[Cadacious]

Summary

Plan-only PR for the ammo campaign (#1026). Adds docs/analysis/ammo/:

• README.md — scope, the D-AM01..06 decisions table, open questions (including AM-01's findings and a flagged design decision for the owner), packet status, the parallelization plan, a UAT checklist outline, and known issues.
• audit.md — current-state audit with file:line citations: the bandolier/reload model, requestAmmoChange's validation gap (and how it relates to open PRs fix(inventory): reject requestAmmoChange when weapon def is not cached #602/docs(inventory): cite the declared sgw_inventory PK in bandolier TOCTOU comments #603), the damage path's ammo-type blindness, items.sql ammo stats, candidate pistol/SMG families to widen, the cooked-data push mechanism, and the GM-tooling gap.
• work-packets.md — restructured, at the maintainer's request mid-task, around one small serial gate (AM-F: reserved item ids, the ammo_item_types/ammo_modifiers contracts, a real+tested AmmoReserve module, pre-split hotspot files, a feature flag) feeding Wave 1 (6 packets in parallel: AM-02 reserve, AM-03 validation, AM-04 damage framework, AM-05 loot, AM-06 GM tooling, AM-07 client push) and Wave 2 (6 packets in parallel, one per remaining ammo family, gated only on AM-04). A file-ownership matrix pins which files each packet may touch so none collide.
• handoffs/session-resume.md — resume note for the next coordinator session.

Also updates docs/readme.md's campaign index.

AM-01 landed mid-task

Packet AM-01 (reverse engineering, PR #1040) merged into main while this plan was being written; its findings are folded in throughout:

• No finite-ammo reserve model exists anywhere in the 2009 client schema — D-AM01 (items in bags) is recorded as a restoration design choice, not a recovered mechanic.
• Widening a weapon's ammo_types needs no client push — the client reads a live, server-populated container cache, not cooked data. This narrowed AM-07's scope to only the 15 new ammo item definitions.
• Toggle abilities 715/719 are architecturally independent of requestAmmoChange — flagged as an explicit open question with a recommendation (apply the damage modifier directly, never cast the ability) for the owner to confirm before AM-04 ships.
• /gmgiveammo's native wire event has no recovered server-side receiver, so AM-06 now ships as a .-console command instead of a native-opcode handler.
• New-item-id cooked-data injection is architecturally sound but unproven for a genuinely new id — AM-07 now spikes one item and UATs it live before batch-authoring the rest.

New-item-id feasibility (what the plan needed to settle)

Contained by design: every packet after AM-F resolves an ammo item through the ammo_item_types mapping table by EAmmoType, never a hardcoded item id. If AM-07's spike fails, only two seed files need repointing at existing placeholder ids — no other packet's code changes.

Test plan

• markdownlint (warn-only) — spot-checked headings/links, not run through the full tool
• Docs-only change; no code, no runtime behavior — no unit/live-DB/wire-format tests apply per TESTING.md
• Verified every internal cross-reference anchor (README.md#..., audit.md#..., work-packets.md#..., the RE finding doc's #q1..#q5) resolves to an actual heading slug

🤖 Generated with Claude Code

https://claude.ai/code/session_01Wgx9Q5HHQtsvQ1tyaRraqz

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: Repository: SandboxServers/Cimmeria/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: d98f635b-f781-4136-96a6-b37828dc964c

📥 Commits

Reviewing files that changed from the base of the PR and between d7b9185 and 29be595.

📒 Files selected for processing (7)

• .claude/agent-memory/items-systems-advisor/MEMORY.md
• .claude/agent-memory/items-systems-advisor/project_ammo_campaign_plan.md
• docs/analysis/ammo/README.md
• docs/analysis/ammo/audit.md
• docs/analysis/ammo/handoffs/session-resume.md
• docs/analysis/ammo/work-packets.md
• docs/readme.md

 ____________________________
< Bunny-grade bug detection. >
 ----------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Cadacious]

Token profile for this PR. Estimated list-price USD, a plan-usage proxy, not a bill (D-TP1).

est. USD (list)	requests	human prompts	agents	peak ctx	wall clock	CI rounds
$7.45	90	0	1	391,252	0.4 h	0 failed of 2

Attribution: worktree, confidence 0.94; 0.0% of the same sessions' spend over this PR's window is unattributed. Review rounds: 0. Follow-up fix PRs: none. Reverted: no.

Claude Code's own cost-state total is per process and is not split by PR; over the 1 sessions this PR's spend came from (100% of it), the profiler was 15.5% under it, so read the USD as a floor.

Profiler 36225a9, Claude Code 2.1.283, price table 2026-10-03. How this is measured

{"schema":"cimmeria-pr-stats/1","pr":1041,"profiler":"36225a9dd1ee","claude_code":"2.1.283","price_table":"2026-10-03","window":{"first":"2026-09-28T22:38:08.772Z","merged":"2026-09-28T23:04:59Z"},"attribution":{"method":"worktree","confidence":0.944,"unattributed_share":0.0},"usd_est":7.4493,"by_model":{"claude-sonnet-5":6.4144,"claude-opus-5-5":1.0349},"tokens":{"input":182,"output":113649,"thinking":35900,"cache_read":26061700,"cache_write_5m":359516,"cache_write_1h":16822},"requests":90.0,"turns":4,"human_prompts":0,"event_triggers":{"background_completion":1},"agents":{"items-systems-advisor":{"count":1,"requests":80.0,"usd_est":6.4144}},"context":{"peak":391252,"p50":311725.5,"compactions":0,"idle_gap_write_share":0.0058},"tools":{"bash_chars":133621,"read_chars":207145,"top_exposure":[{"tool":"Bash","fingerprint":"grep","exposure_chars":3744608},{"tool":"Read","fingerprint":"docs/analysis/bank-vault/work-packets.md","exposure_chars":2999724},{"tool":"Bash","fingerprint":"sed -n","exposure_chars":2809527},{"tool":"Read","fingerprint":"docs/analysis/bank-vault/handoffs/session-resume.md","exposure_chars":2424006},{"tool":"Read","fingerprint":"docs/analysis/bank-vault/README.md","exposure_chars":2236728}]},"quality":{"ci_fail_rounds":0,"review_rounds":0,"followup_fix_prs":[],"reverted":false,"ci_rounds":2},"diff":{"additions":588,"deletions":0,"files":7},"campaign":null,"cost_state":{"usd":null,"reason":"Claude Code's cost-state total is per process, and a process spans several PRs, so it is not split by PR","sessions":1,"covered_share":1.0,"sessions_gap_share":-0.1547}}