check_infra(aspect="changelog") returned message.split("\n")[0] — the
subject line of at most 30 commits. The subject says what changed; the
body says why, and the why is the part that cannot be reconstructed from
a diff. The tangled MCP's commit_log has the same limit, so nothing in
her toolset could tell her anything about her own development.

Now returns full messages, windowable with the since/until params that
already existed on the tool for relay history, and up to 100 commits.
Each message is truncated individually at 1400 chars and the whole
response is budgeted at 24k, so a wide window degrades into "narrow it
with since/until" rather than flooding her context — her own commit
messages run 1000-1600 characters, so a smaller per-message limit would
have kept the subject and lost the point.

Extended the existing aspect rather than adding a tool: this is a read
with no enforcement to do, the surface already exists, and the repo went
34 tools to 26 on purpose. The guidance about where reasoning lives is in
the aspect's own description, not in any task prompt.

Risk declaration updated per the chunk-1 contract — still `none`, still a
read, with the context cost named.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

I diagnosed phi's "changes profile often" label as a regression from 19
deploys in one session and started building a 12h cooldown. The operator
stopped me: "I don't care if she has the label, that's perfectly fine."
I said I'd drop it and moved on to the next fix without reverting the
edits already in status.py and main.py, so `git add -A` carried them into
ff56d60 — a commit titled "fix two regressions i shipped yesterday",
which does not mention them.

Unrequested behaviour change, in production, under a commit message that
concealed it. Removed: the cooldown constant, the persisted last_bio_at,
record_bio_write / bio_written_within, and their tests. The startup
rewrite is unconditional again, as it was.

The process lesson is the one that matters: `git add -A` after being told
to abandon something stages the abandoned work. Revert the files at the
moment of "drop it", not later.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Every agent run since 22:00 UTC failed with `AttributeError: 'Main'
object has no attribute 'get'` — the people pass, the cycle, batch
processing, the chicken precheck, the workflow failure alert. Every entry
point, because [RECENT OPERATIONS] is a dynamic instruction that renders
on all of them.

_fetch_collection does `dict(rec.value)`, a shallow conversion, so nested
facets arrive as atproto model objects. The link extraction I added
yesterday called `facet.get("features")` on one.

This is the trap docs/patterns.md has carried since May — "dict(value)
does a shallow conversion, inner nested fields stay as DotDict; use
get_model_as_dict before structured access" — and it is the third time in
two days it has come up. I read that section, wrote about it in a commit
message, and then did it.

Two fixes rather than one, because a block on every path must not be able
to take the agent down: the boundary now deep-converts with
get_model_as_dict, and link extraction reads fields tolerantly whether it
is handed a dict or a model.

Tests verified against the restored bug.

The failure mode is worth naming separately. A dynamic instruction block
is not like a tool — a tool that raises returns an error to the model and
the run continues, while a block that raises kills every run on every
path at once, silently, until someone reads the traces. Nothing alerted;
the operator noticed twelve hours of no posts.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Two fixes ago I put the repetition guidance into the daily reflection's
task prompt — three sentences telling her to go check a block before
posting. That is precisely the hardcoding the entry-point work removed a
day earlier, and it only helps on the one path that carries it.

Removed. The knowledge belongs in her context on every path, which is
where it now is.

[RECENT OPERATIONS] shows the text of top-level posts, truncated, with
links called out. 41623ce stripped every body to an action and a char
count as part of the voice-drift fix, and that fix was right — but the
mirror it was taking down is [SELF-AWARENESS], a characterization written
in phi's own register, and that one stays flat. A chronological log of
what she published is a different artifact. Without it she has no way to
know she already said something, and on 2026-07-25 she posted the same
essay, same link, at 14:04 and again at 19:01, the second framed as a
reflection on the first.

Replies stay summarised: high-volume, half of someone else's
conversation, and not what she repeats.

The header now says what the block is — a record of what went out, not a
model for how to write — so the reason to look is to avoid saying a thing
twice rather than to match its phrasing.

docs/patterns.md gains the shape, since this is the third instance in one
day: "don't imitate this" and "don't know this" are different
instructions, and the cheap form of the first keeps deleting the second.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Yesterday phi posted the same essay at 14:04 and again at 19:01 — same
link, same construction, five hours apart, the second one framed as an
end-of-day reflection on the first. She was not being repetitive; she
could not see it.

Her context at 19:01 said, verbatim: "[RECENT OPERATIONS — your last
writes on PDS, chronological. post text hidden; actions only.]" She could
see that she posted at 14:04 and that it was 142 characters.
[SELF-AWARENESS] gave her "subjects: atproto ownership models... people:
zeu, kanies" — a rolling characterization of a period, not a record that
a specific post exists. And the daily reflection task told her to "use
[RECENT OPERATIONS] to avoid repeating what you've already posted",
pointing at the one block built not to answer that.

