mirrors/z3
3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2026-05-15 14:45:33 +00:00
Code Activity

Add action/expectation/result structure to all skill definitions

Browse source 
Each step in every SKILL.md now carries labeled Action, Expectation,
and Result blocks so the agent can mechanically execute, verify, and
branch at each stage. Format chosen after comparing three variants
(indented blocks, inline keywords, tables) on a prove-validity
simulation; indented blocks scored highest on routing completeness
and checkability.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This commit is contained in:
Angelica Moreira 2026-03-11 19:51:59 +00:00
parent d349b93d1d
commit 9d674404c8
10 changed files with 364 additions and 48 deletions
Download patch file Download diff file Expand all files Collapse all files

41
.github/skills/solve/SKILL.md vendored
View file

@ -7,12 +7,31 @@ Given an SMT-LIB2 formula (or a set of constraints described in natural language
# Step 1: Prepare the formula
If the input is already valid SMT-LIB2, use it directly. If it is a natural language description, use the **encode** skill first to produce SMT-LIB2.
Action:
Convert the input to valid SMT-LIB2. If the input is natural language,
use the **encode** skill first.
The formula must include `(check-sat)` at the end. Append `(get-model)` for satisfiable queries or `(get-unsat-core)` when named assertions are used.
Expectation:
A syntactically valid SMT-LIB2 formula ending with `(check-sat)` and
either `(get-model)` or `(get-unsat-core)` as appropriate.
Result:
If valid SMT-LIB2 is ready, proceed to Step 2.
If encoding is needed, run **encode** first and return here.
# Step 2: Run Z3
Action:
Invoke solve.py with the formula string or file path.
Expectation:
The script pipes the formula to `z3 -in`, logs the run to
`.z3-agent/z3agent.db`, and prints the result.
Result:
Output is one of `sat`, `unsat`, `unknown`, or `timeout`.
Proceed to Step 3 to interpret.
```bash
python3 scripts/solve.py --formula "(declare-const x Int)(assert (> x 0))(check-sat)(get-model)"
```
@ -27,14 +46,20 @@ With debug tracing:
python3 scripts/solve.py --file problem.smt2 --debug
```
The script pipes the formula to `z3 -in` via subprocess (no shell expansion), logs the run to `.z3-agent/z3agent.db`, and prints the result.
# Step 3: Interpret the output
- `sat` followed by a model: the formula is satisfiable; the model assigns concrete values to each declared constant.
- `unsat`: no assignment exists. If `(get-unsat-core)` was used, the conflicting named assertions are listed.
- `unknown`: Z3 could not decide within the timeout. Consider increasing the timeout or simplifying the formula.
- `timeout`: the process was killed after the deadline. Try the **simplify** skill to reduce complexity.
Action:
Parse the Z3 output to determine satisfiability and extract any model
or unsat core.
Expectation:
`sat` with a model, `unsat` optionally with a core, `unknown`, or
`timeout`.
Result:
On `sat`: report the model to the user.
On `unsat`: report the core if available.
On `unknown`/`timeout`: try **simplify** or increase the timeout.
# Parameters