β.3+β.5: AGENTS.md + vetted template + chat→solve→inspector slice (closes #3)

[Rchari1]

Closes #3. Stacked on #8 (β.2) → base retargets to main when #7/#8 merge.

Implements spec-20260615-113328-agents-template-inspector: the β.3 AGENTS.md+template merged with the β.5 single-run inspector into the full chat→solve→inspector vertical slice.

What this is (7 units)

• AGENTS.md rewritten for D9/D10 — the LLM authors a Julia script from the bundled template, fills params in-script, runs amico-run <script>. Relocated to packages/extension/ (the repo-root copy was invisible to the bundled extension → opencode was silently getting the stub fallback; fixed).
• templates/solve_template.jl — vetted single-transmon X-gate template: TransmonSystem → EmbeddedOperator → SmoothPulseProblem, Ipopt callback emitting AMICODE_ITER + per-iter control-amplitude PNG, atomic result.toml, pulse.jld2, DONE.
• Delivery — prepareOpencodeProject copies AGENTS.md (with the Julia project substituted in) + the template into the opencode session.
• amico-run on PATH — esbuild stages the launcher+bundle into bin/ (VSIX-self-contained); dev resolves the workspace sibling. .vscodeignore ships it.
• Runs-root reconciled — extension default /tmp/amicode-runs → ~/.amico/runs/default, matching amico-run's β.1 default, so the watcher sees what the agent's solve writes.
• Watcher reworked to the β.1 contract — keys completion on FINISHED (previously regex-scraped result.toml and never read FINISHED), identity from manifest.toml, unbounded iter_<N>.png digits. Pure vscode-free ingestRunDir (unit-tested) + incremental live path.
• Inspector/status-bar — CSP localResourceRoots now includes the runs-root (Q69, else blank inspector); status-bar renders failed/aborted and shows fidelity (was showing iter count mislabeled as F).

AC → evidence

Issue AC	Evidence
Template unmodified → FINISHED{completed} + pulse + ≥1 iter_*.png + AMICODE_ITER	slow-tier vet through real amico-run: F=0.99994, 16 iter PNGs, pulse.jld2
AGENTS.md: param-in-script (D10), run-dir contract, load_traj warm-start, no flag CLI	agents_md.test.ts guard
Artifacts delivered; stub-fallback fixed	relocation + opencode_config.test.ts
amico-run on PATH; project in-session; no cross-boundary env	opencode_paths.test.ts + AGENTS substitution
Watcher+amico-run agree on ~/.amico/runs/default; FINISHED-keyed; manifest identity; unbounded digits	watcher_contract.test.ts (completed/failed/aborted/no-promote) + opencode_paths.test.ts
Inspector renders; no CSP blank; failed/aborted render	inspectorResourceRootDirs test + watcher contract test

Fast tier: 44 amico-run + 26 extension tests green from a clean install. Slow tier (AMICO_TEST_JULIA_PROJECT=… pnpm --filter amicode-v2 test:slow) passed on the dev machine.

Notes

• API drift corrected (verified against Piccolo 1.19): load_pulse→load_traj; vetted on Piccolo 1.19 (was 0.11 in dev); fidelity measured via fresh rollout over the computational subspace (unitary_fidelity(U, op.operator; subspace=op.subspace)) — full-space fidelity caps at 4/9 for an embedded gate and was the initial red herring. The amico:transmon skill + Piccolo's canonical multilevel_transmon.jl drove the fix (drive_bounds 0.2, T=10).
• Verification boundary: every mechanical link is tested locally; the one unproven link is the chat-prompt→script LLM hop (needs opencode + creds) — user-verified at the β.6 demo.

🤖 Generated with Claude Code

[jack-champagne]

Partially reviewed — holding substantive feedback until #7 and #8 are reconciled and merged.

I've gone through this in the context of the stack and it reads as coherent with the β.3/β.5 plan, but I don't want to give line-level feedback yet, for two reasons:

1. The base is still moving — this PR retargets onto main once β.1: amico-run thin process orchestrator (closes #1) #7/β.2: vendor + pin opencode (closes #2) #8 land, so the diff I'd review against isn't settled.
2. Two threads from the earlier reviews land squarely in this slice and should resolve first:
   • the IpoptSolverExt reach-through (β.1: amico-run thin process orchestrator (closes #1) #7) — whether the β.3 AGENTS.md template canonizes that fragile methods(solve!) introspection, or seeds something sturdier;
   • the AMICO_* env reconciliation (β.2: vendor + pin opencode (closes #2) #8) — whether chat→solve drives amico-run purely via argv (per β.1's S37/Q90) or carries the spike env forward.

Once #7 and #8 are addressed, green, and merged, I'll do the full pass — correctness, the chat→solve→inspector slice, and a local run — against the settled base.

