# Agent workflow refinement — decisions and evaluation

Date: 11 October 2026. Scope: implemented documentation/policy refinement with the
verification and limits recorded below. This record does not
establish gameplay quality, faster agent execution or any new accepted game checkpoint.

## Authorized design and mini-plan

The user requested implementation of the research recommendations, then explicitly
selected: high-risk slices use three independent technical/gameplay/design-UX reviews
before and after; medium uses relevant specialists; low uses lead inspection with
escalation. Isolated prototypes get focused checks, with normal integration gates on
promotion and separate human acceptance. Evaluation is documentation checks and read-only
walkthroughs only; real coding/gameplay trials are deferred.

Lead owns all writes. Three independent read-only pre-reviews ran on `gpt-6.1-sol`,
medium reasoning, clean contexts: technical, gameplay and design/UX. They recommended
the split with concrete preservation, precedence, risk and wiki-link safeguards.
No implementation agent, Unity operation, game test/build or player launch is needed.

1. Preserve the original working-copy plan byte-for-byte in a clearly historical
   same-directory archive, keeping its relative links and original heading anchors.
2. Make AGENTS.md a shorter entry point preserving project-specific invariants and
   routing to current brief, design/review workflow, exact testing and setup.
3. Retain accepted requirements/roadmap in the current brief, replace accumulated
   chronology with linked current evidence and keep unfinished acceptance prominent.
4. Add one compact experiment/review/verification packet, concrete risk thresholds,
   prototype isolation/promotion rules, model policy and instruction-evaluation method.
5. Add contextual testing/setup/wiki routes and a prospective amendment to the active
   pack plan. Preserve existing test gates, commands, contracts and historical approvals.
6. Check preservation, links/anchors and wiki output; review the actual integrated
   documents independently; reconcile actionable findings and stop.

## Verification packet

- Changed boundary: agent instructions and documentation navigation/chronology only.
- Files: AGENTS.md; PROJECT_PLAN.md; new PROJECT_HISTORY.md, AGENT_WORKFLOW.md and this
  record; TESTING.md; SETUP.md; active pack-plan amendment; wiki development route and
  refreshed linked source/provenance output. No game, tool implementation, package,
  scene, asset or setting changes are planned.
- Focused proof: compare the archived payload with the pre-task plan SHA-256; preserve
  all prior headings; compare testing body/commands/defaults; inspect policy invariants
  and walkthrough routing; resolve local links/anchors; compare pre-task file hashes.
- Integrated gate: existing documentation test/integrity command and static wiki build/
  output-link check because the source route/exported text changes. No Unity gates for
  prose. No new tests that merely duplicate the wording.
- Baseline: unique directory
  `TestResults/agent-workflow-20261010T143030Z-7dcf9081` records revision and hashes of
  1,755 existing tracked/untracked files, plus original AGENTS/plan/testing bytes.
  Its UTC directory date is 10 October; the user-facing Sydney task date is 11 October.
- Reuse: existing game evidence remains historical with its own identity; none is
  called fresh or combined into a new accepted checkpoint by this documentation task.
- Duration: unknown; bounded to one documentation slice and relevant delta checks.
- Escalation: changed runtime/assets/settings/tool implementation, broken evidence
  identity, lost requirements/anchors, reviewer disagreement over material intent,
  or a need for live trials. Do not expand scope or launch Unity to repair prose.
- Stop: required policy and routing implemented, preserved history/invariants verified,
  documentation checks pass and three final perspectives have no unresolved blockers.

## Reconciled design decisions

- Review risk and test scope are independent. Calling a change small, visual or a
  prototype cannot lower authority/lifecycle/progression/persistence risk.
- Keep three perspectives for consequential work; use relevant independent reviews
  for bounded behavior and lead review for small mechanical changes. A perspective
  checklist is not a claim of independent review when only the writer inspected it.
- Keep technical/rendered/network/human/hardware outcomes separate. Solo prototype
  results cannot establish four-player balance. "Keep for further testing" avoids
  treating an interesting experiment as accepted feel/balance.
- A shared task packet serves design, review and verification; no mandatory duplicate
  spec/plan/evaluation forms for every minor task.
- The root and current plan link one project model policy. Its explicit precedence
  resolves the earlier global default for this project; the user's global file and
  other projects are untouched.
- Current risk policy supersedes generic review defaults in old plans prospectively.
  Specific feature reviews/contracts, completed approvals and test/acceptance gates
  remain binding. The active pack plan has a conspicuous amendment.
- PROJECT_HISTORY.md includes the entire original plan as an exact byte payload after
  an archival marker. Keeping it under Docs avoids relative-link rewrites. Current
  plan headings remain for existing bookmarks, with chronology forwarding links.
- Wiki snapshots require explicit source links; new workflow/history/evaluation
  references are exposed in the authored development page without changing the collector.

