mr-onion

Deferred Work Register

Date: 2026-10-19 Status: Living document. Every piece of work a decision record deferred, parked or left open, in one place, with its source, its status and what it waits on.

Triage result (2026-10-19)

Each item below gets its own decision record, locked with Emil before Code builds it. Phase numbers are given as each record is written.

Build order after Phase 39 (updated 2026-10-20, with R5’s three records placed). Phase 39 runs as locked. R4 comes after it, or fits between its gates: TFP-206, TFP-209, TFP-216, TFP-217 Part A and TFP-222, with TFP-222 before Phase 39’s TFP-155.

  1. D26: the local copies move into WSL.
    • Testsmith’s checkout moves into WSL’s Linux filesystem, so starts stop crossing the Windows mount.
    • The workarounds are re-checked.
    • The README’s start timings are corrected.
    • Failure sets and score are proven unchanged, and the run times become the new baseline.
    • Moved into Phase 39 by Gate D’s ruling 3 (2026-10-26); built and independently verified 2026-10-27 (benchmarks/2026-10-27-d26-wsl/). The run-time baseline comes from Gate D’s re-measurement. Then, before item 2: Phase 47, Mr Onion in a target’s CI pipeline (added 2026-10-27, Emil: after Phase 39, before Phase 43; numbered here without renumbering the items below, which the tables’ Order columns cite). - Standard, read-only, against an app the CI job starts on localhost; deployed URLs capped until Phase 43. - The crawl and generate logic moves into src/lib/pipeline/ first (TFP-263); then the CLI, baselines, a Docker image and a GitHub Action. - Tickets TFP-263, TFP-245 to TFP-248; TFP-249 after Phase 43 (D31). - Versioning (TFP-257 steps 1–5) comes before its image. - Paused at TFP-247 for Phase 48 (Emil, 2026-11-08): TestForge becomes Mr Onion (D39) on branch phase-48-mr-onion, merged to main in one go, then tagged 0.3.0. TFP-247 and TFP-248 then ship mr-onion. Tickets TFP-275 to TFP-281; no measurement session while the branch is open. - Resumes at TFP-247 (owner decision, 2026-11-12): Phase 48 is complete (2861a4c, v0.3.0). - Complete 2026-11-13 at TFP-248 (owner decision, 2026-11-13), Gate A passed; released as 0.4.0. TFP-249 stays deferred as D31, after Phase 43.
  2. Phase 43 (D8): the ownership confirmation, once per site.
    • Before it: TFP-287 (owner decision, 2026-11-14), the read-only guard’s dormant wrapper, so the covered-host guard is built on it. The record was reviewed against main on 2026-11-14; see its “Review before building” section.
    • Saved with the site: exact hosts, contact email, 90-day expiry, active and load ticks.
    • It lifts Phase 39’s caps.
    • With a run’s MRONION_ACTIVE switch on, it opens the active gate and the script-payload test runs again.
    • Uncovered runs still run, with their active tests Not run.
    • Tickets TFP-218 to TFP-221.
  3. Phase 46: the app screens.
    • A site lookup by URL, with exact match only, no list of sites and the email masked.
    • The confirmation form.
    • The candidate promotion screen: approve, then Claude Code commits.
    • The active switch in the run form.
    • Tickets TFP-232 to TFP-235.
  4. D29: a retention policy for saved runs and their evidence.
    • How long saved runs and their videos, traces and screenshots are kept, and a “clear old runs” action on Phase 46’s screens.
    • Its own short decision record with Emil’s choices.
  5. D1: the SEO Standard set. The draft set from the coverage proposal, brought into the repo and written once on Phase 39’s shared visit.
  6. D2 with D9: the 30 WAI-ARIA APG patterns.
    • Functional checks for each pattern, shipped in batches if its record says so.
    • INP timed on those interactions and Phase 41’s, as banded findings only.
  7. D3: the rest of accessibility and security Standard. With D30, the solo performance pass.
    • The passive checks need nothing more.
    • The attack-shaped ones sit behind Phase 43’s active gate.
  8. D19: the failure-triage gate.
    • Each failure judged as a real defect, test bug, environment, or inconclusive, with a confidence score and a threshold below which a person decides.
    • Calibrated on the labelled failures.
  9. D27: Mr Onion measured against QA My App and Playwright’s Test Agents.
    • A comparison benchmark on the local Toolshop copies, run once Standard is complete (after D3 and D19).
    • Scored with TFP-154’s rule against the answer key.
    • Every Mr Onion miss is tagged “planned in Dx” or “not planned”. Only the unplanned ones become work, and they can reorder the Comprehensive and scenario steps that follow.
  10. D28: a security sweep of Mr Onion’s model-code paths.
    • Zone by zone, each finding proven by a failing test.
    • Covers the gap writer, candidates and bound candidates, TFP-202’s code check and stripped environment.
    • Done before Phase 45 builds the sandbox.
  11. Phase 45: the Docker sandbox.
    • Every model-written test runs in a container: gap tests, candidates and scenario tests.
    • Read-only mounts, no secrets, a non-root user.
    • Network only to the run’s target hosts.
    • No Docker means they don’t run.
    • At most 60 s per run.
    • Tickets TFP-228 to TFP-231.
  12. Phase 44: scenario gaps, slice 1.
    • Double submit, back-and-resubmit, and retry after a timeout.
    • Built from a vetted harness, with Claude Code’s judgement.
    • Each scenario type is proven able to fail on a planted-defect fixture.
    • Outside the verdict until measured (D6).
    • Interactive gate, plus the confirmation on remote hosts.
    • Tickets TFP-223 to TFP-227.
  13. D10: authenticated and active security (GitHub issue #3), with Phase 44’s slice 2 (two tabs on one account, session expiry mid-flow).
  14. D7: Comprehensive’s negative tests. By ISTQB technique, with a technique-coverage measure, including scenario #8 (valid but unusual data).
  15. D4: Comprehensive SEO, then the highest-risk areas from run history.
  16. D5: all-areas Comprehensive runs, scheduling, per-area jobs.
  17. D24: the opt-in load-testing run. Designed with the ISTQB Performance Testing syllabus, and needs Phase 43’s load tick. Last for now (Emil, 2026-10-19); moves up if needed sooner.