[Rchari1]

Both threads you flagged as blocking the full #9 pass are resolved — ready when you are:

1. IpoptSolverExt reach-through → gone. The template now uses the public Piccolo.Callbacks module (1.19), no methods(solve!) introspection. Re-vetted on Piccolo 1.19 end-to-end through the real amico-run: F = 0.99994 (transmon X, 7 iter PNGs, pulse.jld2, FINISHED{completed}). Two API corrections that came out of the 1.19 re-vet, both committed:
   • fidelity is now unitary_fidelity over the computational subspace from a fresh unitary_rollout — full-space fidelity against the embedded goal reads ~0.44 (the solve doesn't pin identity on the leakage level), and the raw final-propagator block read >1 from ~1e-6 integrator norm-drift. Subspace-rollout is the lab-standard metric and stays ≤1.
   • load_pulse → load_traj in the AGENTS.md warm-start idiom (the former doesn't exist in 1.19).
   • Also left a comment in the template pointing at the solver-agnostic AbstractIntermediateCallback/LivePulsePlotCallback path for when the default solver moves to MadNLP/Altissimo — directly per your note.
2. AMICO_* reconciliation → done (Task 4): chat→solve drives amico-run purely via argv; no env propagation (see my β.2: vendor + pin opencode (closes #2) #8 reply).

The base is still #8 → retargets to main once #7/#8 land. Branch force-updated from the restack (β.1 #7 fixes underneath). 45 amico-run + 26 extension tests green; slow tier green on the 1.19 project.

[jack-champagne]

Re-reviewed against the issue ACs (#3, #5) and pulled the branch locally (build/typecheck green, fast tier 45+26). The slice is well-built and the run-dir-contract consumption is solid — but two issue-AC items don't hold yet, both small fixes:

• #3 AC2 (warm-start): the documented load_traj idiom can't reload a template-produced pulse — the template saves JLD2 key trajectory but load_traj reads key traj, so the reload throws.
• #5 AC3 (status bar): solving (iter N) is never shown — the status bar only updates at completion; live iters go only to the inspector.

A few should-fixes are inline: the live-stats regex silently drops Inf/NaN (the stagnation/blow-up case those stats exist to flag), a replay→tail seam can drop a burst of iter lines, runs lose param/lab provenance, and the cold-start promote prompt can re-fire for an old run. There's also no test covering the live status-bar / incremental inspector path, which is why the status-bar gap slips past CI — one watcher test would catch it.

Solidly right: FINISHED-keyed completion reusing amico-run's validators, runs-root agreement with amico-run's default, no env propagation, no MCP/HTTP, the CSP fix, and scope stays in lane (no scheduler / sizeClass / HDF5 / multi-run). The promote→catalog prompt is a pre-existing stub from the β.2 base, not new here.

Requesting changes scoped to the two AC items (both tiny); everything else is should-fix / nit. (Minor, not inline-able since the file isn't touched here: server_manager.ts's ServerOptions.env comment still references a removed AMICODE_EXTENSION_URL / MCP-callback env — worth deleting.)

[jack-champagne]

Saves JLD2 key trajectory, but the documented load_traj reads key traj, so warm-start reload throws (#3 AC2) — use JLD2.save("pulse.jld2", prob.trajectory).

[Rchari1]

Fixed in a5121a9 — the template now saves JLD2.save("pulse.jld2", "traj", prob.trajectory), so the documented load_traj (which reads key traj) reloads it. Warm-start round-trips.

[jack-champagne]

Live iters go only to the inspector here — setRun fires only at completion, so the status bar never shows solving (iter N) (#5 AC3).

[Rchari1]

Fixed in a5121a9 — LiveRunSink.iter() now calls statusBar.setRun({status:"running", latestIter: rec.iter}) on every iter line, so the status bar shows solving (iter N) live, not just at completion (#5 AC3).

[jack-champagne]

This float regex rejects Inf/NaN, so stagnation/blow-up iters are silently dropped (#5 AC2) — and amico-run's own classifier keeps them, so the two disagree.

[Rchari1]

Fixed in a5121a9 — the NUM group now accepts Inf|-Inf|NaN, matching amico-run's classifier, plus parseAmicoNum maps them to JS Infinity/NaN. Stagnation/blow-up iters are kept. Guard test in watcher_contract.test.ts.

[jack-champagne]

Setting offset to the current size can skip AMICODE_ITER lines appended between ingestRunDir's read and this attach.

[Rchari1]

Fixed in a5121a9 — ingestRunDir now returns the run.log byte count it consumed, and the LogTailer attaches at that startOffset (not current EOF), so lines appended between the replay read and the tail attach aren't skipped. Covered by the 'returns the run.log byte offset' test.

[jack-champagne]

Invocation omits --lab and the template writes no params, so a run can't report which lab/params it actually solved.

[Rchari1]

Two parts: params — fixed in a5121a9, the template now writes a [params] table (delta/levels/T/N/drive_max/max_iter) into result.toml, so each run is self-describing. --lab — added in e21539c (on #23, stacked above): the AGENTS.md invocation now passes --lab default so the run records its lab.

