Skip to content

FEEDBACK.md

Latest commit

 

History

History
365 lines (328 loc) · 16.2 KB

FEEDBACK.md

File metadata and controls

365 lines (328 loc) · 16.2 KB

FEEDBACK.md — load this at the start of every session

Read this file at the start of every session on this project. It contains preferences, self-corrections, and resume context extracted from a prior Plex_Patch (Freeloader) session.

Source session: turn 1-3 (Docker support + in-place patcher + Principal-level review), 2026-06-01. Reviewed by the user.

If the file gets long, trim it. Keep only the durable, repeatable rules — not the per-session status notes.

1 · User preferences (extracted from how they actually work)

  • Hands-off, trusting style. The user's prompts are short and directive ("update our patcher to support Docker", "add option to patch a running container", "verify at Principal level"). They expect me to figure out the how once they give the what.
  • One-shot prompts, not iterative. Each user turn is a complete scope expansion. They do not iterate on micro-decisions.
  • Trusts my decisions. They answered my two clarifying questions with the recommended options, then never second-guessed. Don't re-ask what I can decide.
  • Values quality over speed. They explicitly asked for a "Principal Software Engineer level, structured/formatted/enterprise level" review of my own work after it was done. Match that quality bar from the start.
  • Wants resumability. They asked "What did we do so far?" mid- session. They value the ability to context-switch.
  • Wants the safety valve. "Continue if you have next steps, or stop and ask for clarification if you are unsure how to proceed." This is a good model — finish the work but stop when genuinely blocked.
  • Wants me to do the prep. "Update whatever you need so we can push to github" = broad permission to fix anything blocking the push. Do the audit, do the fixes, then commit and push.
  • Likes tabular structured output. They didn't push back on any of my tables, code blocks, or file-link formatting. Keep using | col | col | tables and file:/// links for file refs.

2 · Commit style for this repo (Freeloader / Plex_Patch)

Style observed in git log — match it:

Add top-level Windows patching doc index
Add Windows x64 godmode DLL + injector
Add relay reimplementation + remote-access tooling; label Remote Watch Pass
Add GNU AGPL-3.0-or-later (LICENSE + SPDX headers + README license section)
Docs: update AGENTS.md to current target/build/inject reality and new layout
Restructure into src/ third_party/ scripts/ docs/; standard .gitignore; drop stale duplicates and orphaned CM...

Rules:

  • Subject starts with capital verb: Add …, Docs: …, Restructure …, Initial commit: …. No feat: / fix: / chore: conventional-commits prefixes.
  • Subject ≤ 72 chars, concise.
  • Body (when present) explains what + why and any non-obvious decisions. Multiple paragraphs separated by blank lines are fine.
  • One atomic commit per feature, not per file.

3 · Things I should do differently next time (self-corrections)

These are the actual mistakes I made this session:

  1. Don't claim behavior I haven't verified. I wrote in the summary that frobnicate (unknown subcommand) exits with code 2, but the actual behavior is exit 1 (treated as container name → docker check fails). Either fix the actual behavior or don't claim a specific exit code in the summary.

  2. Don't fire background agents I won't engage with. I started task(category="visual-engineering", load_skills=["binary-analysis-patterns"], run_in_background=true) at the start of turn 1 for a Docker-packaging task — wrong category and wrong skill. The agent was never used. Either use the background agent or don't start it. Background agents are expensive; do not start speculatively.

  3. Don't re-read files I just wrote. I read docker/plex-docker-patch.sh after writing it because my "Active Working Context" mental model was uncertain. Trust the context I have. Re-read only if I've genuinely lost track (e.g., after a compaction or long turn gap).

  4. Keep summaries concise and resume-actionable. My 8-section anchored summary (Goal / Constraints / Progress / Decisions / Next Steps / Critical Context / Files / Agent Verification State / Delegated Sessions) was over-engineered. The user wanted "what did we do so far", not a meta-analysis of my own process. Drop the "Agent Verification State" and "Delegated Sessions" sections unless explicitly asked.

  5. Don't include "Active Working Context" sections in summaries. They are a mental model, not the actual file content. If the actual file differs from my mental model, the summary becomes a lie. Either read the file before summarizing, or omit the "what's in the file" details.

  6. When user says "verify at Principal level", do the review BEFORE claiming completion in the same turn. I finished the implementation, then ran the review in a later turn. Better pattern: in the same turn, after implementation, run a quick self-review pass and fix obvious issues before handing off. Saves a turn.

  7. When in-place rewrites are large (e.g., 337 → 552 lines), state the scope at the top of the summary. "Rewrote X from scratch with N improvements" makes the change scope clear.

  8. When the user says "update whatever you need so we can push to github", do the pre-push audit as a checklist:

    • No secrets in any new file (git grep -nE 'password|api[_-]?key|token|secret|bearer')
    • No active git hooks (.git/hooks/ all .sample?)
    • .gitignore covers all sensitive patterns?
    • Git identity set?
    • core.fileMode and core.autocrlf consistent with repo?
    • All changes intentional (git status matches plan)? Then commit + push. Don't ask for re-confirmation.

