recurring patterns#

lessons that keep re-earning themselves across the git history. read this before "fixing" something that looks accidental — it may be a deliberate answer to one of these.

subtraction is a first-class move#

built, evaluated, and deleted: the approval system, curiosity queue, feed scanner, tag/episodic memory layer, voyage migration, observation lexicon, LLM-based workflow classification, the process rail. when a subsystem stops earning its complexity, remove it fully (code, lexicon, docs) rather than letting it rot. refactors should propose deletions alongside additions.

feedback loops are the recurring enemy#

the hardest production bugs are phi consuming her own outputs:

• extraction feedback loop — memory extraction re-ingesting phi's paraphrases as facts (answer: append-only observations, trust labels on memory sources)
• bot-hallucination loop — hallucinated claims getting stored then recalled as ground truth (answer: hardened extraction prompt, trust hierarchy)
• voice-drift loop — phi's own recent posts in context dragging her register (answer: 41623ce)
• tool loops — check_urls / tool-call runaway (answer: caps + dedup tracking)

when adding any path where phi's output can become phi's input, name the loop and decide how it terminates.

atproto DotDict bites repeatedly#

get_record(...).value is a DotDict, not a dict: isinstance(x, dict) is False and .get() misbehaves. use get_model_as_dict before structured access. this caused real bugs at least three separate times (1dad5d6, c39a64a, d3c0166). also: cross-repo getRecord/getBlob must resolve DID→PDS via the DID doc — bot_client/bsky.social only works for bsky-hosted accounts.

personality drifts toward voice, prompts toward structure#

the personality file has moved from behavior contract → voice description → disposition, repeatedly stripped of rules, glossaries, and sticky phrases phi parrots verbatim. meanwhile the system prompt moved the opposite way: named blocks ([GOALS], [RESIDUE], [WORKFLOW STATE], ...) each backed by real state, documented in system-prompt.md. don't put rules in the personality or vibes in the context blocks.

prefer deterministic code over an LLM step#

where a step has a closed set of outcomes (workflow-state classification, bio status marker, schedule selection), it has eventually been rewritten as plain code. reach for the LLM only where judgment is actually required.

a block that can render empty can also fail silently#

context blocks return "" when their input is absent — the documented empty-when-unset contract in system-prompt.md. the cost is that a broken read and absent data are indistinguishable from the outside.

on 2026-07-24 an audit found two blocks that had rendered zero times across 611 consecutive production requests. [SELF STATE] read createdAt off a typed atproto Record that serializes to created_at, so the lookup returned "" — written broken in the same commit that added the block, dead its entire 3.5 month life. [DISCOVERY POOL] was fetching a hub endpoint that had gone behind cloudflare access; response.json() raised on the HTML login page and the handler logged it at debug.

• when a block can legitimately render empty, its failure path logs at warning, never debug.
• anything a block depends on belongs in SERVICE_CHECKS (tools/_helpers.py) — hub was absent from it, which is why check_services could never have caught this.
• to audit: strpos for each block label over attributes->'gen_ai.system_instructions' in logfire, counted across a wide window. the hone-prompts skill carries the query. a position of 0 over hundreds of requests means dead, not quiet.

fixing one of them proved it redundant — [SELF STATE]'s surviving line duplicated [RECENT OPERATIONS], so it was deleted rather than kept. subtraction again.

prescription accumulates where nothing inspects it#

the eleven process_* entry points in agent.py carry ~17k characters of hand-written instruction in string literals. that prose is the least-inspected in the system: not drift-checked like docs/system-prompt.md, not surfaced by /api/abilities, and not rendered by the hone-prompts skill, which reads composed instructions and never sees a task prompt.

it is where over-prescription collects. the workflow-alert task ordered name every failure, state that these are terminal run events, do not suppress, do not diagnose — five clauses, each removing a judgment — and produced 15 of phi's last 25 top-level posts as syslog with a mention attached. the cycle task spent two thirds of its 3,514 characters on a decision table mapping workflow classifications to fixed responses.

• guidance about a tool belongs in its docstring; semantics of a block belong in that block's header, next to what they describe. an entry point's task should say what phi is doing right now, and stop.
• a signal should reach phi as perception, not as a command. [WORKFLOW STATE] was already a block she read and judged; the alert path carried the same facts as an order. same information, two shapes, and the ordered one wrote in machine register.
• when a task prompt grows past a few sentences, ask what part of it is describing a tool, a block, or phi's voice — those have homes.

phi writes about what she is woken up to look at#

the 2026-07-25 diagnosis of "phi sounds dry" ended nowhere near personalities/phi.md. every scheduled wake pointed at machine state: cycle opened on [WORKFLOW STATE], two daily passes on the chicken market, reflection on her own metrics, editorial on coral's record. nothing ever woke her up to read a person. so she wrote status reports even where she chose the subject, and [SELF-AWARENESS] accurately reported mode: mostly operational alerts and incident reports.

the tempting fix was to stop counting alert posts in that inventory. that edits the mirror. the diet is the cause — hence the people pass, which takes one of the four daily thought slots.

before changing a block that describes phi to herself, check whether it is lying. if it is accurate, the thing it describes is what needs to change.

defences against voice contagion compound into an agent that never reads#

three separate measures, each right on its own, each protecting against a real failure this repo actually hit:

• [RECENT OPERATIONS] strips post bodies to a char count, "so this block doesn't double as voice training" — the voice-drift loop (41623ce).
• [SELF-AWARENESS] is written deliberately flat, "do not imitate this block's register" — exemplar pressure beats abstract rules, so a mirror that spoke in phi's register would reinforce it.
• [DISCOVERY POOL] said "do not copy their phrasing" — attribution.

together they left the pool samples as nearly the only human writing in phi's context, flagged do-not-imitate, and the only unfiltered writing anywhere in her context is people talking to her in [NEW NOTIFICATIONS] — which requires someone to speak first.

"phi sounds dry" had this underneath it. she was never allowed to read anything. the fix was to separate don't lift someone's sentences (real, kept) from don't learn from writing (accidental, removed), and to name the pool as what it also is: posts the operator chose to like, which is a taste signal, not only a list of leads.

when adding a guard against imitation, say precisely what is forbidden. "don't copy this" and "don't learn from this" are different instructions, and a system that only ever issues the second produces an agent with no voice to protect.

the same block now also says humor does real work in ordinary communication and points at the samples as evidence of it working. that is the shape that gets past the four failed attempts at voice-in-the-prompt: a claim about how people talk, plus real exemplars, plus a task that is analysis — work out how someone landed one, they're often quiet about it. a handed-down voice gets parroted; a noticed one gets learned. never state the register you want.

"don't imitate this" and "don't know this" are different instructions#

three times in one day the same shape turned up: a guard against imitation, written in the cheap form, that also removed knowledge phi needed.

• [DISCOVERY POOL] said "do not copy their phrasing", which also made the only human writing in her context off-limits to learn from.
• [RECENT OPERATIONS] stripped post bodies so it could not double as voice training, which also meant she could not tell she had already posted something — and she posted the same essay twice in one day, five hours apart, in near-identical words.
• both were correct defences against real failures this repo had actually hit. neither was wrong to exist. both were wrong in scope.

when writing a guard against imitation, say precisely what is forbidden. copying someone's sentences, and knowing what is in front of you, are not the same act. a system that keeps issuing the second instruction produces an agent that has read nothing and remembers nothing.

and note where the fix went: into the block header, not the task prompt. a scheduled task saying "check [RECENT OPERATIONS] before you post" is the same hardcoding the entry-point work exists to remove — phi should always know what she recently said, on every path, rather than be told to go look on one.