06 — Current rising tools and composition
Current scan of rising context tools and safe composition patterns across output, shell, retrieval, memory, and compression layers.
Summary
Composition is safe only when each tool owns a distinct interception boundary and exactly one layer controls recovery, memory, and shell rewriting. Cutoff: 2026-08-12. Inclusion requires current maintenance plus either fast attention growth, material adoption, or a mechanism distinct from the four core tools. Stars are discovery metadata, not an endorsement.
Question and scope
Which rising tools add distinct mechanisms, and under what boundary rules can multiple tools compose safely?
Method
The scan admits maintained tools with a distinct mechanism or material adoption signal, then maps overlap, recovery authority, schema cost, and failure interaction before considering composition.
Findings
Current rising set
| Tool | Current snapshot | Mechanism | Best current fit | Hard caveat |
|---|---|---|---|---|
| pxpipe | v0.13.1; 7,055★ / 25 subscribers; created 2026-05-20 | Re-encodes dense stable text/history/tool output into image pages behind a proxy | Very long, model-gated contexts where vision representation is cheaper and proven legible | Lossy modality transform; young project; current outcome evidence is first-party |
| Context Mode | v1.0.169; 19,802★ / 86; created 2026-02-23 | Keeps raw tool data in a sandbox; exposes summaries plus SQLite FTS5/BM25 retrieval | Data/research sessions that need searchable local artifacts rather than raw dumps | Elastic License 2.0; mandatory routing changes agent behavior; no current independent latest-release A/B found |
| Paritok | tag v1.3.3; 1,031★ / 78; created 2026-07-15 | 4B coding-context compressor, tool-schema semantic filtering, history summarization, recovery references | Long read-heavy or MCP-heavy sessions with local GPU capacity | Current first-party result reports 86.5% quality retained at 25.7% compression ratio—real quality loss; 24GB-class bf16 deployment |
| jCodeMunch MCP | v1.108.271; 2,542★ / 19 | Indexes code and retrieves precise symbols/relationships | Symbol-centric navigation across many languages | MCP/index overhead; current release reports its own Codex net-token benchmark as negative/inconclusive |
| Serena | v1.7.0; 27,874★ / 86 | Language-server-backed semantic retrieval and editing | Prevent broad source reads with exact symbol operations | Context engineering, not automatic compression; language-server quality varies |
| fff | v0.10.3; 9,949★ / 20 | Resident typo-tolerant path/content search with frecency and git-aware ranking | Fast repeated lexical discovery without broad file reads | Retrieval, not compression; very low subscriber depth; current task-level token effect unproven |
| codedb | v0.2.5838; 1,357★ / 9 | Indexed symbols, outlines, callers, dependencies, and task-shaped context | Structural relationship queries that lexical search cannot answer | Broad MCP/installer policy surface; current outcome evidence is first-party |
| sem | v0.21.0; 3,294★ / 6 | Entity-level diff/history/impact plus context attention ledger | Avoid unchanged entity rereads and retrieve structural history | Extremely low subscriber depth; current outcome claims remain first-party |
| mq | v0.8.2; 1,005★ / 2 | jq-like structural query/transform language for Markdown | Fetch TOCs, sections, tables, and code blocks instead of whole docs | Retrieval primitive, not automatic task optimization |
| Entroly | 435★ / 6; active 2026-08-11 | Budgeted context selection, code intelligence, recoverable receipts | Experimental local evaluation where recoverability matters | Small project; current self-benchmark discloses an accuracy drop |
| claude-context | no tagged release; 12,384★ / 57; active 2026-07-14 | Embedding/vector retrieval for code through MCP | Semantic code search when exact symbol tools are insufficient | No stable release contract; embedding model/vector-store overhead; not a compressor |
| Repomix | v1.18.0; 27,776★ / 70 | Packs a repository into one AI-friendly artifact | One-shot review, audit, archival handoff, offline prompt assembly | Packer, not adaptive per-turn optimization; can create a huge prompt |
Metadata sources: pxpipe, Context Mode, Paritok, jCodeMunch, Serena, fff, codedb, sem, mq, Entroly, claude-context, Repomix.
Which tools are genuinely different
pxpipe: change modality
pxpipe does not summarize text. It renders selected stable context into dense images and relies on a model's vision encoder. This can attack a context region text compressors treat differently, but correctness is model-specific. A legibility test with exact identifiers, tables, code, and punctuation is mandatory.
Context Mode: keep raw data outside context
Context Mode places tool execution and raw artifacts inside a sandbox, then gives the model summaries and retrieval tools. This is closer to a session-local data plane than a compressor. Its key question is behavioral: does mandatory retrieval reduce data exposure without adding searches and turns?
Paritok: spend local inference to compress remote context
Paritok uses a specialized 4B model and semantic tool-schema filtering. It can shrink a broader class of content than rules alone, but adds hardware, latency, model-quality risk, and cache questions. Its current benchmark openly reports less than full quality retention.
Retrieval tools: prevent the read
Serena, fff, codedb, sem, jCodeMunch, mq, and claude-context try to return the relevant path, symbol, relationship, entity, section, or semantic match instead of a whole file. They are often safer architectural answers because they reduce irrelevant material without lossy rewriting of selected source. Their risk moves to retrieval recall: a missed dependency can be as harmful as an over-compressed payload.
Current composition map
OUTPUT
Caveman skill
OBSERVATION / INPUT — overlapping; select one primary owner
Caveman v2 | Headroom | RTK | lean-ctx
PREVENT THE READ
Serena | fff | codedb | sem | jCodeMunch | mq | claude-context
REPRESENTATION / DATA PLANE
pxpipe | Context Mode | Paritok
ONE-SHOT PACKING
RepomixSafe composition rules
- One owner per observation. Do not pass the same shell result through RTK, Headroom, Caveman v2, and lean-ctx without an explicit experiment.
- Output is separate. Caveman's skill can compose with one input owner, but measure it independently.
- Retrieval before compression. Select the right symbol/section first; compress only a still-large result.
- Raw escalation is mandatory. Exact edits, errors, config, security, and identifiers require original or bounded authoritative input.
- Preserve attribution. Every layer needs a bypass/control mode and separate counters.
- Measure cache classes. Smaller fresh input can still lose when stable cached prefixes are rewritten.
- No star-based default. Repository attention cannot replace an accepted-task A/B.
Implications for jackin❯
Rising-tool verdict
The strongest current growth signals belong to pxpipe, Context Mode, and Paritok, each with a distinct mechanism. The mature “prevent the read” set spans semantic navigation (Serena), lexical search (fff), structural relationships (codedb/sem/jCodeMunch), and documents (mq). None has current evidence strong enough to become a universal default.
Next: 07 — Current evidence and evaluation.
Limitations and unknowns
Tool behavior, releases, licenses, and outcome evidence can change; reverify the current version and integration boundary before adoption.
Sources
Evidence and repository references are cited inline beside the claims they support.