Wave table pending
The static fallback stays quiet until generated status data is present.
Local binary document infrastructure for live agents
Docpipe
Give every document the right shape for the next step.
Docpipe is a local document layer for agents. It turns heavy PDFs, Markdown, and HTML into smaller working packages while keeping exact facts, source locations, and the original available when you need to check them.
after download
docpipe skill install --all
--
Local A/B token cut
--
Exact facts kept sharp
54.2%
PDF prompt-light cut
docpipe - agent session
$ docpipe pack ./2607.12056.pdf --profile prompt_light --out .docpipe-demo
document_id: pdf_ccd97b569aabb51f
payload: 86,999 tokens, 54.2% smaller
dense pages: 5 / 11 visual gist pages
factsheet: 176 exact facts plus recoverable text
agent handoff return document_id, package path, and input sha256 upstream.Generated build status
The demo keeps its surface tied to real CLI output.
Status data refreshes from the current build before publish.
- profilesList supported output and model profiles.
- docpipe-cli
- rust toolchain pending
- Baseline data loads from eval/ab/baseline.json.
| Wave | Gate | Status |
|---|---|---|
| local | pending | Generate status.json for the live table. |
Classic application download
Install the binary; let the skill wire itself into agents.
The product path is an install file plus a prebuilt local binary. Source checkout is not part of the demo acquisition story.
Install once, then let the skill route documents.
The binary installs the agent skill so harness instructions and CLI behavior stay version-locked.
- No network in the deterministic path.
- Exact values cite source blocks and pages.
- Lossy dense mode is blocked until model verification exists.
- Handoffs return document_id, package path, and input hash upstream.
1
release path Download a checksummed package
Ship artifacts resolve to a platform binary plus checksum manifest. Code signing remains a separate release gate.
download docpipe for macOS or Linux
2
Install agent surfaces from the binary
docpipe skill install --all
3
Run a document through the local package path
docpipe pack ./2607.12056.pdf --profile prompt_light --out .docpipe-demo
The agent receives a smaller prompt payload plus exact recoverable facts.
1
pending Installer endpoint
download install file
The install file detects OS and architecture, fetches the matching binary, and verifies release checksums.
2
pending Updates and health check
docpipe self-update && docpipe doctor
Updates stay explicit: no silent network calls in the document conversion path.
How to use Docpipe
Three steps from a heavy file to useful context.
Docpipe stays out of the way until a document is large, visual, or exact enough to need a safer representation.
-
01
Install the local layer
Download the binary once, then install the skill for your agent surface. The deterministic path stays local.
-
02
Pack the document
Point Docpipe at a PDF, Markdown, or HTML file. It chooses a representation based on size, structure, and exactness risk.
-
03
Work small, recover when needed
Give the agent the compact package. Exact facts cite their source, and recoverable handles lead back to the original.
Invisible modular routing
Agents should call Docpipe only when the document asks for it.
Docpipe is a quiet module in the coordinator loop: inspect the input, package the right representation, then return compact context with enough provenance to trust it.
PDFs and provenance
Invoke for PDFs, page-aware extraction, source maps, deterministic packages, or any workflow that needs a re-resolvable artifact.
Facts that must survive
Invoke when IDs, hashes, amounts, dates, citations, clauses, or quotes need to be preserved outside the agent's fuzzy working memory.
Large doc-shaped text
Invoke above the threshold for long Markdown, reports, specs, and handoffs. Skip tiny files and code-like edits that are cheaper to read raw.
Return the package, not a paste
After large outputs are written, return document_id, package path, and input hash so the coordinator can continue with less context.
Markdown and text gate
| < 8 KB | Read raw for token savings; packaging overhead usually loses. |
|---|---|
| 8-16 KB | Invoke only for doc-shaped structure or exact-token work. |
| > 16 KB | Invoke for long documents; local large-doc rows show 80%+ savings. |
Dense PNG pages are visual gist, never exact OCR. Exact text stays in factsheet and recoverable records.
Frugal PDF result
2607.12056.pdf as a prompt-light package.
The intended payload is much smaller than the source text, while the full archival package stays available off-prompt for recovery and audit.
PDF source
189,855
tokens from docpipe pack
Prompt-light payload
86,999
54.2% smaller
Dense pages
5 / 11
176 exact facts
102,856 tokens saved by the intended prompt-light payload.
75.3% savings if dense pages are counted alone, but they remain gist context only.
Package pdf_ccd97b569aabb51f / 343e22009b479259.
Dense PNGs are visual gist, not exact OCR. Exact tokens stay in factsheet and recoverable records.
Threshold evidence
Old A/B rows become routing guidance.
The backtest is useful because it shows both sides: large document-shaped inputs pay off, while small fixture batches can lose tokens.
--
estimated token reduction across the local presentation corpus, with the negative small-file row kept visible.
Loading backtest caveat...
Three-channel routing story
Exact facts do not ride in images.
IDs, hashes, quotes, clauses
High and critical exact-risk blocks stay text-first so agents can quote and compare safely.
Bulky background context
Wave 3 Only verified model profiles can receive dense visual pages, and only as gist context.
Recoverable originals
Wave 3 Every compressed block can point back to exact original text when the task needs it.
Profile picker copy
Plain-language outcomes, not config forms.
Each package carries a stable document ID, source map, and package path so another agent or later step can pick up without re-reading the whole source.
Give me Markdown
Live now Clean Markdown plus IR, manifest, diagnostics, and source map. Markdown input only today.
Build a prompt
Wave 2-3 Ready-to-paste bundle: instructions, contents, exact facts, excerpts. Exact-risk content stays text.
Token-light package
Wave 3 Smallest possible model input. Lossy by design; requires measured model verification.
RAG assets
Wave 3+ Chunks, metadata, source maps, and images for retrieval. Inferred summaries are marked.
Audit
Wave 3+ Full internals, diagnostics, and metrics for QA, legal, and regression review.
Download status
Start with the local document layer.
Download the mirrored CLI archive, install the agent skill, and let large documents route through compact packages with exact facts and recovery paths intact.
Remaining release hardening: code signing, minisign public key, install-file verification, and explicit update commands.