Not placed yet: D25, a summary email after a site run. Its record is still to write.

Alongside, in Cowork:

Standing rule (Emil, 2026-10-19): any site Mr Onion hosts locally under Docker keeps its code in WSL’s Linux filesystem, not in a Windows folder mounted into Docker. See D26.

While the Standard sets are built (D1, D2 with D9, D3): test the section, not the whole app (Emil, 2026-10-20).

The same rule for the Comprehensive work (Emil, 2026-10-20). It covers Phase 44, D10, D7 and D4.

Every new Standard set (D1, D2 with D9, D3): Phase 39’s rule still holds. Each record states the set’s run-time cost against the margin measured after Phase 39 (with-bugs at TFP-208: 1,306 s of 1,800 s).

The rest:

Statuses:

Test suites not yet built

“Order” is the place in the triage result’s build order.

# Item Source Emil (2026-10-19) Status Order Waits on
D1 SEO Standard set (the draft set in the Phase 38 proposal) Phase 38, deferred Include Planned 2 D8, and Phase 39 complete, so its checks are written once on the shared visit (TFP-212). The record copies the draft SEO set from the coverage proposal (Claude Docs, 2026-10-05) into the repo, and states its run-time cost
D2 Functional checks for the 30 WAI-ARIA APG patterns Phase 38, deferred Include Planned, with D9 3 D1; TFP-214 (the “in place” kind of control). The record decides whether it ships in batches
D3 The rest of the accessibility and security Standard sets Phase 38, deferred Include Planned 4 D2. The passive parts need nothing more. Attack-shaped checks sit at the active gate, which D8 opens
D4 Comprehensive SEO, then the highest-risk areas from run history Phase 35, deferred Include. Triage: after D7 Planned 8 D7; reuses D1’s checks. The highest-risk areas are chosen from run history at that point
D5 All-areas Comprehensive runs, scheduling, per-area jobs Phase 35, deferred Include Planned 9 D4; Phase 39 Gate D (workers)
D6 Model-written tests counting toward the verdict Phase 35 Decision 5, deferred “Depends on what these model-written tests find”. Decided 2026-10-20: outside the verdict until measured, per scenario type Decided in Phase 44 Decision 6: a scenario type joins the verdict when every planted defect of its type is found on the fixture and there’s no unlabelled failure on clean fixture, clean Toolshop or 3 real sites, with Emil’s approval 9 (with Phase 44) Phase 44’s measurement
D7 Comprehensive’s negative tests, by ISTQB technique, with a technique-coverage measure Phase 38, deferred No problems. Triage: before D4 Planned 7 D10. Includes scenario #8 (scenario-gap-candidates-2026-10-19.md)
D8 The ownership confirmation and the active gate, with the script-payload test. The load-testing run is split out to D24 Phase 35 Decision 8; Phase 38 A tester confirms once, not repeatedly. Triage: see “D8, confirming once” below Planned 1 Phase 39 and R4. Locked as Phase 43 (phase-43-ownership-confirmation-decision-record.md, Emil 2026-10-19; TFP-218 to TFP-221). It unblocks D3’s active parts, D10, D24 and Phase 39’s ownership row
D9 INP measured through real interactions Phase 38, deferred Needs explaining (below). Triage: with D2, banded findings only Planned, with D2 3 D2; Phase 41’s GET-only interactions
D10 Authenticated and active security Phase 35, deferred; GitHub issue #3 Not sure it should be parked entirely. Triage: the first Comprehensive item Planned (sequence only; no design yet) 6 D8, D3, D19. Overlaps scenarios #3, #4 and #9; R5’s scenario-gaps record says which side owns each
D24 The opt-in load-testing run, designed with the ISTQB Performance Testing syllabus Phase 38, deferred; split from D8 (2026-10-19) Split out of D8. Placed last for now; moves up if needed sooner Planned 10 D8’s load-testing tick

