DeepSeek Harness Plugin

Anseya/flow-tower-context#flow-tower

Stars ★ 0 Category Memory Added 2026-11-16

Compaction-driven project memory for DeepSeek Harness: deterministic compilation of compaction deltas into a searchable cited dictionary, adaptive dual injection, and a portable compaction-takeover preset.

Install

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

dsh plugin --profile web add github:Anseya/flow-tower-context#path:/plugin/flow-tower

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

Version 0.66.4 · DeepSeek Harness >=0.2.2-rc.3 <0.3.0 · public package flow-tower-context · MIT

English · 中文 · Install · Compatibility · Publishing · Project overview · Paper (Word) · Design notes

1. What it is

An observational, automatic, project-local context-memory plugin for DeepSeek Harness (DSH). It watches DSH compaction/summary events and deterministically compiles each compaction delta into a searchable, verbatim-traceable project dictionary, then injects the long-range project picture back into the model context on demand. It does not re-implement the host's compaction transaction; it adds a searchable long-term memory layer on top of it.

  • Compaction-driven compilation — the compaction event is the only write anchor; zero extra LLM calls, no write amplification;
  • Ten entry kinds — file / api-symbol / feature / decision / task / command / error / topic / chapter / person;
  • Verbatim replay — every entry carries seq citations; dict_cite replays the original messages, not a rewrite;
  • Typed relations — parent / derivedFrom / cites / replaces / mergedFrom / sibling / relatedTo / taskContext;
  • Dual injection — auto selects dual / dual-lite / long-only / short-only from task type, checkpoint freshness and context pressure;
  • Multi-agent sharing — subagents and fresh sessions without their own dictionary volume read a shared project digest with cross-volume open tasks;
  • Project-local storage — a single SQLite file (WAL + project write lock), default <project>/.flow-tower/flow_tower.db;
  • Authentication first — the panel page, search/citation APIs and settings writes all sit behind DSH Connection admission: unauthenticated requests receive 401/403, and when the Connection service is unavailable the plugin fails closed with 503 instead of exposing project memory anonymously.

2. Quick start

2.1 Requirements

  • DeepSeek Harness 0.2.2-rc.3 (Desktop or web profile);
  • No external service is required; optional embedding and DeepSeek smart rerank are off by default.

2.2 Install

After release (npm / plugin market):

# A. Plugin market (recommended)
#   Plugins → Experimental plugins → enable Custom plugin indexes
#   Plugins → Discover → Add index source:
#   https://cdn.jsdelivr.net/npm/flow-tower-context/market/index.json
#   Search "Flow Tower Context" → install → restart DSH

# B. Install the npm package directly
plugin_manager install_bundle target=flow-tower-context@0.66.4

Local development / before release:

# A. DSH UI → Settings → Plugins → Add plugin, choose plugin/flow-tower
# B. Build the release directory first, then install it
npm run build:publish        # writes dist/flow-tower-context
plugin_manager install_bundle target=<repo>\dist\flow-tower-context

Requires DSH >=0.2.2-rc.3 <0.3.0; restart DSH after installation so the new JS module generation is loaded. See docs/INSTALL.md for verification and uninstall, and docs/COMPATIBILITY.md for the compatibility statement.

2.3 Taking over compaction at the session-preset level (optional, recommended)

DSH plugin patches only affect the host plane, while a session's compaction service is assembled inside the preset isolate group. The published bundle registers an agent preset Flow Tower (compaction memory): it mirrors the shipped cordis preset and replaces only the compaction-basic row inside its compaction group with flow-tower-context/compaction (FlowTowerCompactionEngine extends the host BasicCompactionEngine and overrides only summarize(); triggering, range selection, transactions and pruning stay unchanged).

Select that preset when creating a new session to inject project memory into the compaction request; without it the plugin works in observer/injection mode. The preset plugin list is a snapshot of the shipped cordis preset and must be regenerated after a DSH upgrade — see docs/COMPATIBILITY.md.

2.4 Verify the installation

# Status route (port from the GUI URL; note the /flow-tower prefix)
Invoke-RestMethod http://127.0.0.1:<port>/flow-tower/api/status | ConvertTo-Json -Depth 5

Key fields: version, compactionBackend, lastSummarize.memoryChars, lastInjection.mode, compactionDiagnostics.sessionEngineVerified. A successful takeover shows memoryChars > 0 and sessionEngineVerified = true. Panel: Plugins → Flow Tower (dictionary / timeline / graph / tasks / settings).

3. Core mechanisms

3.1 Compaction-driven compilation

The plugin observes compaction/start|summary|end on session/event, reads the original messages and summary for the shadowedSeqs range of compaction/summary, and extracts entries, relations and citations. Updates are incremental, deduplicated and archived, with a maintenance pass (archive cited summary noise, attribute modules). Compilation is deterministic and linear in the delta size.

