e7738de6d2
CI / Deep Native Runtime Cases (1/6) (push) Has been skipped
CI / Native Preflight (push) Failing after 1s
CI / Native Runtime Cases (1/2) (push) Failing after 0s
CI / Native Runtime Cases (2/2) (push) Failing after 1s
CI / Native Metadata Reports (push) Failing after 0s
CI / Native Direct Backend Artifacts (push) Failing after 0s
CI / Native Sanitizer Smoke (push) Failing after 1s
CI / Command Contract Snapshots (push) Failing after 1s
CI / Deep Conformance Suite (push) Has been skipped
CI / Graph Build Perf (push) Failing after 1s
CI / Deep Native Preflight (push) Has been skipped
CI / Deep Native Runtime Cases (2/6) (push) Has been skipped
CI / Deep Native Runtime Cases (3/6) (push) Has been skipped
CI / Conformance Suite (push) Failing after 1s
CI / Workspace Checks (push) Failing after 0s
CI / Deep Native Runtime Cases (5/6) (push) Has been skipped
CI / Deep Native Runtime Cases (6/6) (push) Has been skipped
CI / Deep Native Runtime Cases (4/6) (push) Has been skipped
CI / Deep Graph Build Perf (push) Has been skipped
95 lines
2.8 KiB
Markdown
95 lines
2.8 KiB
Markdown
## Semantic Edits
|
||
|
||
In Zerolang, text is an excellent human review format. It is not an ideal agent edit format.
|
||
|
||
When an agent edits text, it must keep many details in its context at once:
|
||
syntax, indentation, imports, symbol names, surrounding control flow, stale
|
||
file contents, and the compiler's current capabilities. A small mistake can
|
||
produce invalid code that is only discovered after another tool call.
|
||
|
||
Graph edits carry more intent:
|
||
|
||
```text
|
||
replaceFunctionBody main
|
||
check world.out.write "hello\n"
|
||
end
|
||
```
|
||
|
||
That patch says which function body changes and what rows should become the new
|
||
body. The compiler parses those rows, builds graph nodes, checks graph shape,
|
||
updates repository metadata, and can export a stable projection for humans.
|
||
|
||
```json-render
|
||
{
|
||
"messages": [
|
||
{
|
||
"role": "user",
|
||
"text": "add a help command and then show me the source"
|
||
},
|
||
{
|
||
"role": "assistant",
|
||
"text": "I’ll add the help branch, run it, and then show the projection."
|
||
},
|
||
{
|
||
"role": "tools",
|
||
"calls": [
|
||
{
|
||
"command": "zero patch --check-only /tmp/help.patch",
|
||
"output": "program graph patch ok (check-only)"
|
||
},
|
||
{
|
||
"command": "zero patch /tmp/help.patch",
|
||
"output": "program graph patch ok"
|
||
},
|
||
{
|
||
"command": "zero run -- help",
|
||
"output": "usage: zero run -- <command>"
|
||
},
|
||
{
|
||
"command": "zero export",
|
||
"output": "repository graph export ok\nwrote: ./src/main.0"
|
||
}
|
||
]
|
||
}
|
||
]
|
||
}
|
||
```
|
||
|
||
## Text Edit Failure Modes
|
||
|
||
Common text-edit problems for agents:
|
||
|
||
- editing the wrong overload or similarly named function
|
||
- losing an import or closing brace
|
||
- creating syntax that looks plausible but is not accepted by this compiler
|
||
- formatting code that later changes the span the agent intended to patch
|
||
- using stale file contents after another edit
|
||
- changing source while the graph store remains the actual compiler input
|
||
|
||
Zero does not remove all errors. It moves the primary edit operation closer to
|
||
the compiler's semantic model so more errors are caught at patch time.
|
||
|
||
## Graph Patch Guardrails
|
||
|
||
Graph patches can include:
|
||
|
||
- graph hash expectations
|
||
- node hash expectations
|
||
- field expectations
|
||
- typed operation names
|
||
- function body or block body replacement
|
||
- dry-run and check-only modes
|
||
|
||
Those guardrails make stale edits explicit. They also make the failure useful:
|
||
the agent can query the graph again and patch the current node rather than
|
||
guessing from text.
|
||
|
||
## Where Projections Fit
|
||
|
||
Projection text is still part of the system. It gives humans a compact review
|
||
format and a manual edit escape hatch. It also gives diagnostics stable spans
|
||
and makes examples readable.
|
||
|
||
The rule is simple: agents write the graph by default; humans review
|
||
projections by choice.
|