Pounce

pounce — solver post-mortem and pre-flight

This skill wraps the pounce-studio CLI for AI-assisted post-mortem of pounce solve reports plus pre-flight inspection of .nl and .gms input files. It is the CLI-and-skill counterpart to the pounce-studio-mcp MCP server (in studio/mcp/); both backends ultimately call the same Rust analysis core, so findings agree.

Use this skill when the user asks any of:

• "diagnose this solve report"
• "did pounce converge?"
• "compare these two runs"
• "what does inf_pr mean?"
• "is this .nl file going to be a problem?"
• "what happened in iteration 47?"

Tools at your disposal

All output is pretty-printed JSON on stdout — pipe through jq to slice. If pounce-studio is not on PATH, invoke it via its absolute path (e.g. ~/.cargo/bin/pounce-studio or ~/.local/bin/pounce-studio, depending on how it was installed).

Post-mortem on a JSON solve report

pounce-studio summary <report>                # headline outcome
pounce-studio diagnose <report>               # failure-mode heuristics
pounce-studio find-stalls <report>            # stalled-progress windows
    [--min-window N] [--max-progress P]
pounce-studio convergence-trace <report>      # per-iter trajectory
    [--columns iter,inf_pr,inf_du,mu,d_norm,...]
pounce-studio get-iterate <report> <k>        # full iter k record
pounce-studio restoration-windows <report>    # restoration entry→exit cycles
pounce-studio compare <r1> <r2> ...           # side-by-side
    [--labels A,B,C]
pounce-studio linear-solver-summary <report>  # FERAL post-mortem
pounce-studio inspect <report> [--json]       # Markdown summary by default;
                                              # --json dumps the whole report

The report must be at JSON detail full for the iter-level tools (convergence-trace, get-iterate, find-stalls, restoration-windows). The summary and diagnose tools work on either detail level.

Glossary

pounce-studio explain <term>                  # column or finding code
pounce-studio citations [--topic T] [--key K] # curated papers

Known column names: iter, objective, inf_pr, inf_du, mu, d_norm, regularization, alpha_dual, alpha_primal, alpha_primal_char, ls_trials, log10_*, plus the linear-solver summary fields (n_factors, n_pattern_reuse, max_fill_ratio, …).

Known finding codes: converged, max_iter_exceeded, restoration_used, restoration_loop, mu_stuck, heavy_line_search, hessian_regularized, convergence_stall.

Pre-flight on an input file

pounce-studio analyze-nl <path.nl>            # AMPL .nl header
pounce-studio analyze-nl --builtin <name>     # builtin problem metadata
pounce-studio analyze-gms <path.gms>          # GAMS .gms header + suggestions
pounce-studio parse-gams-listing <path.lst>   # GAMS .lst SOLVE SUMMARY
pounce-studio list-gams [--suite S]           # bundled .gms instances
pounce-studio list-builtins                   # builtin --problem options

The suggestion blocks are advisory — they are never auto-applied. Decide based on the user's intent whether to pass any of them to a fresh solve.

Starting-point pre-flight (check-x0, on the pounce binary)

pounce check-x0 <path.nl> [--json]            # evaluate the model at x0
pounce check-x0 <path.nl> --x0-file cand.txt  # ... at a candidate point
pounce check-x0 --builtin <name> --json

Evaluates the model once at its starting point (no solve) and reports NaN/inf evaluations (exit 21 — a solve WOULD abort with Invalid_Number_Detected), bound violations, how far the bound_push interior clamp will move the point, initial constraint violation per row, and derivative scale spread. Run it before the first solve of a new model, after any Invalid_Number_Detected, and whenever a warm start saved no iterations.

Running a fresh solve

pounce-studio does the analysis; the pounce CLI does the solve. The skill's job is to chain them.

From a builtin or .nl file

pounce --problem rosenbrock \
    --json-output /tmp/report.json --json-detail full \
    max_iter=500 mu_strategy=adaptive

# or
pounce path/to/model.nl \
    --json-output /tmp/report.json --json-detail full \
    tol=1e-7

CLI option key=value pairs go at the end (the same format as upstream ipopt). --json-detail full is required for per-iter tools.

