WIP night spike: pulse-designer Layer-0 — interview → entities → solve, e2e green by aarontrowbridge · Pull Request #75 · harmoniqs/amicode

aarontrowbridge · 2026-07-03T06:50:29Z

What this is

The Layer-0 pulse-designer track from tonight's spec (spec — drafted in Aaron's personal vault, spec-20260703-020630-amicode-opencode-thin-fork; will be socialized before team-canon promotion). It implements the front-of-chain interview (UX1 #46's storyboard) as opencode-native config: an agent + instructions + a bookkeeping tool pack. Nothing in the phased build moves — amico-run, run-dir contract, schemas, inspector, scheduler/Δ-cloud tracks are all untouched; the only modified prod file is opencode_config.ts (additive keys).

Commits

feat(L0) interview spine in AGENTS.md — 8 stages, one-question-at-a-time, scoped to the pulse-designer persona so the default build agent behaves exactly as before
test(L0) tiered e2e suite (A registration / B plugin load / C live turns / D full chain)
feat(L0)amicode_* tool pack v0 + registration (plugin abs-path + agent block via OPENCODE_CONFIG_CONTENT — both verified live at 1.17.3)
tier C + D enablement

Evidence (all on this branch, tonight)

Fast suite: 122 passed (was 91 on main)
e2e 4/4 tiers green incl. tier D full chain: 10 interview turns (free anonymous provider — zero creds on the machine!) → all four tools fired (System → Formulation → Run entities on disk) → agent authored solve.jl from the vetted template → detached amico-run launch → run-dir → FINISHED → F = 0.9998977, 60 iters
Repro: AMICODE_E2E_LIVE=1 AMICODE_E2E_FULLCHAIN=1 npx vitest run test/slow/interview_e2e.test.ts

Not in this PR

The opencode fork (branded binary — L1 "AMICODE" + build recipe) lives in a local-only clone, deliberately unpublished pending discussion. This branch runs green against the stock vendored binary too — the L0 layer is binary-agnostic by design.

Known follow-ups

.vsix packaging of opencode-plugin/ unverified (TODO in headers); set_model merges via json sidecar; Rydberg is recorded-honestly (template is transmon-only); free-model phrasing variance noted in the tier-D transcript.

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T07:01:22Z

T16 landed after the PR opened: amicode_to_hardware + amicode_calibrate guided stubs (no device I/O; gate/checks values pinned by the serializer). The pack now covers all 8 interview stages — 6 tools registered, 128 tests green. Also since opening: e2e re-verified 4/4 against the refreshed fork binary (fast suite + smoke + tiers; one tier-C flake traced to free-model sampling noise, passed on rerun — live tiers are env-gated off in CI).

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T08:20:04Z

Live-demo-driven additions (Aaron driving the dev host, ~03:00–04:00): Amico identity + proactive interview kickoff (fc7b13d) and amicode_ask — option-stage questions as button payloads, 7 tools registered (903ce76). The matching UI (entity rail, collapsed chips, ask-button cards, start-screen wordmark, en-default locale) lives in the local fork's patch stack (11 commits, unpublished — see the handoff note in the personal vault). All L0 changes here remain binary-agnostic: on a stock binary, amicode_ask degrades to a plain-text question with listed options.

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T09:13:16Z

This branch now additionally carries: the scores runtime (#76 merged, interview-as-data — score #0 v2 with ask-discipline + anchor-on-recorded-state), and the fixed opencode port (amicode.opencodePort, default 43117 — Remote-SSH users forward once). Suite: 203/203. Note for reviewers: 6471ed9 was momentarily red (hardcoded score-version literals in the stacked tests), fixed in 7f28414.

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T09:27:01Z

Final night addition: consolidated on opencode's native question tool (turn-blocking form — structurally prevents the self-answer/prose-repeat failure modes) as the one ask mechanism; amicode_ask deprecated-but-kept (8e9c89e). Score #0 at v3. Fork brands the native form and ships a 5-chip start screen with a repertoire-chips TODO (chips become one-per-entitled-score once a small endpoint exposes the compiled repertoire).

