects / PRs across months", that is this square.

What you get

Piece	What it does
project-portfolio skill	The full workflow: template routing, evidence-first bootstrapping, watchlist-driven incremental sync, priority ranking, discipline rules. Appears in the skill catalog like any other skill.
portfolio_status tool	Read-only, one call: full _watchlist.md + a headline inventory of every project file (first heading, Status/Goal/Next step/last_checked lines, last-modified date).
portfolio_log tool	Appends one dated history line to a project file's History section (creates the section if missing); optionally refreshes last_checked. Slug-validated so it can only address files inside the portfolio directory.
templates/	The four project templates (generic, GitHub PRs/issues, job hunt, side projects) plus the watchlist template, shipped as the skill's resource base.

Install

Into a profile, as a bundle (the package declares dsh.bundle.patch, so the plugin mounts automatically):

dsh plugin add dsh-project-portfolio --profile web

Verified end-to-end on 0.1.1-rc.2: install pins the package into the profile, and the next dsh web boot registers the skill and both tools with the packaged templates resolving from the installed location.

Or mount it for a single run with a patch overlay (note: --patch is a launcher flag — keep it before web-app flags like --port):

# my.patch.yml
- insert:
    - id: project-portfolio
      name: 'dsh-project-portfolio'   # or an absolute path to lib/index.js for a checkout
      config:
        portfolioDir: '~/project-portfolio'

dsh web --patch ./my.patch.yml

Configuration

Field	Default	Meaning
enabled	true	Mount switch.
portfolioDir	~/project-portfolio	Absolute path (or ~/-prefixed) to the portfolio directory. Created content is plain markdown; the directory is yours.

Compatibility

Tested by loading into dsh web on 0.1.1-rc.2 and 0.1.2-alpha.2 (2026-08-31: boot, skill catalog listing, skill content load, tool registration — zero errors on both), and exercised end-to-end on 0.1.1-rc.2 against the official DeepSeek API (deepseek-chat, 2026-08-31): the skill auto-loads from the catalog on a plain "where do my projects stand?" request, and both tools execute in real agent rounds with correct on-disk results. The optional-service degradation (skills not mounted → pure tool plugin) is by design; the tools service is required.

DSH is a v0.1 developer preview whose documentation states that breaking changes will happen. Any DSH release may break this plugin. Profiles pin plugin versions, so upgrading the host does not silently upgrade this plugin — after a host upgrade, expect to retest before trusting it.

Since v0.1.1 the package ships zero runtime dependencies, and that is deliberate, not minimalism for its own sake: the host mounts its own packages (including the tool registry) as by-name plugin entries and resolves module names from the profile directory. A plugin that lists a host package in dependencies therefore installs a second copy into the profile, which can shadow the host's own service resolution and split module-scoped symbol identity — v0.1.0 had exactly this bug (dsh plugin add installs crashed at the first tool call), found and fixed the same day. If you are writing your own plugin: keep host packages out of dependencies.

Exit clause / 退场条款: this plugin exists to cover one square the official skill subsystem leaves to the ecosystem — packaged, tool-backed distribution of one specific skill. If an official capability (or a clearly better-established community standard) covers this square, this repo yields: it will be archived with a pointer to the successor rather than competing with it.

Data flow & secrets

Read this section before pointing the plugin at real data.

• The plugin makes no network calls and collects no telemetry. Everything happens on the local filesystem.
• On the host side it touches only ctx.skills and ctx.tools. It never reads or writes anything under $DSH_HOME or any other DSH-internal state.
• On disk it reads and writes only the configured portfolio directory. portfolio_log accepts a single-segment slug ([A-Za-z0-9_][A-Za-z0-9._-]*); path separators and dot-prefixed names (..) are rejected, so tool calls cannot be steered outside the directory.
• Everything portfolio_status returns enters the model context — which means it is sent to whichever model provider your DSH profile is configured to use. Treat the portfolio directory accordingly:
  • never keep API keys, tokens, passwords, or other credentials in portfolio files;
  • a portfolio attracts exactly the kind of content you may not want leaving your machine — salary numbers, recruiter contacts, offer terms, personal goals. Decide deliberately what belongs in these files, and redact anything you would not paste into a chat with your provider;
  • the reader caps how much it ingests (_watchlist.md at 32 KB, per-file headline scan at 64 KB), but a size cap is not redaction — it bounds volume, not sensitivity.

Relation to the original skill repo

mc856/project-portfolio remains the harness-agnostic source (Claude Code / Codex / Cursor installs). This repo is its DSH-native form: same workflow and templates, plus the two tools that only make sense where a tool registry exists. Workflow fixes land in the original first and are ported here.

Deferred (known, deliberately not in v0.1)

• Session-end auto-logging via the session-query seam — "log where this session left off" without being asked. The seam exists; wiring it responsibly (what to log, what never to log) needs more thought than this version wanted to rush.
• Parsing watchlist triggers in code (the skill currently owns that reasoning).
• An evidence-report tool (compiling impact one-liners across completed records).

License

MIT