After the solve, immediately diagnose:

pounce-studio summary /tmp/report.json
pounce-studio diagnose /tmp/report.json

If diagnose reports convergence_stall or mu_stuck, drill in:

pounce-studio find-stalls /tmp/report.json
# pick a midpoint iter K from the window
pounce-studio get-iterate /tmp/report.json <K>

Through GAMS

When the user has a .gms file and the gams CLI installed (gams must be on PATH or set GAMS_BIN), drive it like so:

cd /a/working/dir
cp /path/to/model.gms .

cat > pounce.opt <<EOF
print_level 0
mu_strategy adaptive
tol 1e-6
json_output $(pwd)/model.report.json
json_detail full
EOF

gams model.gms NLP=POUNCE optfile=1 lo=2

GAMS writes a listing file model.lst and a log model.log. To extract the SOLVE SUMMARY block:

pounce-studio parse-gams-listing model.lst
pounce-studio summary model.report.json
pounce-studio diagnose model.report.json

Pre-flight the .gms first if the user wants option suggestions:

pounce-studio analyze-gms model.gms

Workflow recipes

"Did this solve work?"

pounce-studio summary /path/to/report.json

The status field (SolveSucceeded, MaximumIterationsExceeded, …) and final_kkt_error tell you. If the status is not SolveSucceeded, follow up with diagnose.

"Why is this solve stuck?"

pounce-studio diagnose /path/to/report.json

The findings carry codes the user can ask you to explain via pounce-studio explain <code>. Walk through them in severity order (error → warning → info).

"Compare these runs"

pounce-studio compare --labels baseline,adaptive a.json b.json

Returns one row per report with status, iter count, objective, KKT error, restoration calls, elapsed seconds.

"What does this column mean?"

pounce-studio explain inf_pr
pounce-studio explain alpha_primal_char

Returns a definition, typical numeric range, what abnormal values mean, and see_also citation keys.

"This model won't start / initialize" — the initialization playbook

Reference: the Initialization and Warm Starts chapter of the docs (docs/src/initialization.md).

1. Preflight. pounce check-x0 model.nl --json. "fatal": true means the model does not evaluate at its start — fix that first (in-domain values, or bounds that keep the clamp in the domain). Not fatal: read the warnings in order.
2. Interpret.
   • all-zeros x0 → the model supplies no start. Set values (Pyomo Var.value / GAMS x.L), or generate them: pounce.generate_starts(...) (Python), pyomo_pounce.initialize_missing_values(model) / pyomo_pounce.block_initialize(model) (Pyomo).
   • on-bound components + clamp moves on a re-solve → the previous solution is being discarded; use the warm-start recipe: ws = pounce.WarmStart.from_info(x, info); prob.solve(warm_start=ws) from Python, or the warm_start_init_point=yes mu_init=1e-7 warm_start_*=1e-9 option set on the CLI.
   • very large initial infeasibility → least_square_init_primal=yes, or repair the point: pounce.project_to_feasible(...).
   • large derivative scale spread → a scaling problem wearing an initialization costume; go to the scaling recipes.
3. Re-solve with --json-output ... --json-detail full and pounce-studio diagnose. restoration_used findings that disappear after re-initialization confirm the diagnosis; a clean preflight plus a failing solve means the start is NOT the problem — switch to the troubleshooting recipes.

The MCP twin (suggest_initialization in pounce-studio-mcp) runs step 1 and emits these suggestions mechanically; treat its output as advisory, exactly like the analyze-* suggestion blocks.

"Cite this for me"

pounce-studio citations --topic interior_point
pounce-studio citations --key wachter2006
pounce-studio citations                       # list all topics + keys

Notes on JSON detail levels

• --json-detail summary (default for pounce): writes outcome, statistics, no per-iter history. Compatible with summary, diagnose, linear-solver-summary, and compare. The iter-level tools (convergence-trace, get-iterate, find-stalls, restoration-windows) return empty arrays / errors.
• --json-detail full: adds per-iter history and suffix blocks. Required for the iter-level tools.

When kicking off a fresh solve to feed into the post-mortem tools, always pass --json-detail full.

Troubleshooting

pounce-studio: command not found The binary is not on PATH. Either re-run cargo install --path crates/pounce-studio-core (it installs into ~/.cargo/bin/) or pass the absolute path to the binary: ~/.cargo/bin/pounce-studio summary report.json.

pounce: command not found (when running fresh solves) The pounce solver binary is missing. Build with cargo build --release --bin pounce and copy target/release/pounce onto PATH, or run cargo install --path crates/pounce-cli --bin pounce.

unexpected schema "..." The JSON file is not a pounce.solve-report/v1 document — perhaps an AMPL .sol, a different solver's output, or a hand-written file. Re-run the solve with --json-output <path>.

summary works but convergence-trace returns empty arrays The report was written at --json-detail summary. Re-run the solve with --json-detail full.

gams: Could not spawn solver The GAMS POUNCE link is not installed. See gams/Makefile for build/install instructions.