Files

Sylpheed RE agent 797765a0b2 agent: move the loop prompt's memory into the corpus, and let it push

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>

2026-08-28 15:46:59 +02:00

5.1 KiB

Raw Blame History

Method traps already paid for

Each line cost an iteration at least once. They are general — they are not about Sylpheed, they are about how this kind of measurement goes wrong.

Like REFUTED.md, this list had been living in the autonomous agent's loop prompt, i.e. nowhere durable. See README.md for the ✅/🟡/❔ confidence convention itself.

Controls

Every result needs a control. A control that fails kills the instrument.
Run the known-positive through a new filter FIRST. Three filters have been killed by their own control. When one fails, read the known-good's disassembly before assuming a shape.
A measured negative is a result — but a negative is only as strong as the route you ran, so state its reach.
A null result needs its cause shown to have happened.
A result with NO unknowns is suspicious.
Census the whole set; always run the other population as the control. Zero partials is stronger than a majority.
A 2×2 partition is the sharpest general tool — both off-diagonals empty is a law.
Re-derive a doc's own numbers as the control.

Inference

Never conclude from ONE sample.
A law proved on one population is a hypothesis on the next.
Finding one exception does not imply a family.
Consistency is not proof. A suggestive coincidence is a coincidence until measured. An analogy is not a measurement.
Same layout ≠ same instance. Same record-name set ≠ same object.
A marker is only proven by what it leaves out.
A high-confidence SCORE is not a high-confidence MECHANISM.
Knowing HOW MANY is not knowing WHICH.
Round numbers matching is weak evidence — unless you read the constant.
My own last-turn result is a hypothesis too.
A global partition can understate a per-owner one.
A residual is measured against a population — name it.

Searching and tooling

A search that returns thousands has no power; state the reach.
A substring match is not a hit. A regex miss looks like a null result — print one raw sample before believing a zero.
A derived table can be a cross product — measure its shape first.
After refuting an instrument, sweep everything that depended on it.
Before measuring how wrong a tool is, read what the tool actually does.
The instrument must pass its own control.
Classify a bulk before mining it. The residual is the prize.
Rank by similarity — the cliff is the finding. But read the values before trusting the rank.
Grep the nouns before designing the experiment — and believe it.
Grep gives you a file list — read every file on it.
The answer is often already in the doc that owns the subject — read it end to end. A 🟡 often names its own route.
Before re-trying a blocked idea, check whether the blocker's own doc already tried it.
Ship a regenerator with every artefact. An artefact that moves by a pure reorder is a tool bug.
Never print per-entry lines from a disc-wide sweep — aggregate.

Reading the data

Read what a loader NAMES, not where it stores.
A field the disc never values still gets named by the loader.
An indexed read beats a deduped-pool adjacency read.
Check the whole string set, not the one matching word.
A dict keyed by record name across a multi-entry pak is a lie.
A set-difference over names hides reuse — join per USER.
A self-index names records, not files.
Case-insensitive hashing means two spellings can be one entry.
An "unresolved" name may be the wrong kind, namespace or prefix — or part of a cut asset.
A garbled value may be a real string in another encoding.
Two of my own counts disagreeing is a grammar clue.
A game's own typo is a join key.
A bias constant in the code is a join key.
A prefix trap: enumerate maximal [A-Za-z0-9_] runs, not startswith.
Re-deriving a format is not a finding — asking whether its values resolve is.

Mechanics that have bitten

Never hand-convert a decimal VA — print hex().
grep -c counts LINES — use grep -o | wc -l.
Counter.most_common() tie-breaks by insertion order — use sorted().
Raw grep cannot see inside compressed pak entries.
Commit messages go in a file (git commit -F); a literal | in a table cell needs escaping; git log --all -- <path> can hang.

Runtime / emulator

Look at the PNG — and check its dimensions.
"Animating" is not "still in a mission".
Dedup entity enumerations by position value.
Do not diagnose timing or liveness under gdb. ps %cpu is cumulative.
Classify screens by whole-image statistics, not named pixels — a named pixel is only valid while the image sits at a known place, and nothing errors when it moves.
Do not poll faster than the guest updates — it manufactures a clean curve out of noise.
A probe that never performs the action will "prove" the action does not exist.

5.1 KiB Raw Blame History Unescape Escape