e1dcc689bc
The autonomous agent's loop prompt had grown to ~9 000 words of accumulated
findings, refutations and traps. That is a symptom, not a style: it was doing
the job the corpus should do, in the one place that dies with the container.
Three concrete failures followed from it, and each gets a structural fix rather
than a louder instruction.
1. IT REPEATED WORK IT HAD ALREADY DONE.
The "do not revive" list and the method traps existed ONLY in the prompt, so
nothing pointed a fresh iteration at them. Extracted verbatim into two
tracked files:
docs/re/REFUTED.md 105 claims tested and dead, grouped by subject so a
grep for your noun finds the neighbourhood
docs/re/METHOD.md the traps already paid for -- controls, inference,
searching, reading data, runtime
Both are linked from INDEX.md, and the loop prompt now opens by requiring
them to be read. This is the fix for "re-derived something already known":
the knowledge is now where the next iteration looks, not in a context window.
2. IT FORGOT TO ARM THE NEXT WAKEUP.
The prompt tried to solve this by shouting at itself in the first line. The
real fix is to stop asking: `loose` now defaults to a FIXED interval (45m),
so the harness owns the cadence and a forgotten ScheduleWakeup cannot end the
run. SYLPH_LOOP_INTERVAL= (empty) restores self-pacing.
3. IT COULD NOT PUBLISH, SO THE WORK ONLY EXISTED IN THE CONTAINER.
New `push-work`, plus a read-only credentials mount
(SYLPH_GIT_CREDENTIALS, default ~/.sylph-git-credentials). It pushes the
CURRENT branch only, refuses anything that is not auto/*, and never
force-pushes -- so the consolidated line stays a human's decision and a
confused iteration cannot rewrite history. The loop prompt now requires a
push on every iteration that commits, rather than at the end of some longer
arc, which is exactly when a container dies.
The prompt itself drops from ~9 000 words to 85 lines and, more importantly,
stops accumulating: findings go in the corpus, and the prompt points at it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Reverse-engineering knowledge base
This directory is the spec-side of the clean-room: it records what the original Project Sylpheed binary does (behaviour) and how its data is laid out, so that the Rust port can be implemented from these specs without re-deriving anything and without ever copying original code.
It exists to answer one question fast: "do we already know how X works, and how sure are we?"
The one rule that matters
Never document a claim more confidently than the evidence supports, and never paste original code here.
A wrong-but-confident note is worse than no note: someone builds on it and the bug hides for weeks. Every entry therefore carries an explicit confidence and its evidence. This mirrors the project method — measure the oracle, never infer; refute before believing.
Clean-room firewall
- ✅ Allowed: behaviour descriptions, field offsets/types, formulas, state machines,
observed input→output pairs, and references to the original by address
(
sub_821B68C0) or toxenia-rs/sylpheed.db. - ❌ Forbidden here and in
crates/: pasted decompiled C/C++ or verbatim disassembled function bodies presented as the thing to reimplement. Cite the address; describe the behaviour in your own words. Disassembly is a tool for understanding, not a source to copy.
Confidence levels
| Level | Meaning | Bar to reach it |
|---|---|---|
CONFIRMED |
Behaviour verified against ground truth. | ≥2 independent observations or one observation cross-checked against an oracle (canary framebuffer, a known-correct value, a second code path). |
PROBABLE |
Strong single-source inference. | One clean observation, or an unambiguous static read of the disassembly. |
HYPOTHESIS |
Educated guess, not yet tested. | Anything else. Must say what would confirm/refute it. |
The status markers
The table above is the confidence scale. The markers that appear in
BACKLOG.md and the structures/ pages are a separate, and until now undefined,
vocabulary. They mean:
| Marker | Meaning |
|---|---|
| ✅ | Confirmed — verified against ground truth. |
| 🟡 | Partial: true as far as it goes, or true under a stated assumption. |
| ❔ | Open question. Nobody has answered it yet. |
| 🔴 | Refuted — shown false — or blocked by something the container cannot do. |
| ❌ | A specific claim that was tried and failed. Prefer 🔴. |
| 🚧 | Work started and not finished. |
🔴 never means "we have not run it yet." That is ❔ or 🚧. Reserve 🔴's "blocked" sense for a real limit of the box — no push credentials, no hardware Vulkan (lavapipe only), or a decision only the user can make. The box can run the emulator, script input, screenshot, read guest memory, and build and test Rust, so "needs a run" is never a blocker. This paragraph exists because the marker was undefined for 98 uses and three of them were mislabelled that way.
Promotion requires new evidence, not re-reading the old evidence. A HYPOTHESIS that
"looks right again" is still a HYPOTHESIS. Only an independent check promotes it.
If evidence later contradicts an entry, demote it and record the contradiction — do
not silently edit the conclusion.
When to document
- Right after a function/structure crosses from
HYPOTHESISto at leastPROBABLE— before moving to the next code path, so the knowledge isn't lost or re-derived. - Whenever confidence changes (up or down) — append to the Evidence log, don't overwrite.
- Not while it's still a pure guess with no evidence — a one-liner in the relevant backlog/plan is enough until there's something to stand on.
What to document
- Functions/code paths →
docs/re/functions/<name>.md(one file per function or tight cluster). - Data structures / formats →
docs/re/structures/<name>.md. - Keep the index in
INDEX.md(one line each: name · confidence · one-line summary).
Use the templates: _TEMPLATE.function.md,
_TEMPLATE.structure.md.
How we find and confirm code paths (the toolchain)
Everything joins on the guest virtual address (PC) — code addresses are fixed by the XEX load, identical across our emulator and canary.
- Static (cheap, try first):
xenia-rs/sylpheed.db(DuckDB: 25 481 functions, xrefs, strings, vtables, imports). Query withxenia-rs/zq.py—zq.py grep <str>,zq.py xref <addr>,zq.py dis <lo> <hi>,zq.py fn <pc>. Entry points are usually a string (zq.py grep MSG_DEMO) or an import (movie/XMA API) xref'd back to the loader. - Dynamic (when static is ambiguous): run
xenia-rswith its probe suite —--pc-probe/--audit-pc-probe-hex(fires at block entry),--mem-watch(mid-block reads/writes of a VA),--lr-trace(call/return chains),--trace-instructions,--dump-addr(read guest memory). These already exist; prefer them over hacking canary. - Oracle (correctness ground truth): canary — the Wine cross-build
xenia-canary/build-cross/bin/Windows/Debug/xenia_canary.exe(the native Linux ELF crashes / does not run — do not use it). This is the only emulator that reaches the in-game menu; ourxenia-rsnever got past the intro video. Use canary to observe output (capture its framebuffer for texture colours), not usually to instrument code — though itsbuild-crosstoolchain does compile, so small C++ probes + rebuild are possible when needed. Run muted, one emulator process at a time, point it at the real ISO (not the symlink).
⚠️ VA-equality caveat: join code by PC (fixed), but never assume a data VA holds the same bytes across emulators — allocators differ. Compare data by content/layout.