Features asked for, not yet planned

# Item Source Emil (2026-10-19) Status Order Waits on
D26 The local Toolshop copies move into WSL’s Linux filesystem, with the README’s start timings corrected Emil, 2026-10-19, from TFP-211’s slow with-bugs start (about 6 min, composer once 10.5 min) Wanted, after Phase 39; and the rule for every locally hosted site Built 2026-10-27, awaiting verification (benchmarks/2026-10-27-d26-wsl/) Moved into Phase 39 by Gate D’s ruling 3 (2026-10-26), before the re-measurement Phase 39 complete: the move also speeds up the Toolshop’s API, so mid-phase it would shift the run times Phase 39 compares. What it covers is below
D27 A comparison benchmark: Mr Onion against the “QA My App” plugin (Anthropic Directory, community) and Playwright’s Test Agents (planner, generator, healer; added 2026-11-02) on the local Toolshop copies Emil, 2026-10-20 Wanted: use outside testing tools as yardsticks, not inside Mr Onion Planned 8, after D3 and D19, once Standard is complete (Emil, 2026-10-20: earlier, it would mostly point at work already planned) D26 (both tools run on the WSL setup), D1 to D3, D19. What it covers is below
D28 A security sweep of Mr Onion’s own model-code paths, with the “Verified Security Sweep” plugin (Anthropic Directory, community) or an equivalent review Emil, 2026-10-20 Wanted, before the sandbox is built Planned 9, after D19, before Phase 45 Nothing; it informs Phase 45’s design. What it covers is below
D29 A retention policy for saved runs and their evidence Emil, 2026-10-21, after 278 MB of leftover run folders were cleared Wanted: stop cluttering with records and evidence nobody opens Planned 4, straight after Phase 46 Phase 46 (the action belongs on its screens). What it covers is below
D30 A solo performance pass, so real sites without the ownership confirmation still get performance judged Emil, 2026-10-21, after TFP-157’s ruling that timed checks are Not run under the request cap Wanted Planned, with D3 7, with D3 TFP-157 (Phase 39). What it covers is below
D25 A summary email after a run against a site, to the tester’s inbox, with the way back to the full report Emil, 2026-10-19, with TFP-217 (the same for Mr Onion’s own nightly CI) Wanted: a quick view for whoever is in charge of testing that site To plan Not set A decision record of its own. It is most useful with D5’s scheduled runs. The record decides the questions below
D31 Active and interactive gates against deployed URLs in CI (TFP-249) Phase 47, deferred (2026-10-27) Wanted, after Phase 43 Planned After Phase 43 Phase 43, TFP-248
D32 An npm package for the CI runner Phase 47 Decision 8 (2026-10-27) Deferred: the image pins Chromium, npm wouldn’t Open - TFP-247
D33 Models (gap writer, local model, OpenAI) in CI Phase 47 Decision 9 (2026-10-27) Off in CI for now Open - Phase 45’s sandbox
D34 Comprehensive all-areas runs in CI Phase 47, deferred (2026-10-27) Wanted, with D5 Planned, with D5 With D5 D5, TFP-248
D35 Native integrations for other CI systems (GitLab component, Azure DevOps task) and SARIF output for code scanning Phase 47, deferred (2026-10-27) Not before someone asks Open - TFP-248
D36 Workers that back off when a site slows down (adaptive concurrency) Phase 39 re-measurement ruling 1 (2026-10-27) Deferred: a fixed limit of 3 holds for now Open - Phase 39 complete
D37 The stale re-check’s re-runs judged for cap starvation (it has no timing reporter, so a reused test it re-runs on a capped host keeps its re-run result) Phase 39 ruling 4 amended (2026-10-28) Deferred: rare path, documented as a known limitation Open - TFP-264