3.2 Dual injection

After compaction two information sources coexist: the short-range <compacted-summary> checkpoint (authoritative, precise) and the long-range digest/project map (overview, index). Three signals select the mode:

Mode Injected Trigger (injectionMode: auto) Budget factor
long-only full digest / map no checkpoint in context 1.0
dual-lite goals / open tasks / modules / traversal only, recent sections dropped fresh checkpoint (≤2 turns) and coding/research/history task 0.6
dual full digest planning task; or stale checkpoint (>2 turns) 1.0
short-only nothing; native checkpoint only general chat with a fresh checkpoint; or pressure ≥0.85 0

The budget adapts to pressure (base 2000 tokens, scaled at 0.55/0.7/0.85, floor 400), and the header carries a machine-readable marker, e.g.: [flow-tower] Project map · project · 10 open / 20 done · inject: dual · short: checkpoint 1t ago · task: coding

The injection budget is per compaction cycle (v0.66.3): a compaction replaces the durable injected messages, so the per-volume counter is reset on compaction/summary; a long session never silently stops injecting after 40 injections.

3.3 Retrieval

  • Default hybrid: lexical-first (CJK bigrams + ASCII tokens + path segments, headword ×2, alias ×3, bilingual synonym expansion) with sparse-semantic fill, avoiding pure rank fusion pushing correct hits down;
  • /api/search and dict_search support mode (hybrid/lexical/semantic), kind, module, status, since/until, sort;
  • The panel's "smart rerank" optionally calls DeepSeek for query rewriting + candidate reranking (results cached for 10 minutes); embedding is optional and off by default.

3.4 Storage and evidence

  • Project-local SQLite: entries, relations, citations, summary tree, groups, metrics, injection cursor;
  • Run log compaction-runs.jsonl (engine line with memoryChars/injectedMessages, event line with compactionId) and engine debug log $DSH_HOME/flow-tower/compaction-engine-debug.log;
  • The status route merges both lines and exposes compactionDiagnostics (engine log summary, sessionEngineVerified).

4. Configuration defaults (excerpt)

Key Default Meaning
storageMode project global / custom also supported
maxEntries 6000 entries per volume, rolling eviction
digestMode on-change inject only when the dictionary version changes
digestStrategy brief large projects switch to the project map
injectionMode auto dual / off also supported
digestBudgetTokens 2000 further scaled by pressure
summaryMemoryChars 1800 project-memory cap inside the compaction request
mapThresholdEntries 200 switch brief → module map above this size
relationsMode on typed relations
llmCompiler off shadow / on (proposals need seq-citation validation)
proposalMode off LLM proposal pipeline

Three layers: bundle defaults → $DSH_HOME/flow-tower/settings.json → <project>/.flow-tower/settings.json. See docs/injection-policy.md and the panel settings page.

5. Tools and panel

Six model-facing tools: dict_browse, dict_lookup, dict_cite, dict_timeline, dict_search, dict_relate (plus dict_shadow_probe when llmCompiler != off). The read-only panel has Browse / Tree / Timeline / Tasks / Graph / Settings; the HTTP API under /flow-tower/api/* is read-only; an optional stdio JSON-RPC MCP server lives in tools/mcp-server.mjs.

6. Evaluation and tests

cd plugin/flow-tower
npm test                  # 85/85 tests + module-graph --check (68 modules / 0 cycles)
npm run eval:search       # recall: lexical/hybrid Recall@5 = 0.94, sparse semantic 0.66
npm run eval:cost         # context cost: 5.26M chars of tool output, pruner bound ≈315k tokens, avg injection 767 tokens
node eval/txt-corpus-eval.mjs --file=<corpus.txt>   # long-text stress (8.44M-char novel → 2,846 entries, 100% fidelity)

Other scripts: eval/offline-eval.mjs, eval/real-session-eval.mjs, eval/long-session-eval.mjs, eval/relations-bench.mjs, eval/ci.mjs. Paper numbers come from a real development session (8 compactions, 3,881 entries, 23,676 active relations) and the evaluations above.

Release tooling:

npm run build:publish          # build dist/flow-tower-context (public manifest, LICENSE, bundle patch, index, DETAIL)
npm run publish:check          # pre-release checklist (name/engines/locale/market index/absolute paths)
npm run publish:check:strict   # also fail on the Anseya placeholder

7. Known limits

  • Weak semantic recall: lexical + synonyms + sparse semantics only; cross-language/paraphrase recall is limited without optional smart rerank or embeddings;
  • No compaction, no compilation: short sessions without compaction produce no memory;
  • Heuristic task/decision extraction may produce long or fragmented entries; mitigated by quality gates, maintenance and cite replay;
  • Single-writer storage (file lock + WAL), no multi-master merge;
  • The preset-level override is a static snapshot and must be regenerated after a DSH upgrade;
  • Engineering debt: LOC exemptions for store.js, client.js, web-page-script.js; no browser automation tests yet.

8. Links, paper and license

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.