PDD Alpha 0.1.10

PDD Alpha 0.1.10: drawings as the screen spec, Try tools for your agent, fakes for the outside world, clearer access answers, and safer workflow recovery and replay reports.

Update

npm install -g github:promptdriven/pdd
npx skills add promptdriven/pdd#alpha -g -a codex -a claude-code -y
pdd --version      # 0.1.10; warns when your skills are older
pdd setup          # in each repo; asks before registering Try tools

These update commands apply after the reviewed release reaches both main and alpha. Start a new agent chat after updating so it loads the new skills.

What is new

• The drawing is the spec. Linked .ui.svg drawings name a screen's parts, words, ids, order and states. The skills use them during build and verification. pdd check --lint --local [STORY...] reports local drawing warnings with no sign-in, network or AI calls. pdd check --screen URL --drawing PATH compares text, ids, regions and spatial order in Chrome or Chromium with 15% position tolerance on each viewport axis. It uses no AI calls and does not compare fonts, colors or pixels. Set PDD_CHROME_PATH when browser discovery needs help.
• Try tools for your agent. pdd mcp offers try_open, try_state, try_run, try_check and try_save_case over stdio. Use them when you ask your agent to try behavior. Runs use the Try budget. Saved cases are separate observations for review, not approved expectations. pdd setup shows proposed registrations for supported agents and asks before writing them. Missing runtime data snapshots leave state unavailable and prevent saving a case.
• Run on request in Try. The local agent can fill fields and run an action when you explicitly ask. Otherwise you press Run. pdd advertises its supported actions when connecting; older pages keep manual controls. This release's proof does not submit a live chat turn.
• Fake the outside world. RUNBOOK §21 and six packaged examples show databases, files, hashes, services, decisions and factories behind faked ports. The prompting, feature and bug skills direct outside effects through those ports.
• More precise checks. Binding coverage reports which selected modules have bindings, or says coverage is unavailable. Local oneOf advice distinguishes overlap, miss and uncertainty; numeric advice compares related quantities. pdd check --read CALL_ID reads an existing result without sending its action again.
• Clear Cloud access answers. no-cloud-access and cloud-unavailable stop immediately with access or retry guidance. New refusals exit 5; loss of access while reading an already submitted check preserves its recovery command and exits 3. Nested error details take precedence over compatibility wrappers. Default program update guidance installs from main, while skills install from alpha.
• Workflow recovery from linked checkouts. Verified normal and bare repository parents can host issue worktrees started from a linked checkout. Confirmed startup refusals recover the same allocated run; uncertain replies and identity conflicts stay explicit.
• Structured replay reports in code space. New bundled feature and bug code areas validate an agent's bounded replay manifest against saved original vector bytes and current source identity. Missing or stale evidence appears as BLOCKED or NOT REPLAYED above prose and prompt-runtime results. Actual assertion failures stay visible and can be published for review. Updates mark an existing report pending before pushing changed source. Agents execute tests and supply the record; pdd validates structure and identity, not independent execution or equivalence. Saved workflows without the opt-in retain their behavior.
• Final published cases stay in replay scope. Prompt-space publications refresh the scenario anchor through final code approval, then freeze it. Code publication independently lists current changed scenarios and shows BLOCKED if any are absent from that anchor or the inventory cannot be read. Unpublished scenario edits are refused before code entry. Exact Git paths retain non-ASCII and whitespace-bearing names. This fixes #2553; a passing report for an older saved subset cannot hide a newly added scenario.
• Safety and reliable tests. Approval context reading avoids stalls on repeated wrappers; Try launch paths are passed as data. The Herdr transport fixture declares its module scope so an ESM parent cannot break it.

Known limits

• Formal not-run and not-applicable outcomes remain explicit; replay PASS does not claim a formal proof.
• Live Try state/saving depends on snapshot support from the runtime. This bundle does not deploy runtime changes.

Checks

• Final integrated code: 3587e0cc46e45ba8b54d23979165b4a059b80aa7. Both final corrected full suites on the runner VM passed 892/892, with zero failures or skips: 102.03 seconds and 105.29 seconds.
• Real Chrome test results within each suite passed 82/82 across four groups, including the four group results, with zero failures or skips. PDD_CHROME_PATH selected /usr/bin/google-chrome.
• A fresh verifier ran 94 focused and neighboring checks on the correction with zero failures or skips. All four original regression assertions also failed against a throwaway old-code copy, including the actual wrong PASS; this is a semantic control, not a crash. All 43 generated prompt headers match; package, lock and five skill versions are 0.1.10.
• A user-style GitHub branch install prints 0.1.10. All 153 packaged files match the final source. The documented skills command, using the release branch, installs five 0.1.10 skills for both Codex and Claude Code; all 60 installed skill files match and the older-skills warning clears. The pre-existing old pdd still comes first on the default PATH; the proof uses the installed Alpha executable directly.
• With that freshly installed CLI, pdd check fake-hash --max-calls 18 on the live runtime passed in 24.2 seconds: 1/1 criterion, 1/1 check, 2 AI calls; requested model gpt-6.1-sol. Draft check report. A DRAFT pass is not approval. The report link expires after seven days. Binding coverage is unavailable without a binding manifest. This smoke test does not prove native workflow replay semantics.
• The initial full run exposed three environment-dependent test-fixture failures. Fixture isolation fixed them without changing expected assertions; tracked in #2551. The #2553 reproduction is now covered by checked-in native tests. A fresh verifier confirmed all four original regression assertions fail against the old code and pass after correction. No formal proof is claimed; these module prompts have no Runtime interface.
• The corrected-head public GitHub check passed 892/892, zero failures or skips, in 129.88 seconds on its first attempt. The earlier pre-correction Chrome startup timeout remains tracked in #2555.
• Runtime AI calls for this correction proof: 2; total for the original release and correction proofs: 4. No live chat turn, new MCP execution proof or runtime deployment is claimed.

Promotion verification

• Promoted to main in #2565, commit b2d275a64da48a5eb8f71e53209798974f96c98b. The main and alpha release trees match. The reviewed issue-reporting guide and links from #2541/#2542 are included.
• A fresh npm install -g github:promptdriven/pdd from main prints 0.1.10. All 154 packaged files match main and alpha: 148 unchanged from the tested bundle and six reviewed documentation files, including one new guide. The documented alpha skills install matches all 60 installed files across Codex and Claude Code.
• Promotion main CI and both CodeQL language checks passed. The tag identifies this verified main release commit.