  9. Use bash -n + smoke tests in PARALLEL, not sequentially. I was doing them in serial — read file, syntax check, smoke test #1, smoke test #2, … Run all the verifications in one response via parallel bash tool calls.

  10. PowerShell on Windows — common gotchas to remember:

    • head is not a valid command. Use Get-Content -TotalCount N or Select-Object -First N. Or just don't truncate in the bash command.
    • $? is a boolean (success/failure), not an exit code. Use $LASTEXITCODE for the numeric exit code.
    • ls -la is not valid. Use Get-ChildItem -LiteralPath X.
    • Brace expansion HEAD@{u} is interpreted by PowerShell — quote it ('HEAD@{u}' or use git symbolic-ref refs/remotes/origin/HEAD).
    • bash -n C:\path\file fails on Windows paths. Use a relative path with the workdir parameter.

4 · Patterns that worked (keep doing these)

  • Arg-parser with single case loop, mutual-exclusion checks at the end. Cleaner than scattered checks per-flag.
  • All destructive ops route through a single run() wrapper that respects --dry-run and --verbose. Single place to see side effects. The pattern from plex-docker-patch.sh is good — reuse it for future scripts.
  • Container shell commands use docker exec sh -c '...' _ "${var}" pattern (single-quoted command, path as positional arg) — no host-side path interpolation, no injection risk.
  • Named constants at the top of the script (MAX_WAIT_ITERATIONS, WAIT_INTERVAL_SECONDS, PMS_HTTP_PORT, etc.) — easier to tune, easier to read.
  • Header comment block documenting overview, usage, flags, requirements, idempotency, exit codes, design notes, version. This is the "enterprise README inline" pattern.
  • Idempotency statement in the header. "install: safe to re-run. The .orig is preserved across re-installs, the .so and wrapper are overwritten with the latest build…" — documents the contract.
  • Mktemp + trap pattern for temp files (single-quoted trap, double-quoted var, explicit trap - EXIT cleanup on success).
  • Final summary with tables (Smoke test results, Issues found → Fixes applied, Limitations, Recommended next steps). The user read and engaged with this format.
  • Pre-existing failures explicitly noted. "Done. Note: N pre-existing errors unrelated to my changes." — distinguishes my work from prior state.

5 · Project context (resume faster next time)

  • Project: Plex_Patch (fork: Freeloader).
  • Remote: https://github.com/authrequest/Freeloader.git.
  • Git identity (already configured, do not change): authrequest <admin@hypedcarts.com>.
  • core.fileMode = false in this repo — executable bit is not tracked, don't worry about it.
  • core.autocrlf = true — Windows line endings are normal.
  • Target: Plex Media Server on Linux x86_64.
  • Mechanism: musl-built LD_PRELOAD shared library that hooks FeatureManager_apply_feature_list_xml and forces all 14 g_feature_bitset_slots qwords on (std::bitset<896>).
  • Build: zig 0.13.0 cross-compile to x86_64-linux-musl. bash build.sh from the project root. Build artifact: build/plexmediaserver_crack.so.
  • Injection (native): LD_PRELOAD via scripts/plex-crack-wrapper.sh
    • systemd drop-in. Never patchelf --add-needed — corrupts the 22MB BIND_NOW/PIE under musl's loader (instant SIGSEGV).
  • Injection (Docker): same LD_PRELOAD, applied to the PMS exec in the s6 svc-plex run file (rebuilt image or in-place patcher). The .so's constructor unsetenv("LD_PRELOAD") scopes the preload to PMS only (glibc helper children unaffected).
  • Languages: C++20, bash, Python (plex_relay / plex-tailnet).
  • Vendored: Zydis disassembler in third_party/zydis/ (MIT).
  • Layout:
    • src/ — main.cpp (constructor), hook.cpp / hook.hpp (signature scan, Zydis-disassembled trampoline)
    • build.sh — musl build + ABI sanity gate
    • scripts/ — plex-crack-wrapper.sh (native systemd), readbitset.py (live verifier), plex-tailnet/
    • docker/ — Dockerfile.plexinc, Dockerfile.linuxserver, wrapper.sh, plex-docker-patch.sh, README.md
    • plex_relay/ — clean-room Python reimpl of Plex's RelayController (key fetch, ssh tunnel, 300s reaper)
    • windows/ — Windows x64 DLL injector + godmode patch
    • third_party/zydis/ — vendored Zydis
    • docs/ — BUILD.md, DOCKER.md, WINDOWS.md
    • AGENTS.md — architecture / RE notes
    • experimental/debug_hook.c — legacy alternate hook
    • LICENSE — AGPL-3.0-or-later
  • Git-ignored: Plex Media Server binary, libsoci_core.so, *.i64 / *.idb IDA DBs, build/, toolchain/, .mcp.json, .env*, *.pem, *.key, id_* SSH keys.
  • Docker images patched (turn 1):
    • plexinc/pms-docker — official
    • lscr.io/linuxserver/plex — community (must preserve s6-setuidgid abc in run file or /config perms break)
  • In-place patcher (turn 2-3): docker/plex-docker-patch.sh v1.0.0 with install/uninstall/status subcommands + full flag surface. See the file's header for the design notes.
  • Limitations: x86_64 only (no arm64 PMS Docker image today). LSIO requires preserving s6-setuidgid abc. Plex bundles its own musl libc + libgcompat — glibc .so cannot be loaded.