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T10:46:06Z

Ready for review — merge gated on @jack-champagne's eyes (Aaron's call: everything ships except the merge).

Since the last update: scores runtime merged in (#76 content — interview-as-data, score #0 v3, guard, router, usage funnel), fixed port 43117, curated capabilities answer, native-question-tool consolidation, H-robot brand mark + spinner, and the distribution chain is live: the vendored binary now comes from the private mirror release harmoniqs/opencode v1.17.3-amicode.1 (linux-x64 + darwin-arm64), clean-fetch verified byte-identical.

Review pointers (the seams):opencode.lock.json + scripts/fetch_opencode.mjs (repo/tag + authenticated gh path), src/opencode_config.ts (additive agent/plugin/permission injection), opencode-plugin/ (Bun-side, single-export), scores/ (splice never touches source AGENTS.md — fallback intact). Suite: 206 fast + slow gates green, CI green both platforms. Setup for testing: root AGENTS.md (agent-forward) or packages/extension/TESTING.md (human).

Known-not-done, honestly: Rydberg CZ template is EXPERIMENTAL (vetting stalled, marked in-file); amicode_ask deprecated-not-removed; ask-card/bridge legacy code pending pruning; repertoire-driven start chips need a server endpoint.

🤖 Generated with Claude Code

…the per-solve "ask" nag (#79) AGENTS.md instructs the agent to READ run artifacts (FINISHED/result.toml for results, run.log for failure tracebacks) under the runs root — but the injected permission.external_directory only granted the templates dir + /tmp/amicode-work. opencode defaults external_directory to "ask", so every read-back fired a permission prompt: one per solve, and repeatedly on failures (the 2026-07-03 live test hit this hard — four failing runs, agent in run.log each time). Thread the resolved runsRoot (respects the amicode.runsRoot setting) into buildOpencodeConfigContent and grant `${runsRoot}/**`. Still least-privilege: scoped to the one additional root AGENTS.md actually sends the agent's file tools into; amico-run's own subprocess writes never needed a grant. Tests: unit (grant present + still path-scoped), real-opencode integration (grant survives the 1.17.3 deep-merge alongside the user's global permission keys). 93 pass, typecheck clean. Note: Aaron's #75/#76 restructure the same permission block (adding entities/ scores grants — correctly); whichever lands second has a one-line rebase. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Scoped to the pulse-designer persona (never hijacks the default build agent), one-question-at-a-time, 8 stages PLATFORM→CALIBRATE, transmon fully wired / Rydberg recorded-honestly, amicode_* tools as bookkeeping with bash+amico-run still the launch mechanism. Tests: interview block + no-unknown-placeholder assertion (only {{TEMPLATE_PATH}}/{{JULIA_PROJECT}} substituted). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…gin load / live turns) Tier A green vs stock 1.17.3; B gates on the plugin file, C on creds (none on this machine tonight — morning step: opencode auth login). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Four bookkeeping tools (pick_system / set_model / formulate / solve-record) executed by opencode's embedded Bun runtime; entities as TOML under ~/.amico/runs/default/_entities ($AMICODE_ENTITIES_DIR override). Registration via buildOpencodeConfigContent: plugin abs-path + pulse-designer agent block + entities-dir permission grant (T8 probe: both mechanisms live at 1.17.3). Plugin stays outside the extension bundle/tsconfig; load line on stderr (debug-config stdout purity — regression-guarded). 99→122 tests green. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Free anonymous provider (opencode/big-pickle) resolves without auth.json → AMICODE_E2E_LIVE=1 forces the creds gate. Cadence assertion = stage-batching check (multiple '?' inside one platform question is fine); transcript saved to tmpdir for the experiment note. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Real-home mode now substitutes AGENTS.md via prepareOpencodeProject (stage 6 needs real template/julia paths). Keyword-routed answers, bounded turns, run-dir poll. Result: 10 turns on the free model → all four amicode_* tools fired (System→Formulation→Run stub) → authored solve.jl → amico-run launch → F=0.9998977/60 iters. AMICODE_E2E_FULLCHAIN=1 gates it. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

amicode_to_hardware (device_session stub; gate + checks PINNED by the serializer so a stub can never claim an approval that didn't happen) and amicode_calibrate (ILC stub, auto-refs device_session.toml). No device I/O — explanations say so explicitly, mirroring AGENTS.md stage 8, which now names both tools. 6 tools registered (verified via /experimental/tool/ids). 122→128 tests green. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

