Roadmap

Each milestone has a verifiable exit criterion — the loop’s stop condition. Don’t start a milestone until the previous one’s criterion is met.

Milestone	Status
M0 — Model viability	DONE 2026-07-22 — see m0-results.md
M1 — Scanner core	DONE 2026-07-30 — 40 rules, fixture-tested
M2 — GitHub integration	DONE 2026-07-30 — see m2-results.md
M3 — Report + narration	in progress — parts built, wiring is Queue 2
M4 — Hardening & polish	in progress — download, caching, capability check and options page built; severity tuning and the full audit outstanding

M0 — Model viability spike — DONE

Question to answer: can Gemma 4 E4B (via @litert-lm/core on WebGPU) produce useful, non-hallucinated analysis of Terraform code on consumer hardware?

• Serve spike/model-eval/ (npm run spike), load the model, paste in the 10 fixture samples from spike/model-eval/fixtures/ (known-bad Terraform with documented issues)
• For each sample, record: did the model’s commentary correctly describe the planted issue when given the finding? Did it invent nonexistent issues? Tokens/sec?

Exit criterion: on ≥8/10 fixtures the model produces a correct, specific explanation with zero fabricated resources/attributes, at usable speed (>10 tok/s). If this fails, the project pivots or stops — decide with the user before M1.

M1 — Scanner core — DONE

HCL parsing + rule engine with an initial ~15 rule set (public S3/GCS buckets, open security groups 0.0.0.0/0, unencrypted storage/EBS/RDS, hardcoded secrets, missing versioning/logging, overly-broad IAM). Pure TS, no chrome.* imports.

Exit criterion: every rule has known-bad + known-good fixture tests passing; scanner runs against a real-world sample repo checkout and findings are manually spot-checked.

M2 — GitHub integration — DONE

Fine-grained personal access token (read-only) stored in chrome.storage.local; fetch repo tree + .tf file contents via api.github.com; handle pagination and

100 file repos.

Exit criterion: lists a private repo’s .tf files and runs the scanner on them end-to-end; no request leaves the allowlist (verified in the DevTools network log).

Met 2026-07-30 via npm run verify-m2 against a real private repo — see m2-results.md. Verified in the harness rather than the side panel, because the panel is still a scaffold; the code path exercised is the shipping one (src/lib/github/client.ts → src/lib/scanner).

M3 — Report generation + LLM narration

Assemble scanner findings into a structured report; feed findings + relevant code snippets to the on-device model for explanation, prioritization, and module-level design commentary. Render in side panel with export (markdown download).

Exit criterion: full run on a real repo produces a report where every stated vulnerability traces back to a rule-engine finding ID; model prose reviewed for fabrications on 3 sample repos.

M4 — Hardening & polish

First-run model download UX (progress, OPFS/Cache API caching), WebGPU capability detection with graceful fallback message, rule severity tuning, options page.

Built: the options page downloads a catalogue model with progress and can delete it or revoke the host access; the Hugging Face hosts moved to optional_host_permissions so the install prompt names GitHub alone; the side panel falls back to a downloaded model and structurally cannot start a download itself. Still open: rule severity tuning, and the audit below.

Checked in a real browser (2026-08-19, Chromium 148, unpacked dist/)

• unlimitedStorage is not needed. navigator.storage.estimate() reports a 10.7 GB quota for the extension origin, against 2.97 GB for the largest catalogue model. So the install prompt stays at GitHub alone. Residual: Cache API storage without that permission is best-effort and the browser may evict it under pressure — navigator.storage.persist() would ask for durability and costs no permission.
• The incomplete-download guard fires correctly. A connection dropped mid-transfer, cache.ts caught the byte-count mismatch, discarded the partial copy and reported it. This is the guarantee the module exists for and it had never been exercised for real.
• Still unverified: a successful multi-gigabyte Cache.put, and narration end-to-end. The test machine has no WebGPU adapter under this Chromium, so nothing has yet run a model inside the shipped extension.

Known gap — a failed download starts over

There is no resume. A 2–3 GB transfer that dies at 95% is retried from zero, and on a flaky connection may never complete; the user cannot tell “try again” from “this will never work here”. HTTP Range requests against the partial bytes are the fix.

It is deliberately not a quick one. Resuming means keeping a partial entry on disk, which is precisely what #download’s discard-on-failure exists to prevent — a partial model cached under the real key would be served as the whole thing. Any resume design has to keep partial and complete entries structurally distinguishable, or it trades a slow download for a corrupt model.

Exit criterion: clean install → report on a fresh machine profile without developer intervention; extension passes privacy-auditor full audit.