@leanandmean/coding-agent 0.74.1-scramjet.74 → 0.74.1-scramjet.76

package/docs/terminal-verification.md

@@ -16,6 +16,32 @@ The shared headless terminal feeds actual TUI writes into `@xterm/headless`; its
 
 The native journey fixture is `packages/scramjet/tests/fixtures/interactive-viewport.mjs`. Hosted interaction and graphics drivers live in `.github/scripts/terminal-probe.py` and `.github/scripts/terminal-safety.py`; `.github/scripts/test-terminal-fixtures.py` exercises their evidence predicates. `.github/workflows/terminal-feasibility.yml` owns the hosted matrix. Build the runtime before running fixtures, because they load compiled packages. Windows desktop/clipboard execution through `.github/scripts/windows-terminal-probe.sh` requires separate explicit permission.
 
+## Settings-input diagnostics
+
+The native interaction driver arms the fixture's serialized `settings-trace-dock`, `settings-trace-wheel` or `settings-trace-height` action immediately before typing the synthetic search query, and disarms with `settings-trace-end` in `finally`. The action ID is the transaction ID; `report.json` retains each transaction separately before subsequent input or cleanup can overwrite fixture state. `/settings` entry precedes the armed search window. Exact flushed `> query` readiness and its existing four-second polling deadline remain authoritative: there is one typing attempt, no prefix acceptance, paste substitution or automatic retyping.
+
+`settingsTransactions` records the intended `queryId`, ID, acceptance result, driver-local duration, optional posting receipt and final `fixture` evidence. Missing fixture/receipt evidence is explicit (`diagnosticError`, or an absent posting record if generation failed). `fixture.json` retains only the current or last transaction as `settingsInput`; a new arm resets all observations. A failed disarm retains the pre-disarm snapshot and does not replace the original acceptance error.
+
+Observation boundaries are deliberately distinct:
+
+- macOS key/text stdout contains ordered `keys` receipts with attempted physical codes/flags, down/up HID postings (including modifier events), local monotonic start/end and posting offsets, and frontmost bundle/PID at start plus PID at end. Physical ordering and 20ms pacing are unchanged. `hid-posting-only` means the posting calls returned, **not OS/terminal delivery acknowledgement**. The other Swift consumer, `terminal-safety.py`, ignores key/text stdout; mouse receipts are unchanged.
+- The fixture wraps each `ProcessTerminal.start()` callback once and forwards its original arguments and return value unchanged. `parsed` observations precede TUI preprocessing; they are already downstream of stream framing, terminal protocol interception and parsing, **not raw stdin**. Restart, suspension and external-editor handoffs reuse this wrapper rather than accumulating observers.
+- `listener` observations occur at the fixture's existing TUI listener. Earlier preprocessing/listeners may consume input; observing it here still does **not** prove settings-component receipt. Events include ordered sequence, fixture-local offset, editor-focus and frame-flush state. Final painted-query evidence accepts only prefixes of the expected synthetic query; anything else becomes `other-or-absent`.
+
+New input diagnostics contain only the expected query's allowlisted keys and press/release categories, unrelated-input categories, or `redacted-paste`. Parsed bracketed paste is classified before key matching; even split stream delivery is reassembled by the existing production parser, not a second diagnostic parser. No paste payload, clipboard text, raw stdin or payload hashes are added. These privacy claims concern the new diagnostic fields, not a redesign of pre-existing synthetic fixture snapshots. Each window expires after ten seconds, retains at most 128 events, and reports `expired`, total `observations` and `dropped`; disarmed/expired evidence is frozen until the next transaction.
+
+Correlate by transaction ID and ordered observations, not by subtracting monotonic origins from different processes. Posting without parsed-callback evidence leaves OS delivery, terminal encoding/stream delivery and parsing unresolved. Parsed/listener differences narrow application handling; neither an unchanged-code rerun nor these records establish a runner-flake diagnosis or fix the original lost character.
+
+`phaseSeconds` reports startup/calibration, interaction, cleanup and (on macOS) Swift compilation durations, including failure paths. Compilation is nested within startup/calibration, and settings transaction durations are nested within interaction: do not sum overlapping intervals. These are measured local elapsed durations, not configured allowances or end-to-end input latency. Offline Python/PTY checks verify accounting, transaction privacy/lifecycle and exact incomplete-query evidence, not fresh macOS posting/delivery. Hosted receipt inspection and matched before/after native timing remain required in S3 of issue #617; no live npm publication is authorized as a benchmark.
+
+## Job-local desktop driver reuse
+
+The macOS interaction workflow compiles `.github/scripts/macos-terminal-events.swift` once after its clean exact-revision guard into `$RUNNER_TEMP/native-driver/desktop-events`. It explicitly supplies `--desktop-driver=<path>` to installed and checkout journeys, each installed control and the ARM64 iTerm2 default-off diagnostic. Omission preserves standalone compilation; an explicit empty, missing, directory or non-executable path fails without compiling a replacement. Shared-driver input is macOS-only. Safety jobs remain independent.
+
+Uploaded `native-driver/compilation.json` identifies the source revision/SHA-256, binary SHA-256, compiler version, architecture, executable path and measured compilation duration. Each invocation's `desktopDriver` records mode (`shared`/`standalone`), resolved executable and binary digest, allowing correlation with that job's compilation record. Trust comes from same-job compilation of the guarded checkout, not from a digest alone. There is no cross-job/run/architecture cache or downloaded executable. Shared invocation has no `phaseSeconds.swiftCompilation`; the job's compilation duration is accounted separately, not charged to every invocation.
+
+Per-invocation SDK/capability checks, profile configuration, fresh owned terminal execution, calibration, full interaction predicates, controls and restoration are unchanged. Reuse does not make desktop delivery or cleanup redundant. Native before/after evidence must qualify sample count, runner image/architecture, terminal/profile/toolchain, failures and queueing; compilation-count reduction alone is not measured elapsed benefit. Full release performance remains a separate, independently authorized observation.
+
 ## Recovery input smoke — 2026-10-08
 
 A separately authorized Windows Terminal 1.24.12741.0 / Windows build 22631 / WSL offline production InteractiveMode smoke ran against freshly built Stage 5 changes on the dirty working tree based on `08defe2b1e33b0505e9105d36858d241ac17aa60`. A disposable fixture used isolated in-memory settings, a synthetic provider, and the real terminal-indicator implementation; no paid calls, personal extensions, clipboard writes or terminal preference changes occurred.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@leanandmean/coding-agent",
-  "version": "0.74.1-scramjet.74",
+  "version": "0.74.1-scramjet.76",
   "publishConfig": {
     "access": "public"
   },
@@ -37,9 +37,9 @@
     "test": "vitest --run --passWithNoTests"
   },
   "dependencies": {
-    "@leanandmean/agent": "0.74.1-scramjet.53",
-    "@leanandmean/ai": "0.74.1-scramjet.49",
-    "@leanandmean/tui": "0.74.1-scramjet.46",
+    "@leanandmean/agent": "0.74.1-scramjet.55",
+    "@leanandmean/ai": "0.74.1-scramjet.51",
+    "@leanandmean/tui": "0.74.1-scramjet.48",
     "@silvia-odwyer/photon-node": "^0.3.4",
     "chalk": "^5.5.0",
     "dictionary-en": "^4.0.0",