D38 “Comprehensive” becomes “Break Me” once more areas have Comprehensive sets: UI, report and docs wording Emil, 2026-10-28 Wanted, when more areas are built Open - More Comprehensive areas (D4, D7)
D39 A new application name (many “testforge” repos exist): repo, package, testforge-skills plugin, data folder (%LOCALAPPDATA%\TestForge, one-time move), TESTFORGE_* settings, test-ID prefixes, TFP- prefix, testforgeapp@gmail.com Emil, 2026-10-28; proposed name “Mr Onion” (Emil, 2026-10-30) Wanted. Tagline “Mr Onion: peel back every layer” (Standard the first layer, “Break Me” the core). Ruled out: “Mr Pickles” (Cucumber’s “pickles”, the Pickles docs tool, a TV brand), “Señor Onion” (could read as a caricature; the ñ breaks technical names), “Allium” (JUXT’s spec language with a Claude Code plugin; Alumnium, an AI testing tool), “Tearjerker”, “Core Sample”, “Shallot” (owner’s preference) Done 2026-11-12 (2861a4c, Phase 48 merged 2026-11-11; Phase 48 complete 2026-11-12). Locked → Phase 48 (Emil, 2026-11-08): phase-48-rename-to-mr-onion-decision-record.md. Test-ID and TFP- prefixes stay; the mailbox is D44. Nothing is ever created again as emil3663/testforge, so the old URL keeps redirecting (Phase 48 Decision 10) Phase 48, after Phase 47’s TFP-246, before TFP-247 -
D40 An image a later check’s own action brings in (a lazy-loaded image it scrolls to) charged to that check, not to the test that loaded the page, so a check the cap starved that way is Not run, not Failed Phase 39 rulings on 70a0ad3, item 2 (2026-10-28) Deferred: a known limitation, #16 Open - TFP-264
D41 Runs that really go through a proxy: the browser using it (Playwright’s proxy option) and the lowercase variable names (http_proxy, https_proxy, no_proxy); TFP-209 only passes the uppercase names through TFP-209’s verification (2026-11-02) Deferred, decided with Phase 47 (CI runners behind a corporate proxy) Open With Phase 47 Phase 47’s record
D42 A small System One decision model (Jev first; others benchmarked against it) as a fallback when the deterministic rules can’t settle which Standard tests apply: page type, a form’s purpose, which APG widget a control is, whether a library test fits a component Emil, 2026-11-04, revisiting Martin’s System One pointer from Phase 34 (only the typed-verdict pattern was adopted then, in TFP-119) Wanted, with the reasoning below; Jev because most people he has spoken to use it, with a benchmark of the others at the time; packaging looked at closer to the time, with its pros and cons Open With D2, when its widget recognition is designed D2; the labelled set below. What it covers is below
D43 Ending the child-side double setting names: saved site-suite units and library entries that check TESTFORGE_* names are regenerated or rewritten, so runs stop setting both names Phase 48 Decision 8 (2026-11-08) Deferred: until then every run sets both names Open Not before 1.0’s removal ticket, and not part of it Phase 48 merged; the 1.0 record
D44 A Mr Onion mailbox in place of testforgeapp@gmail.com and its +ci and +<site> aliases, if one is wanted Phase 48 Decision 4 (2026-11-08) Deferred: the old mailbox stays; the nightly’s sender name becomes “Mr Onion nightly”; the nightly’s TESTFORGE_SMTP_PASSWORD secret keeps its name until this lands (Phase 48 Decision 9, amended 2026-11-08) Open - Phase 48 merged

D39, before the name is committed (2026-10-30). “Mr Onion” is the proposed name. First checks: no software, app, GitHub project or testing tool found under it; mr-onion and mronion unregistered on npm (2026-10-30). Before anything is renamed:

  1. Trademark search: CIPC (South Africa); the US (USPTO) and EU (EUIPO) registers if it will be marketed abroad. Classes 9 and 42 (software, software services). Results so far (owner and Cowork, 2026-11-03; a first pass, not legal clearance):
    • US (USPTO), live marks, classes 9 and 42: 340 matching “MR” or “ONION”, none “MR ONION”. The only “MR. ONION” filed (1994, processed onions) is dead. Live “___ ONION” software marks exist: SECURITY ONION (9, 42), TRUST ONION (9, pending), THE ONION TALKS (9, pending), BLUE ONION (42, ×3), RED ONION (42), DIGITAL ONION (42). None covers testing or developer tools, but “onion” is a crowded word in software: easy to register around, weak to defend.
    • TMview (EU, UK and the other offices it holds), “Mr Onion”: 32 marks. None live in 9 or 42: the one class 9 filing (Thailand, 2009) has ended, and there are none in 42. Live “Mr Onion” marks are in other classes: a restaurant chain’s in China (29, 30, 31, 43; its domains may be taken), Malaysia (29, 30, 43), Mexico (31), a Chinese financial services firm (35, 36), Germany’s “Mr. Onions” (35, 41, 43). No live EU or UK mark, and none for Moon and Me’s “Mr Onion” character.
    • South Africa: no company named “Mr Onion” on CIPC’s BizPortal. CIPC’s trade mark search couldn’t be found in eServices. Owner decision (Emil, 2026-11-03): the trademark check is done for the rename. No same name was found in South Africa; the mark will be registered in the EU or the US, not South Africa. Noted: a US or EU registration gives no rights in South Africa, and South Africa’s register itself wasn’t searched.
  2. Domain: decide whether one is wanted, and check mronion / mr-onion in the extensions wanted.
  3. Registries: GitHub repo emil3663/mr-onion (free under the owner’s account), npm mr-onion (re-check just before publishing), container image ghcr.io/emil3663/mr-onion (Phase 47, TFP-247).
  4. Internal names, chosen together in the rename’s own decision record: the settings prefix (MRONION_*, with the TESTFORGE_* names read as a fallback for a while), the local data folder and its one-time move, the plugin name, the test-ID prefixes (TFFU-, TFSE-…), the ticket prefix, and the testforgeapp@gmail.com addresses.
  5. Timing (Emil, 2026-10-30): the rename is its own short phase with its own decision record, after the trademark check and before Phase 47’s TFP-247 publishes an image or action; outside pipelines will pin that name. v0.2.0 is tagged under the current name when Phase 39 completes: tags and releases survive a GitHub repo rename, and old links redirect.
  6. Mascot (optional): a friendly onion character, which also settles the faint Tor (“onion”) association.

D42, a decision model for choosing Standard tests (2026-11-04). Martin pointed at System One models (small, fast models for a single yes/no or pick-one decision: Jev, Kev, Laya, GLiNER2.5-Decide) during Phase 34. Then, only the pattern was adopted: each gate records a typed verdict (TFP-119), and the checks behind it stay deterministic code. Emil asked (2026-11-04) whether such a model could help decide which Standard tests run against a page.

Phases or items not wrapped up

# Item Source Emil (2026-10-19) Status
D11 Phase 28, OpenAI path: Gate A open, no live completion ever succeeded Phase 28 Park Parked (2026-10-19). Revisit if a funded key is used
D12 Phase 29, the crawl skill’s drift from dom-snapshot.ts, and its live verification Phase 29 (draft); Phase 22 Needs to be addressed; discuss Decided in R6 (2026-10-20): the skill describes the method and points to the code for the rules (r6-skills-relook-2026-10-20.md); TFP-237
D13 Reliability’s element-mapping (local model) Phase 33, open Needs explaining (below) Parked with the local-model path (2026-10-19)
D14 The database test’s skip on pages that save nothing (local model) Phase 33, open Needs explaining (below) Parked with the local-model path (2026-10-19)
D15 Template-versus-local-model comparison Phases 31–33, deferred Check it Closed: measured in Phase 34 (TFP-120, phase-34-option-b-measurement-report.md) and confirmed by Emil 2026-09-28. With page data the model was never called (0 gaps on both targets; 15 of 17 cases identical to the template’s suite)
D16 The hollow cookie/session test Phase 34, deferred Not a valid test Closed: replaced in Phase 36 (TFP-125) by a check on the site’s own session cookies (expiry and scope), skipped with a reason when there are none
D17 Reliability gap-filling (local model) Phase 34, deferred Needs explaining (below) Parked with the local-model path (2026-10-19)
D18 A UI control for gap-filling Phase 34, deferred Needs explaining (below) Closed: decided against in Phase 34 Decision 9 (confirmed by Emil 2026-09-28). Revisit only if a site shows gaps with page data
D19 The failure-triage gate (real defect, test bug, environment, or inconclusive) Phase 34, deferred (Martin’s review) Needs explaining (below) Planned (2026-10-19), order 5: after D3, before D10. It is calibrated on Toolshop’s labelled failures (TFP-178), R1’s 5 false link failures, and the new Standard sets’ failures, so it is in place before Comprehensive and the practice sites multiply the failures
D20 Skills used as rules at run time, and proving the skills Phase 36, deferred Needs explaining (below) Decided in R6 (2026-10-20): no new run-time rules; each plugin skill proven by known-answer cases through its real invocation (TFP-239)
D21 TFP-169, the occasional connection reset in a security test Known-bug backlog Can we force it? Open. TFP-208 (verified) didn’t see it in its 6 runs, nor R1 in its 3; 3b44370’s diagnostics stay. TFP-210’s nightly run watches it alongside D23. Forcing it (below) is a ticket if it recurs
D23 performance-bands.test.ts:69, a timing test, failed once in TFP-208’s check run on 02ed0ef (the fixture page built for an LCP of about 3 s measured over 4 s, with the machine loaded); passed alone 2 of 2, and in the 6 full runs before TFP-208’s verification (2026-10-19) — Open: known flake. TFP-210’s nightly run will show its rate; a ticket if it recurs. Second timing test, 2026-10-21: timing.test.ts’s “/busy never goes quiet” settle cap (0 capped of 1) failed once in TFP-155’s full unit run (9efbd5f); passed alone 3 of 3 and in that night’s four earlier runs. Likely the same cause: under load the fixture’s polling slows past the 500 ms quiet window. One data point each; a recurrence of either gets a ticket, and the fixture’s /busy is to be made busy by a mechanism that load can’t slow. Third, not timing, 2026-10-21: route.inspect.test.ts’s closed-port test failed once in a full unit run on a969cbf (600 of 601, 1,144 s): closedPort() frees a port, and before the browser tried it something else answered there (ERR_INVALID_HTTP_RESPONSE, not the ERR_CONNECTION_REFUSED it expects). Passed alone 8 of 8, and in that day’s other full runs. A recurrence gets a ticket: a port that stays refused, not one freed and hoped for
D22 The real-site backlog’s stale header (TFP-162 to 164 shown open) Real-site backlog Update it Closed 2026-10-19: header and order list updated (cefa24d, be688d3, f4ed183, 837de4f, verified 2026-10-07)

Explanations Emil asked for

D8, confirming once. Proposal for the design:

Triage (Emil, 2026-10-19): the proposal stands, with these calls:

D9, INP. Interaction to Next Paint is the Core Web Vital for responsiveness: the time from a click, tap or key press to the next frame the browser draws.

