# Appendix A — Browser AI Field Guide: Setup, APIs, States, Diagnostics and Configuration
# Book: Browser AI From First Principles
# Chapter URL: https://programmer.ie/books/browser-ai-from-first-principles/25-appendix/
# Prose fingerprint (sha256, normalized): 6314f71440a0475431264b3b3831b6f81c29c6bac7702853fa2fa9c06fb25212
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

This appendix consolidates the contracts, states, and procedures distributed through the argument into one working reference. Prose explains; this section is for the keyboard.
Publication snapshot: September 2026. Browser flags, origin trials, API availability, and experimental configuration change. Flag values and external documentation details below reproduce the tested builds of this edition. Anything the book’s own laboratories did not observe is marked as such.
A.1 Development setup Surface Setting / location Value / action Purpose Stability Chrome flags chrome://flags/#optimization-guide-on-device-model Enabled (lab profile) On-device model support for local development Experimental, recheck per build Chrome flags chrome://flags/#prompt-api-for-gemini-nano Enabled where applicable Foundation-model API access for local testing Experimental, recheck per build WebMCP chrome://flags/#enable-webmcp-testing Enabled Local WebMCP development without relying on the origin trial Experimental, recheck per build Extensions chrome://extensions Developer mode, load unpacked Load the Browser AI Observatory Durable workflow Diagnostics chrome://on-device-internals Open model status Inspect on-device model state and errors Browser diagnostic surface JavaScript LanguageModel.availability(options) in DevTools Run with intended options Operational preflight before creation API contract Observatory AI Observatory panel, capability inspection Run inspection Capture the full lifecycle instead of guessing Book tooling Flag values follow the laboratory run-book convention of this edition, not a permanent browser contract.
A.2 Availability and lifecycle states One hierarchy, observed in the book’s laboratories:
Layer State Meaning What to do Interface not exposed Global such as LanguageModel absent Check browser, channel, context, configuration Interface exposed JavaScript interface exists Continue to operational preflight Capability unavailable Requested configuration cannot currently be created Explain, change route, or fall back Capability downloadable Required assets can be acquired Require visible user action Capability downloading Acquisition underway Show progress and wait Capability available Compatible instance can be created Call create() Session creating Browser is preparing the instance Measure creation separately Session ready Resource exists Permit operation Session running Request active Permit cancellation Session destroyed Explicit lifecycle end Do not reuse Request completed Runtime returned successfully Still evaluate output Request aborted Intentionally stopped Do not count as model failure Request failed Runtime operation failed Classify the failure Feature accepted Application validator accepted the output Use the result Feature rejected Model completed but the feature contract failed Repair, retry, or abstain available is not a ready session, and operational completion is not behavioral acceptance.
A.3 Built-in AI API matrix Globals, operations, and streaming support as implemented in the book’s adapter registry:
Capability Global Create Main operation Streaming Output Prompt LanguageModel LanguageModel.create() prompt() promptStreaming() Text, optionally constrained Summarization Summarizer Summarizer.create() summarize() summarizeStreaming() Summary text Writing Writer Writer.create() write() writeStreaming() New text Rewriting Rewriter Rewriter.create() rewrite() rewriteStreaming() Transformed text Proofreading Proofreader Proofreader.create() proofread() — Corrected text with correction evidence Detection LanguageDetector LanguageDetector.create() detect() — Ranked language candidates Translation Translator Translator.create() translate() translateStreaming() Translated text Session verbs observed in code: create, prompt, clone, destroy. The extension worker in this edition executes prompt; other operations return an explicit unsupported marker rather than failing silently.
A.4 Prompt API configuration quick reference Key Where Purpose Important note expectedInputs create() and matching availability check Declares input modalities and languages Must agree with preflight options expectedOutputs create() and matching availability check Declares expected output Treat as capability contract samplingMode Enrolled web experiments Semantic sampling policy Not available on ordinary pages by default topK Extension compatibility surface Numerical sampling control Legacy extension surface; also a named compat alternative temperature Extension compatibility surface Sampling variation Legacy extension surface; also a named compat alternative monitor create() Observe acquisition progress Progress events do not prove fresh bytes transferred signal creation and request operations Cancellation Preserve abort separately from failure responseConstraint prompt operation in laboratory surfaces Constrain output with a schema Laboratory surface; structure still requires validation A.5 Sampling modes Per current Chrome Prompt API documentation (external, September 2026 — not lab-observed except as noted):
samplingMode Intent most-predictable Maximum predictability predictable Strongly predictable slightly-predictable Mildly predictable balanced Balance of predictability and variation (typical default) slightly-creative Mildly varied creative More varied most-creative Maximum variability Book laboratory default: most-predictable. Reason: controlled experimental comparison and compatibility with the runtime observed during the book’s experiments, including the recorded speculative-decoding rejection that named it alongside topK: 1 and temperature: 0. Not a universal recommendation: creative or writing features may legitimately choose another supported profile.