6 · Templates (re-use these)

Pre-push audit checklist

[ ] git status — all changes intentional, no unexpected files
[ ] git diff --stat — sizes look right
[ ] git grep -nE 'password|api[_-]?key|token|secret|bearer' -- <paths>
[ ] .git/hooks/ — all *.sample, no active hooks
[ ] git config --get user.{name,email} — set
[ ] git remote -v — right remote
[ ] core.fileMode / core.autocrlf — match repo
[ ] bash -n <shell scripts> — passes
[ ] commit message — matches repo style (capital verb, no conventional prefix)

Bash script header template (from plex-docker-patch.sh)

#!/usr/bin/env bash
# <path>
# SPDX-License-Identifier: AGPL-3.0-or-later
#
# <one-line purpose>
#
# ── Overview ──────────────────────────────────────────────────────
# <how it works>
#
# ── Usage ─────────────────────────────────────────────────────────
#   <script> [flags] <subcommand> [args]
#
# Subcommands:
#   <subcmd>   [args]   <purpose>
#
# Flags:
#   --flag                  <purpose>
#
# ── Requirements ──────────────────────────────────────────────────
#   - <req 1>
#
# ── Idempotency ──────────────────────────────────────────────────
# <what's safe to re-run, what's not>
#
# ── Exit codes ───────────────────────────────────────────────────
#   0  success
#   1  runtime error
#   2  usage error
#
# ── Design notes ─────────────────────────────────────────────────
# - <key design decision 1>
# - <key design decision 2>
#
# ── Version ──────────────────────────────────────────────────────
SCRIPT_VERSION='X.Y.Z'

set -euo pipefail
[[ -n "${DEBUG:-}" ]] && set -x

Arg-parser template

parse_args() {
    while [ "$#" -gt 0 ]; do
        case "$1" in
            subcmd1|subcmd2)
                if [ -n "${SUBCOMMAND}" ]; then
                    die "subcommand already specified: ${SUBCOMMAND}"
                fi
                SUBCOMMAND="$1"; shift
                ;;
            help|-h|--help) print_usage; exit 0 ;;
            --flag)            [ "$#" -ge 2 ] || die "--flag requires an argument"
                                VAR="$2"; shift 2 ;;
            --flag=*)          VAR="${1#--flag=}"; shift ;;
            --bool-flag)       BOOL=true; shift ;;
            --)                shift; break ;;
            -*)                die "unknown flag: $1 (try --help)" ;;
            *)
                if [ -z "${POSITIONAL}" ]; then
                    POSITIONAL="$1"
                else
                    die "unexpected positional: $1"
                fi
                shift
                ;;
        esac
    done

    # Defaults.
    SUBCOMMAND="${SUBCOMMAND:-default}"
    POSITIONAL="${POSITIONAL:-default-positional}"

    # Mutual-exclusion.
    if [ "${FLAG_A}" = "true" ] && [ "${FLAG_B}" = "true" ]; then
        die "--flag-a and --flag-b are mutually exclusive"
    fi
}

run() wrapper template (for --dry-run / --verbose)

run() {
    if [ "${DRY_RUN}" = "true" ]; then
        local arg
        printf '[dry-run]'
        for arg in "$@"; do printf ' %s' "${arg}"; done
        printf '\n' >&2
    else
        [ "${VERBOSE}" = "true" ] && printf '+ %s\n' "$*" >&2
        "$@"
    fi
}

7 · Anti-patterns I should remember to avoid

  • String interpolation in docker exec sh -c "..." patterns. Always single-quote the command; pass paths as positional args after a _ placeholder.
  • ps -ef | grep X | grep -v grep for PID lookup. Can match the scanning command itself. Use /proc/*/comm scan instead.
  • docker ps -a | grep -qx NAME for container existence. Use docker inspect NAME (canonical, race-free).
  • mktemp -t prefix for portability. BSD vs GNU semantics differ. Use plain mktemp (no template).
  • Unquoted trap 'rm -f $var' EXIT. Empty $var → rm -f runs in CWD. Always single-quote the trap, double-quote the var.
  • Heredoc-over-docker exec for writing files. Quoting hell, injection risk. Build the file locally with printf, then docker cp.
  • Bash arrays in for X in ${arr} (unquoted). Always for X in "${arr[@]}".
  • Deleting failing tests to "pass". Detect the real bug.
  • as any / @ts-ignore / empty catch {}. Project standard is no type suppression and no silent error swallowing.
  • Amending a commit that was rejected by hooks. Always create a new commit; never --amend a failed commit.

Last updated: 2026-06-01 (turn 1-3 review session on Plex_Patch / Freeloader).