dsh-web-missions

Persist-and-find web research for local coding agents in the DeepSeek Harness web UI.

npm package and GitHub repo share the name dsh-web-missions (janpauldahlke/dsh-web-missions).

A mission is a small, budgeted web-fetch unit: the agent pulls a handful of pages through a local browser39 headless browser, every fetched page becomes a durable artifact in a per-project store, and the agent finds it back lexically later — across sessions. Budget, an egress allowlist, and a depth fence bound the work; cookie-banner consent is handled deterministically by the host, not the model. No cloud search, no embeddings, no vector store, no LLM calls inside the plugin.

Related plugins (same dual-face dsh.bundle shape for the web rightbar):

Plugin	Repo
Bounded-file agent memory	dsh-memento
Long-horizon task status	dsh-local-long-horizon
Process & port lifecycle	dsh-agent-processes
NVIDIA GPU util / VRAM / power	dsh-gpu-monitor-nvml
Local LLM endpoint / slot health	dsh-slot-health
Local web discovery (SearXNG / browser39 search)	dsh-web-discovery

This package: dsh-web-missions · source janpauldahlke/dsh-web-missions.

Verified against DeepSeek Harness 0.2.1-alpha.2 (dsh web).

---

Requirements

• DeepSeek Harness web profile (dsh web).
• Node.js ≥ 20 to build.
• @aquintanar/browser39 2.0.0 (the only runtime dep) — a local headless browser whose platform binary ships as an optional dependency, so no system Chrome is required.
• Happy-path companion dsh-web-discovery (same profile, curated version — see that package’s HANDOVER.md). Provides discovery_search (SearXNG and/or browser39 search) as the opener ritual step. Without it, seed URLs + mission_fetch still work (no_engine is supported); do not expect discovery_search in the tool list. Not folded into this package.

---

Screenshots

Light theme, matching the DSH default. Full-width shots (stacked — narrow rightbar crops badly in two-column tables).

In the rightbar (chat + Missions pane):

Board (recents):

Detail (conclusion delivered):

Pages ledger (full / blocked color grammar):

---

What you see

• Web Missions rightbar — board (active XOR recent) and per-mission detail as collapsible cards (goal → conclusion → briefing → pages → budget → hints).
• Color grammar — green = conclusion delivered / page full; amber = thin or hints attention; red = blocked / budget exhausted. Status pills stay grey (lifecycle, not quality).
• Budget — remaining/max fetch_calls plus depth/host fence in the card preview; refusal gates (allowlist, depth, host-breadth) cost zero.
• Consent note — when a cookie banner was auto-accepted, the receipt says so.

---

Install

From npm (recommended)

dsh plugin --profile web add dsh-web-missions
# restart dsh web (or rely on live patch reload), then hard-refresh the browser

From GitHub

dsh plugin --profile web add github:janpauldahlke/dsh-web-missions

From a git checkout (developers)

git clone https://github.com/janpauldahlke/dsh-web-missions.git
cd dsh-web-missions
npm install && npm run build
dsh plugin --profile web add "$(pwd)"

Restart (or boot) dsh web so the host + client faces load:

env -u DSH_WEB_URL -u DSH_SHELL -u DSH_SESSION_ID dsh web --no-open

Uninstall:

dsh plugin --profile web remove dsh-web-missions
# restart the web instance that had the plugin

---

Agent tools

Tool	Purpose
mission_start	Start a mission — store dir + mission.md + recents row. Optional seed URLs (depth 0).
mission_list	Board — active + recents.
mission_status	Manifest + visited ledger (+ optional discovery if the companion recorded candidates).
mission_fetch	Fetch via local browser39 (cache hits free; force re-fetches).
mission_find	Lexical search over stored pages only — never fetches.
mission_read	Read one stored chunk (verbatim text + quality). Zero budget; text is untrusted.
mission_verify	Claim check: citations resolve to full artifacts; numbers appear in cited pages.
mission_stop	Stop the mission and run the same claim check.

Ritual: fetch → find → read → note → verify → stop. Every fetch is a receipted artifact (p-…) with page_quality, chunk_count, next_step_hint, consent flags, and budget. Blocked pages are stored too.