'Who are you?' now answers 'I'm Amico — Amicode's pulse-design copilot' (never 'opencode, an interactive CLI tool'), and a greeting/no-intent session opener triggers the stage-1 PLATFORM question instead of a generic assistant reply. Specific asks keep the skip rule. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Interview option-stages (platform, sim-vs-solve, problem/gate) now go through amicode_ask(question, options[2-6]); the renderer draws buttons from the tool part's input and a click sends the option text as the next user message. Free-form values stay plain questions. 7 tools registered (runtime-verified). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Live-demo findings: the free model repeated the question in prose and answered it itself. Tool return now says STOP HERE explicitly; AGENTS.md mirrors it. New optional details[] arg — one dim qualifier per button (minimalist card, slightly more context). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Pixel <0110> face in brand yellow on the H silhouette (from the website robot); currentColor body serves dark+light. Same asset ships in the fork's logo/favicon (kept in sync manually — canonical geometry recorded there in AMICODE-PATCHES.md). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ited stop Work found staged in the night tree (parallel session / Aaron) — folded in after validating build + full suite green. Fixes the Remote-SSH pain: forward 43117 once, restarts reuse it; stop() now resolves on child exit so a fixed-port restart can't race the old process for the socket. Set 0 to restore per-start ephemeral ports.

…tlement registry)

…kaging)

…N, plugin-sidecar pattern)

…e_manifest transport, scores-root grant, never-brick fallback

…s (score_guard sibling module, entitiesDir transport)

…nel (PASS 25s)

…ded-state (v2) The compiled score is now what runtime users see — the STOP-HERE ask discipline and optional per-option details move in from the AGENTS.md fallback, plus the anti-drift anchor from Aaron's live session (rail said rydberg while the interview asked transmon-omega): re-read the System entity before stage-2+ questions; correct the record first if it's wrong. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ion cross-check Score content bumps (v1->v2 tonight) must not red the suite; instead assert the compiled banner and the manifest agree on whatever version ships.

… deprecated Live finding: 1.17.3 ships a turn-blocking question tool (tool/question.ts) with options+descriptions+custom answers — the model reached for it unprompted and its semantics structurally kill self-answering and prose repeats. Score #0 v3 + AGENTS.md fallback now route all choice questions through it; amicode_ask stays registered but deprecated. One mechanism, upstream-owned. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ine talk The starter chip hit a Webfetch to the engine's website and a flat 3-sentence reply. The canonical question now has a canonical answer: capability bullets (interview, fast-path, live inspector, warm-start/resume, hardware preview), honest scope line, then next-step options via the question tool. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…late fetch_opencode gains repo/tag manifest fields with authenticated gh download for the private harmoniqs/opencode release (206 tests). TESTING.md = the 7-point branch test script for the team. solve_rydberg_cz.jl: 2-atom 3-level QuEra deep-blockade CZ, public-Piccolo-only (fixed-phase NLP + post-hoc virtual-Z scan; free_phase kwarg is Piccolissimo-gated) — vetting solve in flight, score wiring lands with its result. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Point an agent at the repo and it can set up, verify, and test the branch: prerequisites with checks, ordered setup, five verification gates, dev facts (fixed port, fork boundary, plugin runtime, scores), sharp edges.