[jack-champagne]

On launch this re-reveals and can re-pop the promote prompt for a prior session's completed run, since the promote-once guard is per-run and rebuilt each switch.

[Rchari1]

Fixed in a5121a9 — on launch the watcher baselines a prior finished run by setting activeRunDir without a sink (so it renders nothing and can't re-pop promote), and the promotedRuns Set is now shared across run-switches so a run promotes at most once. Only a fresh live completion prompts.

[jack-champagne]

Hardcoded /tmp is now redundant with runsRoot and needlessly broadens the webview's allowed roots.

[Rchari1]

Fixed in a5121a9 — inspectorResourceRootDirs no longer includes /tmp; the inspector only loads frames under runsRoot, so the webview's allowed roots are now just dist/media/runsRoot.

[Rchari1]

Thanks Jack — all addressed in the latest push (force-pushed; #10–#22 rebased on top).

Must-fix

• β.3 — AGENTS.md + vetted solve-script template #3 AC2 warm-start: template now JLD2.save("pulse.jld2", "traj", …) so load_traj (which reads key traj) can reload it.
• β.5 — Minimal single-run inspector (stats row + live plot) #5 AC3 status bar: sink.iter now calls setRun({status:"running", latestIter}) so the bar shows running · iter N live, not only at completion.

Should-fix

• β.5 — Minimal single-run inspector (stats row + live plot) #5 AC2 Inf/NaN: AMICODE_ITER_RE + a new parseAmicoNum accept Inf/-Inf/NaN, so blow-up/stagnation iters aren't dropped (matches amico-run). Tests added.
• log-tailer skip: ingestRunDir now returns the run.log bytes it consumed; the tailer attaches at that offset (+ drains once) — no gap, no double-emit.
• promote re-pop: a watcher-level promotedRuns set makes promote fire at most once per run; a run already finished when switched to (launch-follows-latest / switch-back) is pre-marked, so it never re-pops.
• /tmp root: dropped from inspectorResourceRootDirs (redundant with runsRoot).
• params recording: result.toml now includes a [params] table (δ, levels, T, N, drive_max, max_iter) so a run is self-describing.

Disposition: --lab on the invocation is genuinely multi-lab (Phase-2 lab.toml) — β is the single default lab, and the run now records its params in result.toml, so I left --lab for that phase. Shout if you'd rather I add it now.

[Rchari1]

Thanks for the thorough pass, @jack-champagne. Both AC blockers and all should-fixes are addressed:

Blockers (a5121a9)

• β.3 — AGENTS.md + vetted solve-script template #3 AC2 warm-start: JLD2 key traj (round-trips with load_traj).
• β.5 — Minimal single-run inspector (stats row + live plot) #5 AC3 status bar: iter() now drives setRun(status:running, iter N) live.

Should-fixes (a5121a9): Inf/NaN regex kept; replay→tail startOffset seam closed; [params] provenance in result.toml; promote no longer re-pops a prior finished run; /tmp dropped from webview roots. Stale server_manager MCP-env comment and the AGENTS.md --lab are in e21539c (on #23, stacked above).

Your top-level callout — 'no test covering the live status-bar / incremental inspector path': now covered. I extracted the live newest-wins dedup into a pure SinkDedup in run_dir_reader.ts and unit-tested it (watcher_contract.test.ts). This wasn't academic — that exact path had regressed so the run.log iters raced a shared counter past the lagging PNG frames and silently dropped every frame (blank inspector). The split (frames dedup on frames; log lines advance a separate high-water mark) + tests are in e21539c.

Re-requesting review. 47 tests green.

[jack-champagne]

Re-checked at a5121a9 — the two AC blockers from my last review are resolved:

• β.3 — AGENTS.md + vetted solve-script template #3 AC2 (warm-start): template now saves JLD2 key traj, matching load_traj.
• β.5 — Minimal single-run inspector (stats row + live plot) #5 AC3 (status bar): LiveRunSink.iter() now drives solving (iter N) (file_watcher.ts:52–59, status_bar.ts).

One gap remained on #5 AC1 ("display refreshes for the whole solve"): the live plot frame never advanced during a solve. image() and iter() share one latestIter (file_watcher.ts:37,47,52); the solver prints AMICODE_ITER iter=N before writing iter_N.png, so the tailer bumps latestIter to N and image(N) then hits N <= latestIter and drops the frame. There was also no final-frame push on completion. Stats rows streamed fine — it was the plot that froze.

Fixed in #23 (SinkDedup splits frame dedup from the log high-water mark), now validated end-to-end on the live vertical. Recording the link for history; resolved in the unit. Worth a follow-up: a watcher test that appends to a run dir over time so this path is covered going forward.

[jack-champagne]

Approving as a unit member — my two blockers are resolved here; the live-plot gap is fixed in #23 (validated end-to-end). Details in the comment above.