---

Egress, budget + depth

• Allowlist — off by default. When "allowlist": "on" in <cwd>/.dsh/web-missions/config.json, <cwd>/.dsh/allowlist.txt gates egress (one pattern per line, # comments; ** = allow all). A missing file is seeded with a template; empty patterns → allowlist_empty; a miss → allowlist_denied (message names the off-switch). Cache hits are not re-gated.
• Budget — each mission has a used/total fetch budget. Cache hits and refusal gates cost zero.
• Depth fence — maxDepth (default 2) + maxDiscoveredHosts (default 3) bound crawl breadth/depth.

Config knobs: <cwd>/.dsh/web-missions/config.json (overrides optional ~/.dsh/web-missions/config.json) — allowlist, snippetWindow, previewChars, readChars, sampleLinks, findLimit, maxDepth, maxDiscoveredHosts, defaultFetchBudget, numberFormatCheck, …

---

Privacy

• Local only — fetches run through a local browser39 headless browser and are rendered to markdown locally. No search engine, no cloud API.
• Scoped sessions — browser39 sessions live under ~/.dsh/browser39/sessions/<host>/ (persisted, so an accepted consent survives a refetch); the mission store lives under <cwd>/.dsh/web-missions/.
• Consent is deterministic — cookie-banner acceptance is host code behind the receipt (not a model judgment). Shipped vocabulary: assets/consent_profiles.toml (generic labels for every host; optional [[hosts]] only to prioritize site-unique labels). Per-project override: copy that file to <cwd>/.dsh/web-missions/consent_profiles.toml (replaces wholesale — do not grow an endless host list in the package).
• No LLM calls, no embeddings, no vector store inside the plugin.

---

Surfaces

Surface	Path / name
Mission store	<cwd>/.dsh/web-missions/<id>/ (manifest.json, mission.md, artifacts/, receipts.json, links.json, config.json)
Egress allowlist	<cwd>/.dsh/allowlist.txt
browser39 sessions	~/.dsh/browser39/sessions/<host>/
Consent profiles	assets/consent_profiles.toml (generic + optional hosts; cwd override replaces wholesale)
HTTP	GET /api/web-missions/health · /missions · /missions/:id · /sessions · POST …/stop · …/recents/clear · …/clear-all
UI	Rightbar Web Missions (board + detail)

---

Design in five lines

1. A mission is a budgeted fetch unit — seed URLs + a used/total budget; the store is the unit of persistence.
2. Every fetch is a receipted artifact — chars, page_quality, consent_*, budget — so claims stay verifiable and dead URLs are known.
3. Find is read-only — mission_find ranks the store lexically (no embeddings, no vectors) and never triggers a fetch.
4. Three fences, zero-cost refusals — budget, egress allowlist, and a depth / host-breadth fence bound the work without wasting budget.
5. Consent is host code — deterministic cookie-banner handling behind the receipt, never a model judgment.

Deliberately absent: cloud search, embeddings, vector stores, and LLM calls inside the plugin.

---

Architecture

Dual-face package (same bar as the other web rightbar plugins):

• Host (lib/index.js, ESM) — mission store, browser39 fetch, consent tactic, lexical find, egress / budget / depth fences, HTTP, tools.
• Client (lib/client.js, CJS ModuleLoader factory) — the Web Missions rightbar (board + detail).
• Glue — cordis.patch.yml + dsh.bundle / dsh.client in package.json.
• Build — build.mjs calls dsh-build-tools (buildDshPlugin()); lib/ is tracked so prepare can fall back when the toolchain is unavailable.

The @aquintanar/browser39 runtime dep carries the platform headless-browser binary as an optional dependency; BROWSER39_BIN overrides the binary path.

---

Develop / test

Hermetic checks live under test/ (same bar as other dsh-* plugins / dsh-build-tools). Operator boot scripts are local-only under agent/scripts/ (gitignored).

npm run build
npm run smoke              # tool-face + untrusted + pane + r4 + conclusion-md
npm run smoke:consent      # consent profiles + tactic
npm run smoke:allowlist    # egress allowlist

---

License

MIT — see LICENSE.