TECH Signal 502
CHAP stores human overrides of agent outputs as queryable, content-hashed envelopes with diffs and rationales
CHAP is a protocol that records every human decision on an agent's draft, approval, edit, or rejection, as a structured envelope containing a diff, rationale, and tags, chained by content hash so the full history can be queried instead of reconstructed from scattered logs.
When an agent drafts a code review, triages a ticket, or reviews a contract and a human edits or rejects it, that decision normally lives in chat threads, ticket comments, or someone's memory. CHAP gives teams one queryable store for those decisions, which means post-incident reconstruction stops being forty-five minutes of guesswork across four UIs. Only one feed carried this, so adoption and real-world validation are not yet established.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Agent outputs are artefacts and human edits are structured overrides carrying a JSON Patch diff, a rationale string, and caller-defined tags.
Envelopes chain by content hash, so the full decision history for a task can be queried from a SQLite store rather than grepped across multiple systems.
TypeScript and Python SDKs are available, with TypeScript exposing a typed facade and Python exposing the raw JSON-RPC envelope shape.
THE READ
What the cluster adds up to.
CHAP addresses a gap that teams using coding agents already hit: when a bot drafts a code review or triages a ticket and a human overrides part of it, the override is ephemeral. It lives in a chat message, a ticket comment, or the human's head. Six weeks later, reconstructing why a decision was made requires grepping across multiple UIs and is, in the project's own words, half guesswork. CHAP replaces that with a single envelope format that captures the agent's draft as an artefact and the human's edit as a structured override.
The override envelope is the core unit. It contains a JSON Patch diff (operations like replace at a specific path), a free-text rationale, and caller-defined tags. In the documented scenario, a developer disagrees with a bot's severity warning on a pull request and overrides it with a diff that changes the severity from a higher level to "info," a rationale explaining it as a false positive, and tags like "false-positive" and "framework-pattern-misread." The envelope also carries an intent_preserved flag. Envelopes chain together by content hash, so the entire decision history for a workspace is queryable rather than linear-log-replayable.
The implementation ships with a coordinator that uses SQLite for persistence and supports TypeScript and Python. TypeScript provides a typed facade (coord.api.*) with compile-time checks and autocomplete on every method. Python keeps the JSON-RPC envelope shape visible, with consumers expected to write their own dispatch helpers. Workspaces are created with named profiles such as core/1.0 and review/1.0, and participants join as either human or agent types. The repository also includes a paper and twelve scenarios beyond the pull-request review demo.
What CHAP does not cover, based on the material provided, is the integration layer itself. The code samples show how to emit envelopes from an existing Cursor integration, but the actual wiring to a specific agent platform is left to the consumer. The protocol also appears to be a single-project proposal carried by one feed, with no corroboration from other sources. There is no indication of production deployment, community adoption, or interoperability with other agent frameworks. The profiles (core/1.0, review/1.0) suggest the protocol is designed for extension, but no additional profiles are documented in the material provided.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER