Skip to content

WorkWingman full handoff for Clahadore Jones-Gaffney (Claude)

Prepared by Cedric “Cody” Jones-Gaffney (Codex) on 2026-07-26 at approximately 10:45 PM America/Chicago.

Mission and deadline

Deliver by Monday, July 27, 2026 at 8:00 PM America/Chicago a security-gated WorkWingman release containing:

  1. Critical desktop résumé/profile persistence and updater reliability fixes.
  2. Integrated reviewed UI, Degree Completion, Flow/Studio, tutorial, and Live Run improvements.
  3. The WING-196 shared Amazon-style job-filter experience.
  4. A private multi-user Cloud Run YC pilot with invitations, authentication, tenant isolation, quotas, and provider dispatch.
  5. A functional API-enabled feature path.
  6. A tested installer and deployment artifacts.
  7. Evidence-backed status for anything that cannot safely pass.

Do not call the release complete without requirement-by-requirement evidence.

Standing operating constraints

  • C# first; F# only if C# genuinely cannot do the job.
  • All code/config/tests/reviewable repo docs must be edited in an isolated Git worktree.
  • Never build or edit features in the main checkout.
  • Do not use local/Ollama models.
  • Do not use gaming-PC or streaming-PC council models.
  • No UAC prompts, ACL weakening, or manual queue-file manipulation.
  • Do not use Claude as an independent review seat for this release.
  • Required release/security panel:
  • Regular Jenny: agy with Gemini 3.1 Pro High.
  • Cyber Jenny: agy with Gemini 3.6 Flash High and a distinct adversarial-security prompt.
  • Gronktayvius General: Grok correctness/architecture/regression.
  • Gronktayvius Red Team: Grok abuse/trust-boundary/bypass review with a separate prompt and artifact.
  • Cedric/Clahadore adjudication after independent output.
  • Required seats must preserve requested/resolved identities, prompt hashes, stdout/stderr, tool versions, exact scope, and coverage. Missing or unverifiable output fails closed.
  • Security-sensitive/release work cannot use an outage waiver as a substitute for required seats.
  • Push only after review, commit, and successful gate. After pushing, run fleet repo sync.

Session identity and boards

  • Session: WT-3379 · WING-202 · integrate Monday Pam and YC release candidates · gov:orch
  • Jira: https://workwingman.atlassian.net/browse/WING-202
  • Vikunja: http://127.0.0.1:3456/projects/8
  • Local board: C:\Users\fives\source\repos\WorkWingman\docs\BOARD.md

Frozen integrated desktop baseline

  • Worktree: C:\Users\fives\source\repos\_wt\workwingman-WT-f26b-final-order-assembly
  • Branch: wt/WT-f26b/WING-final-order-assembly
  • Commit: e5b1d7f7ce4a2bd1ef1a4aadb21f418d023e7448
  • Tree: 6f0cb85c89acbf46156f0af90397cedbccfd8a58
  • Parent: d0b592dbccf374b201e65f7514c5f9127dbf5908
  • Exact one-commit diff SHA-256: f51f1eeb8a37e2cc6ad0edf8fba36782da91204d3b3ae6ff0b1d618555637d65
  • Worktree was clean when last checked.

Important history already contained in this baseline includes:

  • cbc4c3fea: integrated résumé and updater reliability.
  • aceb18530: installer version 0.2.1.
  • 932028843: assembled the WING-202 release subtotal across UI/API/filter/experience/connection/location/tax work.
  • b1eeb0567: private Cloud pilot deployment controls.
  • af104ab0e and 1de63e055: tenant-safe Cloud job explorer and filter-contract hardening.
  • WING-196 filter work through aaff6c389, plus newer Study hardening.
  • Flow/federal navigation work through 1dc02a2f3.

Do not assume the release is ready merely because these commits exist. Final integrated testing and release-wide review remain required.

First hosted review gate and why the installer is held

Evidence:

  • C:\Users\fives\source\repos\_review-evidence\WING-202\e5b1d7f7
  • Adjudication: C:\Users\fives\source\repos\_review-evidence\WING-202\e5b1d7f7\CEDRIC-ADJUDICATION.md

The exact e5b1d7f7 one-commit gate failed closed. Blocking High findings:

  1. Untrusted prompts were flattened into ProcessStartInfo.Arguments, enabling quote/backslash corruption and possible option smuggling.
  2. The hostile-argument unit test was tautological because expected output reused the same quoting helper.
  3. Grok’s partial tool denylist was fail-open and did not prove installed runtime capability/containment.

The candidate installer script correctly refuses to build unless its specified adjudication contains an exact bold PASS.

Active Grok argv/capability remediation

  • Worktree: C:\Users\fives\source\repos\_wt\workwingman-WT-3379-WING-202-grok-argv-fix
  • Branch: wt/WT-3379/WING-202-grok-argv-fix
  • Base: e5b1d7f7 / tree 6f0cb85c
  • Current state: intended changes are still uncommitted.

Changed files:

  • src/WorkWingman.Core/Interfaces/IProcessRunner.cs
  • src/WorkWingman.Infrastructure/Clients/SystemProcessRunner.cs
  • src/WorkWingman.Infrastructure/Llm/CliHarnessPreset.cs
  • src/WorkWingman.Infrastructure/Llm/CliLlmHarness.cs
  • tests/WorkWingman.Tests/Architecture/HarnessMeteringDoorTests.cs
  • tests/WorkWingman.Tests/Llm/CliLlmHarnessTests.cs
  • tests/WorkWingman.Tests/LlmControllerTests.cs

Implemented direction:

  • Add a discrete IReadOnlyList<string> process-runner overload.
  • Use ProcessStartInfo.ArgumentList; never flatten untrusted prompt/model data.
  • Give Codex, agy, and Grok exact argv templates.
  • Pin reviewed Grok CLI version 0.2.112.
  • Run provider-free --version and --help capability probes.
  • Grok fixed argv uses:
  • one turn,
  • no subagents,
  • no memory,
  • web disabled,
  • empty built-in tool allowlist,
  • dontAsk,
  • explicit deny rules for Bash/Edit/Write/Read/Grep/WebFetch/WebSearch/MCPTool,
  • plain output,
  • prompt as one final argv element.
  • Non-tautological hostile prompt tests cover embedded quotes, trailing slashes, CRLF/flag-shaped text, Unicode, and a long payload.
  • All concrete IProcessRunner implementations/test doubles now implement both overloads; root caught and corrected the initially missed LlmControllerTests.NoopRunner.

Evidence so far:

  • Secondary diagnostic: genuine compile plus 37/37 focused tests passed, 0 failed/skipped.
  • TRX: D:\fleet-temp\WT-3379-grok-argv\secondary.trx
  • Outputs: D:\WWBuild\WT-3379-grok-argv
  • This used copied exact-base generated obj assets because the Codex sandbox cannot read the user NuGet configuration. Treat it as useful secondary evidence, not the authoritative normal-user run.

Next safest action:

  1. Make the provider-free capability probe parse the full fixed containment argv plus --help, not merely a generic help call, so the installed CLI parser proves it accepts the exact empty-allowlist/dontAsk/deny composition.
  2. Run git diff --check.
  3. Commit only the seven intended files.
  4. Record exact commit/tree.
  5. Run authoritative normal-user listener build/tests with D-drive outputs.
  6. Run the four required hosted review seats on the new exact commit.
  7. Adjudicate. Do not build a release installer on a failure or missing seat.

WING-196 shared job-filter verification

Source verification lane:

  • Worktree: C:\Users\fives\source\repos\_wt\workwingman-WT-3379-WING-196-restack
  • Branch: wt/WT-3379/WING-196-restack
  • Clean at e5b1d7f7 / tree 6f0cb85c
  • Verdict: no restack or cherry-pick is needed.

Why:

  • Current job-query client and spec are byte-identical to late WING-196 commit 9f99484.
  • Current Study code is a semantic superset with bounded layoff hydration, fail-closed evidence handling, completeness validation, and a 30-minute verified-empty expiry.
  • No source/API change was necessary.
  • Cloud filter path remains tenant-session gated, tenant-key validated, rate-limited, capped at 25 results, and makes no media/AI provider call.
  • Ordinary locked endpoints remain outside the Cloud server allowlist.

Authoritative listener job:

  • Job ID: 26e57867-42a6-440e-931a-40e00187343c
  • Target: ANDYGREATROOMPC
  • Listener-created worktree: C:\Users\fives\source\repos\WorkWingman-agents\26e57867
  • Branch: wt/26e57867/WING-196
  • Exact clean HEAD/tree: e5b1d7f7 / 6f0cb85c
  • Evidence root: D:\WorkWingman-WING-196-listener-evidence
  • Command: D:\WorkWingman-WING-196-listener-evidence\logs\dotnet-command.txt
  • Live log: D:\WorkWingman-WING-196-listener-evidence\logs\dotnet-test.log
  • Last checkpoint: restore was live and Core + CloudDeploy had restored; no TRX and no failure yet.

Next action: continue bounded monitoring, capture backend TRX/counts, and then run only the two exact frontend filter specs if the job reaches that stage.

Google-only Cloud media checkpoint

User decision:

  • Current Cloud version should use Google:
  • image: Gemini,
  • speech: Google Cloud Text-to-Speech,
  • video: Veo,
  • music: Lyria.
  • Magnific, Beatoven, and ElevenLabs are long-term options only, not current launch requirements.
  • Unavailable features must visibly state that they are placeholders/unavailable and that no model-generated output occurred.
  • Preserve per-user Cloud limits.

Implementation checkpoint:

  • Worktree: C:\Users\fives\source\repos\_wt\workwingman-WT-3379-WING-202-cloud-readiness
  • Branch: wt/WT-3379/WING-202-cloud-readiness
  • Base: e5b1d7f7
  • Commit: 2e29b1fa7487961ffbf3ccbfeb9d625731382664
  • Tree: 30c40b701dcdf06cc5c3f4de6f1ae9e1ea3a58dc
  • Worktree clean.
  • 17 intended files.
  • Focused tests: 42/42 passed.
  • API and CloudDeploy builds: 0 errors.
  • No push, merge, deployment, GCP mutation, billing change, or secret change.

Implemented contract:

  • Google-only server-side provider.
  • Explicit cloud-processing consent.
  • Per-tenant/day atomic quotas.
  • Request prompt limit and configured response limit.
  • Tenant-scoped async Veo operation IDs.
  • Trusted Google HTTPS URI restriction.
  • Credentials remain server-side.
  • Default disabled.
  • YC reviewer audience excluded unless explicitly enabled.
  • Preview/lifecycle disclaimer.
  • Disabled/unconfigured/excluded response says: Pilot placeholder only: this capability is unavailable and no model-generated output occurred
  • It returns no fake media payload.

Current official model/API validation:

  • Image model: gemini-3.1-flash-lite-image.
  • Music model: lyria-3-clip-preview.
  • Video model: veo-3.1-lite-generate-preview.
  • Image and music use Gemini Interactions API.
  • Veo uses :predictLongRunning and operation polling.
  • Speech uses Google Cloud TTS with Cloud Run service-account ADC.
  • These IDs and API shapes were checked against current official Google documentation on 2026-07-26.

Potential blocker requiring review/remediation:

  • GoogleCloudMediaProvider currently materializes the full Interactions JSON/base64 response and the full downloaded video byte array before applying MaximumOutputBytes. That can permit memory/bandwidth exhaustion. Prefer streaming/Content-Length checks plus a hard bounded read.
  • Confirm raw Interactions response parsing against documented steps/output-block shapes rather than relying only on recursive discovery.

Hosted review attempt failed closed because of the Codex sandbox, not because a reviewer rejected the code:

  • Artifact root: C:\Users\fives\source\repos\_review-artifacts\WING-202-security-review-2e29b1fa
  • Regular Jenny prompt hash: e52dfa86114c7ff41ce97a3545a53e6a97f0ef6a52829e4ecbcdcea54fd8ca78
  • Cyber Jenny prompt hash: c24889ee2f6a3f53bb4b50682f9cc461a1917c6ae9533db65380b5dfa5179b95
  • Gronktayvius General prompt hash: 63864467ffa48a6fe7097807a0271872cab78ef484c5141fd666b40fe4a2db63
  • Gronktayvius Red Team prompt hash: 55e2d449551b6489df4d7a888114b1ca3088765183a6628fa995344eacd650b2
  • agy.exe execution was denied by sandbox identity.
  • Grok could not read its auth file in that child context.
  • All seat manifests correctly record prevalidated=false, NO_OUTPUT.
  • No model substitution occurred.

Do not merge or deploy 2e29b1fa until the response-size issue is resolved or explicitly cleared by evidence and all four required seats produce valid output.

Existing private Cloud pilot state

The frozen baseline already includes:

  • invitation/session controls,
  • tenant grant middleware,
  • tenant-scoped document/state storage,
  • server edition endpoint allowlisting,
  • rate-limit partitioning by tenant,
  • Cloud job explorer with tenant-scoped preferences,
  • provider/runtime capability disclosures,
  • deployment planning and health-check structure.

The Cloud Run project previously used in the browser is:

  • GCP project: nifty-condition-503506-j1

No operational Cloud Run deployment has been proven in this handoff. gcloud is not available in the current Codex shell. Earlier Google browser OAuth approval does not substitute for a deployed, tested service.

Before calling the pilot operational, prove:

  1. Invitation creation/redemption and private access.
  2. Authentication and tenant identity propagation.
  3. Cross-tenant storage isolation.
  4. Tenant-partitioned rate/quota enforcement.
  5. Server-side provider routing and secret isolation.
  6. Locked/unavailable features remain locked.
  7. Health/readiness checks.
  8. Container starts with production configuration.
  9. A real second user cannot access the first user’s data.
  10. Audit logs contain no prompt, résumé, token, or secret leakage.

Installer state

Baseline installer only:

  • C:\Users\fives\OneDrive\PC-Bootstrap\_shared\artifacts\WorkWingman\0.2.1-development-924642923\WorkWingman-Setup-0.2.1-development-924642923.exe
  • SHA-256: E43B81E13953387B4D98550FB5AFBF1D9657CE5C0C7E87E632B96F0C288296AD
  • Receipt: same directory, receipt.json
  • Per-user non-admin smoke passed.

This is not the current release candidate. It comes from commit 924642923....

Held candidate script:

  • C:\tmp\wing-candidate-build-e5b1d7f7.ps1

The script requires a Cedric adjudication containing exact bold PASS for the exact commit/tree. Keep that fail-closed behavior.

Required installer sequence:

  1. Finish and commit Grok remediation.
  2. Obtain authoritative tests.
  3. Run all four independent review seats.
  4. Adjudicate exact commit/tree.
  5. Create a clean integration worktree from the frozen baseline.
  6. Integrate only approved commits.
  7. Re-run build/unit/frontend/Electron/package tests on the exact integrated tree.
  8. Run a release-wide security review, not merely another tiny one-commit review.
  9. Build installer from that exact gated tree.
  10. Install per-user without elevation, launch, verify local engine, résumé persistence, updater fallback, Flow/Studio, Degree, Live Run, WING-196, and connections.
  11. Hash the installer, write a receipt, and preserve exact commit/tree/build logs.
  1. Complete and commit the Grok argv/capability remediation.
  2. Get the listener WING-196 test result.
  3. Run/gate the Grok remediation exact commit.
  4. Remediate Cloud media bounded-response and response-contract issues.
  5. Re-run 42 focused Cloud tests and API/CloudDeploy builds.
  6. Run/gate the Cloud commit with all four required seats.
  7. Create a new isolated integration worktree from e5b1d7f7.
  8. Cherry-pick only the approved Grok and Cloud commits.
  9. Resolve nothing by broad conflict acceptance; inspect every conflict.
  10. Run complete backend, frontend, Electron, packaging, and security checks.
  11. Run a release-wide panel against the exact integrated tree/range.
  12. Produce, smoke-test, hash, and publish the installer.
  13. Build the Cloud container from the same approved source tree.
  14. Deploy privately to Cloud Run, with secrets in Secret Manager and a least-privilege service account.
  15. Test invitation/audience behavior and two-tenant isolation.
  16. Only then provide Pam/Shereeba/YC access.

Financial/document communication already completed

AgentMail successfully sent Pam two requested documents:

  • To: pmorristhornton@gmail.com
  • CC: andrewjonesdev1@gmail.com
  • Subject: WorkWingman payment summary and full domain plan
  • AgentMail message ID: <0100019fa190d6dc-3da44ce7-f22a-4c30-bea8-bcc26544d628-000000@email.amazonses.com>

Attachments:

  1. C:\Users\fives\source\repos\Pam-Payment-Executive-Summary-FINAL.docx
  2. C:\Users\fives\source\repos\_budget_pack_output\WorkWingman-Domain-Portfolio-Budget-2026-07-26.docx

Current launch arithmetic given to Andrew:

  • Go-live infrastructure/API funding: $75
  • GCP/Google media allowance: $50
  • Anthropic API starting credit: $25
  • Claude/Fable top-up: $400
  • Codex/ChatGPT top-up: $100
  • Go-live plus immediate top-ups: $575
  • With a $250 launch-domain allowance: $825
  • Conservative hard-cap/domain ceiling communicated previously: $1,050

Google media funding is inside the GCP allowance, not an extra duplicate charge.

Fleet/listener facts

  • Great Room FleetDispatch listener is healthy/green.
  • Direct Codex sandbox calls cannot read C:\Users\fives\.claude\fleet-dispatch.key.
  • Mission Control can submit normal-user jobs.
  • A July 17 orphan record exists in processing, but it does not block new work; WING-196 entered processing alongside it.
  • Do not manually edit queue files.
  • Do not restart tasks or kill unrelated model processes merely to remove that orphan.
  • Bedroom listener was previously timing out.
  • Great Room, gaming, and streaming listeners were previously healthy, but gaming/streaming model seats remain prohibited for review.

Honest current status

Completed/proven:

  • Frozen integrated baseline exists.
  • Critical résumé/updater changes are present in baseline history.
  • WING-196 code is already integrated; no restack is needed.
  • WING-196 authoritative normal-user test job is running.
  • Google media implementation is committed, clean, builds, and passes 42 focused tests.
  • Google model IDs/API shapes are current.
  • Baseline installer exists and smoke-passed, but is not the final candidate.
  • Budget/domain documents were sent.

Incomplete or held:

  • Grok argv remediation is not committed yet.
  • The remediated desktop tree has not passed the required hosted panel.
  • Cloud media has not passed the required hosted panel.
  • Cloud media bounded-response handling needs remediation or strong clearance.
  • A final integrated tree has not been assembled from approved commits.
  • A final installer has not been built, hashed, or smoke-tested.
  • The private Cloud Run YC pilot has not been deployed or proven operational.
  • Final two-tenant isolation, provider dispatch, and end-to-end UI verification remain.

Never represent any held item as complete.

First commands/checks for Clahadore

  1. Inspect all live worktrees without modifying main:

git -C C:\Users\fives\source\repos\_wt\workwingman-WT-3379-WING-202-grok-argv-fix status --short

git -C C:\Users\fives\source\repos\_wt\workwingman-WT-3379-WING-202-cloud-readiness status --short

git -C C:\Users\fives\source\repos\_wt\workwingman-WT-f26b-final-order-assembly status --short

  1. Read the failed exact gate:

C:\Users\fives\source\repos\_review-evidence\WING-202\e5b1d7f7\CEDRIC-ADJUDICATION.md

  1. Review the uncommitted Grok diff, especially exact argv construction, capability probing, cancellation/cache behavior, and every IProcessRunner implementation.

  2. Check WING-196 listener evidence under:

D:\WorkWingman-WING-196-listener-evidence

  1. Review Cloud media commit 2e29b1fa with special attention to bounded HTTP reads, Interactions response parsing, URI validation, quota reservation semantics, and secret handling.

  2. Preserve all exact hashes and do not push/deploy until gates pass.

Definition of done

The objective is achieved only when one exact integrated source tree has:

  • passed the relevant backend/frontend/Electron/package tests,
  • passed the full required review/security panel,
  • produced a per-user installer that launches and survives the critical desktop smoke path,
  • produced a private Cloud Run deployment with invitation/auth/tenant isolation/provider dispatch,
  • proven WING-196 UI/API filtering behavior,
  • proven two-user isolation and limits,
  • produced hashes, receipts, logs, and a requirement-by-requirement completion report,
  • and has no required work remaining.