…icode.1) Lock repointed to the mirror (both platform shas from the release assets); clean-fetch verified end-to-end via the authenticated gh path — downloaded binary is byte-identical to the deployed build (8aa718ad…) and serves <title>Amicode</title>. Rydberg CZ template clearly marked EXPERIMENTAL (vetting stalled on pathological first-iteration NLP cost; parked) and TESTING.md de-overclaimed accordingly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ot signature The rebase onto main folded in #79's runsRoot param (3rd position). These test call sites predated it and passed pluginPath/scoresRoot into the wrong slots or omitted runsRoot entirely. Insert runsRoot so scores + slow-e2e configs grant the right paths. Fast suite 207 pass; tsc --noEmit clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The pre-solve formulation.toml emit could not run through the agent deploy path: AGENTS.md copies solve_template.jl alone into a scratch dir and runs the copy, so @__DIR__ was the scratch dir and include(emit_formulation.jl) resolved to a file that was never copied — LoadError before solve!, on every run. CI never executes the template, so it stayed green. - Inline emit_formulation into solve_template.jl; delete the sibling helper. Sharing moves to a Julia package (AmicoRunDir.jl) when #75/#76 add a second template — the only share mechanism that survives the single-file copy. - Tighten formulation.schema.json: per-family if/then — transmon requires and type-checks delta/levels/drive_max; formulation requires gate/T/N; Q/R optional; additionalProperties:true per family (unknown families lenient). - Add template_deploy_contract.test.ts: reproduces the AGENTS.md single-file copy and asserts every include() resolves. Red on the bug, green inlined. schema_version stays "1" (additive: no on-disk formulation.toml exists yet).

… warrant-bounds schema (#703) (#706) * feat(schema): the autonomy datum — device {none|ro|rw} documented in warrant bounds, rejection pinned (#703) SEAM 6, amicode's half of the one-autonomy-doctrine (codesign spec spec-20260831-120000, issue #703): - $defs.bounds.device — the enum {none, ro, rw} — now carries the datum's documented semantics: none = no device access; ro = read-only device access; rw = device writes permitted, gated by the real-board-session human gate. The approval branch's description points at the datum (one definition, this schema — the fleet §2.1 pointer formalized). - The three-systems contract stated once in the schema's documentation, each counterpart's criterion NAMED under its own owner: amicode's bounds (this schema), the P4 gate's end-to-end read (merge-gated on strumento #75 MERGED), Telaio's warrant-bounds deserialization conformance (their campaign). The SEAM 4 contract note is not on main yet — the director wires the cross-reference line post-merge. - Rejection tests pin the no-second-knob doctrine mechanically: validateBounds REFUSES any device value outside the enum and any unknown device-permission-shaped field (additionalProperties:false), field-precise errors naming the impostor key; the approval record kind refuses a second device knob the same way. * feat(extension): the approval tool surface validates against the schema's device datum (#703) SEAM 6's tool-surface half (issue #703) — one definition, no free text: - amicode_request_approval's bounds arg now cites the @amicode/schema warrant-bounds schema ($defs.bounds — the one definition) and its execute() validates bounds via validateBounds: a device value outside {none, ro, rw} or a second device-shaped field is refused BEFORE the card renders, so a malformed ask can never become a pressable button. - The plugin twin (retired from the config, still drift-guarded by the projection test) mirrors the same gate: opencode-plugin/warrant_bounds.ts is the dependency-free twin of $defs.bounds (bare package specifiers do not resolve in the plugin runtime — unreachable by construction, per ./ledger_client's sibling-module rules), PINNED to the schema by test/warrant_bounds_parity.test.ts: the corpus (every datum case) must agree with validateBounds on ok AND error strings, and DEVICE_DATUM must equal the schema's own device enum — growth breaks the pin. - The mint path (amicode_service/warrants.ts approveArgv) validates bounds against the schema too, refusing invalid bounds instead of silently dropping them — a dropped key silently under-authorises the minted warrant; test/amicode_service_warrants.test.ts is the flag-mapping pin the export always promised, plus the refusal cases. The three-systems contract line for the PR (the director's): amicode's bounds are this slice (the schema datum, above); the P4 gate's end-to-end read is merge-gated on strumento #75 MERGED; Telaio's warrant-bounds deserialization conformance is their campaign's criterion — each named under its own owner, none claimed as another's.