## Read-only representative walkthroughs

These are policy-routing inspections, not executed experiments, independent agent
performance trials or proof of game behavior. They are reviewed against the final
documents before publication; the expected selections below do not authorize execution.

| Task / condition | Reading and review route | Proof / escalation / stopping rule |
|---|---|---|
| HUD wording/link correction | Root instructions plus affected copy; low risk, lead inspection. | Link/text inspection; no Unity suite by default. Escalate if wording changes an actual mechanic/requirement. Stop after correction and focused checks. |
| Local recoil-cue experiment | Current brief and workflow; bounded presentation is medium, relevant gameplay and UX review before/after. | Preserved isolated baseline, rendered cue inspection and explicit human responsiveness question. Human feel remains pending without playtesting. Shared fire timing/damage changes escalate; promotion requires affected normal gates. |
| Weapon damage tuning | Gameplay requirements, stat/proc rules, workflow/testing entries; medium only if interactions are understood and bounded, gameplay review plus technical when calculation/proc boundaries are affected. | Inspect solo/four-player kill-time and synergy assumptions, use relevant rules/native proof for implementation. Broad proc/refund/defense/encounter effects escalate high. Stop at bounded result with balance uncertainty explicit. |
| Settings/onboarding layout | Existing UI/input flow, workflow and relevant catalogue; medium, UX and technical reviewers for local focus/input behavior while preserving existing settings storage contract/schema. | Render full open/focus/close/return flow, keyboard cues and aspect ratios; human comprehension is separate. Shared pause, session lifecycle or storage contract/schema changes escalate high. Promotion retains affected UI/input/network gates. |
| Launcher/report sidecar | Workflow, TESTING and catalogue entries; medium technical review only if additive and ownership/identity/cleanup contracts stay unchanged. | Controlled parser/report proof and representative affected real cohorts for later implementation. Changing ownership/cleanup/identity lifecycle escalates high. No automatic unrelated progression matrix; omissions/reuse must be explicit. |
| Small RPC/remote movement fix | Current contracts, workflow and exact movement/authority entries; high, all three independent pre/post reviews. | Focused credit/input/lifecycle proof plus affected suites, two/four-owner, warp/lifecycle and normal-stage gates. Line count does not reduce scope. Human/hardware/separate-network acceptance stays separate. |
| Portal/reward "UX simplification" | Accepted reward/stage requirements and affected plan; high despite visual framing, all three reviews. | Preserve participant identity, rewards, progression, input/lifecycle and recovery. Cannot replace approved behavior through prototype shorthand. Full affected integration plus rendered flow; material player-facing requirement changes need user resolution. |
| Keeper primitive prototype | Roster plan and current defense/spawn/state contracts; high, all three perspectives. | Placeholder art does not isolate shared protection, spawn or lifecycle effects. Preserve bounds/authority, satisfy affected normal gates before production promotion, human balance remains open. |
| Missing Natural arrival cells | Current Task5 record and testing campaign procedure. Review retained evidence first; any production repair classifies by actual boundary. | Preserve 19 attempts / 44 captures, four incomplete cells, natural16 gap, contact warnings and unchanged cumulative caps. No new attempts, forced selection or rerolls follow from this walkthrough. Stop with truthful incomplete/manual targets. |

## Verification results and limits

- **History/source preservation:** the original plan payload exactly matches the captured
  working-copy bytes (SHA-256
  `7085a4867cce33815445063f8397beb9931adebb5850388681345e5bec8d2347`).
  All 21 original heading anchors remain in the current plan. TESTING.md is unchanged
  except its new routing paragraph; existing gates, commands and defaults are preserved.
  Baseline comparison found only the six intended existing documentation files changed
  and three intended new Markdown files, with no unexpected tracked/untracked changes.
- **Focused link inspection:** the first attempt caught an incorrect debug-arena evidence
  filename introduced in the brief. Corrected it to the existing DEBUG_TEST_ARENA.md;
  the retained second inspection passes with no missing local links/anchors. Both results
  remain in the evidence directory; the initial failure is not discarded.
- **Existing documentation checks:** `Tools/Wiki/Test-Wiki.ps1` passed all 30 tests
  (11.284 seconds for the test suite), followed by 91-entity source/hash/generated-content/
  identity/link/anchor integrity. No new wording-mirroring tests were added.
- **Static wiki output:** `Tools/Wiki/Build-Wiki.ps1` passed strict generation and built
  links/anchors/image/script/style targets across 122 HTML pages. The first build reports
  6.23 seconds for static generation; this is documentation timing, not agent/game speed.
  An installed-theme advisory and three generated index pages outside nav were reported;
  the pinned build and target checks passed without changing dependencies.
- **Independent integrated review:** technical, gameplay and design/UX perspectives
  approve, with no unresolved blockers. Technical also independently verified exact
  archive bytes. UX copy suggestions were adopted: discretionary delegation is explicitly
  beyond required reviews, and the medium settings example preserves its storage contract.
- **Walkthroughs:** all nine representative routes were inspected against the integrated
  policy and independently reviewed. They cover low/medium/high changes, prototype
  promotion, misleading visual/small-task labels and incomplete Natural evidence.
  No live prototype, rendered game UI observation or human playtest occurred.
- **Final evidence:** the unique directory retains `inspection-initial.json`,
  `inspection-after-link-repair.json`, `inspection-final.json`, the reusable task-local
  `inspect_docs.py`, `wiki-build.txt`, `wiki-build-final.txt` and `final-manifest.json`.
  Final publication rechecks the changed snapshots/output after this evidence update;
  the manifest binds final document hashes and separately identifies the earlier checks.
- **Scope/limits:** no Unity/Editor/gameplay/player/network test or game build was run;
  game source/assets/settings/packages, tool implementations, the user's global AGENTS.md,
  existing failures and campaign caps remain unchanged by this task. No commit or push.
  Human feel/readability/balance, representative hardware, separate-network acceptance,
  Task5 gaps, permanent unlock persistence and roster roadmap remain open. Workflow
  savings/quality are unmeasured; real comparison trials remain deferred by user choice.