Triage (Emil, 2026-10-19): built with D2. It adds no page visits: it times Phase 41’s interactions and D2’s APG checks.

D13, D14 and D17, the local-model items. All three are weaknesses of the optional local model (Ollama), not of the template:

The template covers reliability and database itself, and since Phase 34 the model is never called when page data exists. Parking all three with the local model costs nothing today.

D19, the failure-triage gate. A step that judges each failure as a real defect, a fault in the test, an environment problem or flake, or inconclusive.

D20, skills at run time. Mr Onion’s own app reads none of the skills under skills/. Phase 36 asked whether some should become rules the app applies, for example:

“Proving the skills” is Martin’s idea: test each skill against cases with known answers, as Mr Onion is tested.

D25, what its decision record has to settle. TFP-217 emails Mr Onion’s own CI result from GitHub. A site run is different: it runs in the app, on the tester’s machine.

D26, moving the local copies into WSL.

Why (as it stood before D26 was built, 2026-10-27). Testsmith’s checkout sits in a Windows folder (..\practice-software-testing) mounted into Docker. Every file operation crosses from Windows into Docker’s Linux. The slow steps on each start are:

with-bugs waits for composer to finish (bug 80’s log folder; see scripts/toolshop-local/README.md). One start took 10.5 minutes, and TFP-211’s took about 6. A first install took 20 to 25 minutes. Docker recommends keeping bind-mounted code in the WSL filesystem on Windows.

What the item covers:

The rule from now on applies to every site Mr Onion hosts locally under Docker, such as the-internet’s local copy (Phase 39, TFP-158) and the practice sites (practice-site-targets-2026-10-19.md). Its code lives in WSL’s Linux filesystem from the start.

D27, the comparison benchmark.

Why: Mr Onion’s goal is to find every issue on a site. A comparable tool run on the same target, with the same answer key, shows what Mr Onion misses and what it finds that the other doesn’t. That is measured, not claimed.

Not inside Mr Onion: outside tools are yardsticks only. A third-party plugin in a run would bring someone else’s judgement into the verdict, an update cycle Emil doesn’t control, and licensing questions. Most of these plugins also work from source code, while Mr Onion tests from the outside.

When (Emil, 2026-10-20): once Standard is complete, after D3 and D19. Run at step 2 it would mostly point at Standard sets not yet written. Run at the very end it could no longer change the order.

What it covers:

D28, the security sweep of the model-code paths.

Why: model-written tests turn text read from a site into code that runs on Emil’s machine (Phase 42, batch 2 ruling 5). TFP-202 hardened that path. A review that proves each finding with a failing test checks the hardening before Phase 45 contains it and Phase 44 puts it to real use.

What it covers:

D29, retention for saved runs.

Why (2026-10-21):

What its decision record settles (Emil’s choices):

D30, the solo performance pass.

Why (2026-10-21):

What it covers:

D12, for the R6 discussion. The crawl-and-classify-pages skill is a written copy of how dom-snapshot.ts classified a page on 2026-09-24:

Options:

  1. Sync it, with a “last synced at commit X” marker and a check that warns when dom-snapshot.ts changes.
  2. Make the skill describe the method and point to the code for the rules.
  3. Retire it.

D21, forcing TFP-169. Probably, yes. TFP-208’s fix (decbfad) found that unit tests shared one folder that another run could change mid-start, which is a likely cause of the same kind of start-up failure. A ticket could try to reproduce it on purpose:

If none of these reproduces it on the code after TFP-208, close it as fixed by decbfad, and keep 3b44370’s diagnostics in place.

Also tracked elsewhere