aarontrowbridge · 2026-07-03T06:50:29Z

What this is

The Layer-0 pulse-designer track from tonight's spec (spec — drafted in Aaron's personal vault, spec-20260703-020630-amicode-opencode-thin-fork; will be socialized before team-canon promotion). It implements the front-of-chain interview (UX1 #46's storyboard) as opencode-native config: an agent + instructions + a bookkeeping tool pack. Nothing in the phased build moves — amico-run, run-dir contract, schemas, inspector, scheduler/Δ-cloud tracks are all untouched; the only modified prod file is opencode_config.ts (additive keys).

Commits

feat(L0) interview spine in AGENTS.md — 8 stages, one-question-at-a-time, scoped to the pulse-designer persona so the default build agent behaves exactly as before
test(L0) tiered e2e suite (A registration / B plugin load / C live turns / D full chain)
feat(L0)amicode_* tool pack v0 + registration (plugin abs-path + agent block via OPENCODE_CONFIG_CONTENT — both verified live at 1.17.3)
tier C + D enablement

Evidence (all on this branch, tonight)

Fast suite: 122 passed (was 91 on main)
e2e 4/4 tiers green incl. tier D full chain: 10 interview turns (free anonymous provider — zero creds on the machine!) → all four tools fired (System → Formulation → Run entities on disk) → agent authored solve.jl from the vetted template → detached amico-run launch → run-dir → FINISHED → F = 0.9998977, 60 iters
Repro: AMICODE_E2E_LIVE=1 AMICODE_E2E_FULLCHAIN=1 npx vitest run test/slow/interview_e2e.test.ts

Not in this PR

The opencode fork (branded binary — L1 "AMICODE" + build recipe) lives in a local-only clone, deliberately unpublished pending discussion. This branch runs green against the stock vendored binary too — the L0 layer is binary-agnostic by design.

Known follow-ups

.vsix packaging of opencode-plugin/ unverified (TODO in headers); set_model merges via json sidecar; Rydberg is recorded-honestly (template is transmon-only); free-model phrasing variance noted in the tier-D transcript.

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T07:01:22Z

T16 landed after the PR opened: amicode_to_hardware + amicode_calibrate guided stubs (no device I/O; gate/checks values pinned by the serializer). The pack now covers all 8 interview stages — 6 tools registered, 128 tests green. Also since opening: e2e re-verified 4/4 against the refreshed fork binary (fast suite + smoke + tiers; one tier-C flake traced to free-model sampling noise, passed on rerun — live tiers are env-gated off in CI).

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T08:20:04Z

Live-demo-driven additions (Aaron driving the dev host, ~03:00–04:00): Amico identity + proactive interview kickoff (fc7b13d) and amicode_ask — option-stage questions as button payloads, 7 tools registered (903ce76). The matching UI (entity rail, collapsed chips, ask-button cards, start-screen wordmark, en-default locale) lives in the local fork's patch stack (11 commits, unpublished — see the handoff note in the personal vault). All L0 changes here remain binary-agnostic: on a stock binary, amicode_ask degrades to a plain-text question with listed options.

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T09:13:16Z

This branch now additionally carries: the scores runtime (#76 merged, interview-as-data — score #0 v2 with ask-discipline + anchor-on-recorded-state), and the fixed opencode port (amicode.opencodePort, default 43117 — Remote-SSH users forward once). Suite: 203/203. Note for reviewers: 6471ed9 was momentarily red (hardcoded score-version literals in the stacked tests), fixed in 7f28414.

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T09:27:01Z

Final night addition: consolidated on opencode's native question tool (turn-blocking form — structurally prevents the self-answer/prose-repeat failure modes) as the one ask mechanism; amicode_ask deprecated-but-kept (8e9c89e). Score #0 at v3. Fork brands the native form and ships a 5-chip start screen with a repertoire-chips TODO (chips become one-per-entitled-score once a small endpoint exposes the compiled repertoire).

🤖 Generated with Claude Code

aarontrowbridge · 2026-07-03T10:46:06Z