A.6 Session lifecycle Exposure, creation, ready, usage, clone, destroy, failure — see the table in Chapter 9, Section 6. Application rules: measure creation separately from inference, snapshot usage and quota, branch experiments with clone() rather than reprompting, end with explicit destroy(), and record why destruction occurred.
A.7 Observatory event dictionary Every trace in this book validates against one closed vocabulary. Canonical lifecycle events:
Event Meaning capability.inspect.started Capability inspection began capability.inspect.finished Exposure and availability observed capability.inspect.failed The inspection itself failed model.download.progress Browser emitted acquisition progress session.create.started Creation attempt began session.create.finished Session available to the application session.create.failed Creation failed session.snapshot Usage, quota, and resource state observed session.cloned Branch created from an existing session session.destroyed Resource explicitly ended prompt.started Model request began prompt.chunk Streaming output observed prompt.finished Completed, failed, or aborted fixture.evaluated Machine-checkable evaluation performed feature.validation.finished Application-level acceptance or rejection Coordination, authority, policy, and context events:
Event Meaning coordinator.job.received Job accepted by the coordinator coordinator.job.queued Job awaiting a worker coordinator.worker.selected Idle worker matched capabilities coordinator.lease.issued Work leased, never duplicated coordinator.job.started Execution began (authority consumed here for bound jobs) coordinator.job.completed Result captured coordinator.job.failed Error captured authority.effect.normalized Exact effect canonicalized authority.grant.bound Single-use grant issued with digest authority.grant.consumed Grant consumed once authority.grant.expired Grant lapsed unused policy.compile.finished Portable policy compiled or rejected policy.decision.finished Decision with rule and reversibility recorded policy.run.finished Policy evaluation complete, external effects counted context.manifest.loaded Site context validated context.unit.extracted Bounded unit admitted with provenance context.unit.excluded Omission recorded with reason structured.candidate.received Proposal entered admission structured.validation.finished Gate outcomes recorded A.8 Capture and privacy modes Metrics-only capture is the default: sizes and outcomes are recorded without retaining prompts or page text. Content capture is explicit and visible. Private units are excluded by default and cannot cross an external provider boundary; exclusions carry reasons rather than failing silently.
A.9 Structured-output validation gates Parse, schema, domain, policy, authority, then bounded repair — see the table in Chapter 16, Section 11. A model may propose a structured action; only deterministic policy and appropriate human authority may admit it for execution.
A.10 WebMCP quick reference The live surface is document.modelContext.registerTool(definition, { signal }), withdrawn via AbortController. Registration requires the origin trial or the testing flag, a secure top-level context, and an exposed global; the book’s probe records available or absent rather than assuming. Observed in the tested profile: WebMCP absent — the capstone never depends on it being present.
A.11 Authority states A grant binds one normalized effect — origin, tool, arguments, policy version — under a SHA-256 digest, with a TTL and single use. Lifecycle: bound, consumed exactly once at the coordinator START boundary, or expired. Replay is denied before execution; mutation is rejected on digest mismatch. Policy requests authority; it never manufactures it.
A.12 Diagnostic decision table Symptom First check Next check Likely layer LanguageModel undefined browser, channel, context flags and configuration exposure availability() returns unavailable exact options passed hardware, storage, language support capability downloadable visible user action acquisition monitor model acquisition downloading never completes chrome://on-device-internals restart, storage pressure acquisition available but create() fails creation options sampling and profile compatibility session configuration Session succeeds, no answer prompt trace abort and error events request Prompt completes, result wrong fixture and evaluator trace with captured evidence behavior Worked yesterday, unavailable today model and browser state storage, updates, build browser-managed lifecycle WebMCP global absent origin trial and testing flag browser build WebMCP exposure Visible tool, wrong tool selected selection fixture descriptions, schemas, results agent behavior A.13 Lens modes See the comparison table in Chapter 24, Section 2. One line each: Understand interprets, Explore retrieves bounded evidence, Govern applies user policy reversibly, Act proposes and executes through admission, policy, authority, and coordinator boundaries.
A.14 What to record in every experiment Selection and inputs, provider and data route, evidence with provenance, policy version and decision, authority state for effects, worker and execution lineage, the validated trace — enough that a stranger can replay the run and reach the same verdict about what happened.
A.15 Laminated card UNDERSTAND interpret EXPLORE retrieve bounded evidence GOVERN policy, reversible ACT propose, never self-authorize MODEL CAN: perceive, classify, generate, rank, propose MODEL CANNOT: grant authority, change policy, bypass admission, consume a grant twice, invent provenance, turn uncertainty into certainty AVAILABILITY: absent < exposed < downloadable < available SESSION: create, run, clone, destroy — measure each GRANT: exact effect, once, expiring — consumed at START TRACE: every claim validates or it did not happen