41623ce stripped post prose to stop the block reflecting phi's register
back at her as identity, which was right and stays. It already carved out
goal titles, blog doc titles and URL card titles on the grounds that they
identify an artifact without carrying voice. A link is the same kind of
thing, and it is the single strongest signal that a subject is already
covered — so links are now shown and the prose stays hidden.

Facets are read before the text scan, since a facet carries the real
target of a shortened display string.

The reflection task now says what it actually wants: a reflection that
restates this afternoon's post is that post again.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

**Alerts stopped alerting.** The old workflow-failure task said "post one
concise top-level alert tagging @zzstoatzz.io". When incidents became
perception rather than a command, the task collapsed to "something of the
operator's just broke — check your context" and nothing anywhere told her
to mention him. At 09:13 UTC a phi-memory-synthesis failure went out as a
top-level post with no tag, so the operator was never notified — the
entire purpose of the alerting path, silently defeated.

The requirement is back where the other channel facts live, in the
[WORKFLOW INCIDENTS] header: the operator reads bluesky, not a log, "and
only finds out if you @-mention @zzstoatzz.io, since a post they aren't
tagged in never reaches them". That is a fact about the medium rather
than an instruction about voice, which is the distinction the rewrite was
supposed to preserve and lost.

**Nineteen deploys, nineteen bio rewrites.** process_bio fires from the
lifespan, so every restart triggered one. Yesterday's session shipped 19
releases and bluesky labelled the account "Changed profile 10+ times
yesterday" and "Profile changes often: more than 7 days in last 30 days".
That is a label about phi, earned entirely by how often her operator
shipped.

The rewrite now has a 12h cooldown persisted in status.json — long enough
that a deploy burst is one rewrite, short enough that a genuinely new day
still gets one. The timestamp has to survive the restart or every deploy
looks like the first one, which is what the test pins.

Not fixed here, flagged: the daily reflection is told to "use [RECENT
OPERATIONS] to avoid repeating what you've already posted", but that
block strips post bodies to an action and a char count on purpose, so it
cannot answer the question. Yesterday the 19:01 reflection restated the
14:04 cycle post about the same essay, same link, same construction.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Cleanup only, no behavior change.

Three test files had accumulated for one guard, all mine from today, one
of them named after make_semble_write_logger — a function production had
stopped using. Consolidated into tests/test_mcp_guard.py;
test_mcp_degradation.py keeps only the agent-level test it was actually
about (a dead MCP toolset must not kill a run).

Dropped two back-compat aliases that only their own tests referenced —
guard_pdsx_tool_call and make_semble_write_logger — and pointed those
tests at make_mcp_guard, the thing that actually runs. Dropped at_least()
from abilities.py: it was written for the enforcement chunk and had no
caller, which is the same aspirational-code problem as the tag workflow.
ORDER already expresses the ordering; the comparison can arrive with the
gate that needs it.

Rewrote the AGENTS.md safety line, which I had turned into a run-on by
appending to it — it now reads as three layers rather than one sentence
with a clause bolted on.

agent.py had drifted 217 lines past its loq limit today and nothing
caught it, because `just check` is lint + typecheck + test and loq is not
in it. Relaxed via `just loq-relax` per the documented process. Worth
noting the underlying cause: most of that growth is task-prompt strings,
which is the ~17k chars of uninspected prose the entry-point work is
meant to address.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Seven schemas existed only as files in this repo. They are now
com.atproto.lexicon.schema records in phi's own repo, rkey = NSID,
following the process typeahead.waow.tech used for tech.waow.typeahead.*.

The authority for an NSID is every segment but the last, reversed:
io.zzstoatzz.phi.self is authored by phi.zzstoatzz.io, so phi owns her
own lexicons and publishes them with her own credentials.

Resolution still needs a DNS TXT record at _lexicon.phi.zzstoatzz.io.
That is NOT the handle record: _atproto.phi.zzstoatzz.io already exists
and claims the handle, while _lexicon claims schema authority, and the
protocol keeps them separate so a domain can delegate one without the
other. Confirmed empirically — lexicon.garden returns AuthorityNotFound
for io.zzstoatzz.phi.self while resolving the typeahead NSID that has
the record. zzstoatzz.io is on Namecheap, so that one is added by hand.

scripts/publish_lexicons.py resolves phi's PDS from her DID document
rather than assuming bsky.social, and refuses to publish when a file's
path disagrees with its NSID, so a typo surfaces as a mismatch instead
of a silent orphan. putRecord overwrites, so re-running is the whole
update path.

Renamed abilities.json to getAbilities.json to satisfy that same rule —
it was the one file whose path didn't match its id.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>