fix(desktop): stop dropping 19 of the harness's log targets

[yjc801]

Summary

child_rust_log_filter handed the spawned harness buzz_acp=info. That looks
right and matches almost nothing.

EnvFilter matches on an event's target, which defaults to the module path —
but buzz-acp sets an explicit target: on its diagnostic lines, and an explicit
target replaces the module path rather than extending it. None of those
targets begin with buzz_acp, so none of them ever matched.

That silenced 19 targets under five roots:

Root	Targets
pool::	prompt, session, model, permission, metrics
acp::	wire, update, usage, permission, tool, cancel, thought, stream, session, plan, init
canvas::	fetch
engram::	core
observer	—

Two of the casualties are lines whose absence cost real debugging time: the only
record that a session was created (created session … for channel …), which made
session rotation unobservable, and pool::model's model-override miss, which is
what would say whether a [1m] model ref actually reaches the API. Both questions
went unanswered for days against a log that was structurally incapable of
answering them.

The fix names each root explicitly. With no RUST_LOG set, the harness gets:

buzz_acp=info,pool=info,acp=info,acp::stream=off,engram=info,canvas=info,observer=info

buzz_acp= stays in the list — lines that don't override target: still fall
back to the module path, which does start with buzz_acp.

Why info and not debug — measured, not assumed. Across these targets the
call sites are roughly 11 debug, 7 info, 6 warn, 2 error. At info the
debug lines stay off, including acp::wire's 11 frame dumps and acp::thought,
so this surfaces the lines worth reading without inflating every agent's log.

Why acp::stream=off — it is the one info-level site that fires per chunk,
logging the text of every agent_message_chunk
(crates/buzz-acp/src/acp.rs), so acp=info
alone would copy every agent response into the runtime log verbatim. Nothing
would bound that: maybe_rotate_log runs only inside open_log_file
(managed_agents/storage.rs), which is called once at spawn, so the 10 MB
ceiling is never re-checked for the life of the process. It stays available on
request — RUST_LOG=acp::stream=trace or acp=debug both turn it on.

Defaults are merged per target, not appended. Appending is not neutral:
EnvFilter keeps one directive per (target, span, fields), and DirectiveSet::add
does Ok(i) => self.directives[i] = directive on a binary-search hit while
Ord for Directive compares target/span/fields and never the level. So a trailing
pool=info wins over an operator's RUST_LOG=pool=off, and clips acp=debug
back to info. Instead, a default is dropped when the operator names its target
or an ancestor of it:

RUST_LOG	result
unset	all seven defaults
pool=off	pool stays off; other roots still covered
acp=debug	drops both acp=info and acp::stream=off — debug means debug
acp::stream=trace	drops only acp::stream=off; acp=info still covers the siblings
warn	a bare level names no target, so it suppresses nothing
anything naming buzz_acp	passed through untouched (see Scope)

Directive parsing strips the [span{field=value}] section before splitting on
=, since that section can itself contain one.

Scope

Local agents only. A provider-backed agent's harness is launched by its
backend from a separate env and never through this path, and
get_managed_agent_log refuses remote agents outright. This does nothing for
diagnosing a sprite.

The operator escape hatch is preserved: a RUST_LOG that names buzz_acp
explicitly is passed through untouched, so someone narrowing the filter to one
target doesn't get it widened back out.

Related issue

Upstream block/buzz#3309 — same bug, other side of the seam. Currently open, not merged.

block#3309 diagnoses the identical mechanism (explicit target: replaces the module
path, so buzz_acp=* matches none of it) and fixes it in buzz-acp by
prefixing every custom target with buzz_acp::, so the crate filter reaches
them. That is the better fix and it strictly dominates this one on coverage:

• it fixes every consumer, including provider-backed sprites, which this PR
  explicitly does not;
• it makes RUST_LOG=buzz_acp=debug behave as TESTING.md already documents,
  instead of leaving the docs wrong;
• it needs no per-root list to be kept in sync.

This PR is a fork-local workaround that works today without waiting on an
upstream merge, and is worth having on those terms — but a reviewer should know
it is the narrower of the two fixes.

They are compatible. If block#3309 lands and this fork merges upstream, the
re-rooted targets match the buzz_acp=info clause that this filter still
carries; the extra roots become inert rather than conflicting. The redundant
roots should be dropped at that point, and HARNESS_TARGET_ROOTS in
log_filter.rs is the thing that will flag it — it asserts against the roots
buzz-acp actually uses, so a re-rooting upstream fails that test loudly instead
of leaving dead config behind. acp::stream=off would need re-expressing as
buzz_acp::acp::stream=off at the same time.

Layout

The helper and its tests moved out of runtime.rs / runtime/tests.rs into
managed_agents/runtime/log_filter.rs (281 lines). Keeping them in place
pushed runtime.rs to 1022 against a 1000-line limit and runtime/tests.rs
to 1333 against its merge-base ratchet of 1275, so just ci could not pass.
Both files are back at their base contents.

Testing

cargo test --manifest-path desktop/src-tauri/Cargo.toml --lib log_filter

test result: ok. 11 passed; 0 failed. Full gate: just ci green.

The tests assert resolved behaviour, not substrings. The defect this module
exists to fix was a filter that read correctly and matched nothing, so
filter.contains("pool=info") is not evidence of anything. Each case parses the
produced filter and asks whether a target is enabled at a level, through
tracing_subscriber::filter::Targets — a newtype over the same
DirectiveSet<StaticDirective> that backs EnvFilter's static directives,
parsed and resolved most-specific-first the same way. Added as a
dev-dependency; it was already in desktop/src-tauri/Cargo.lock, so the
lockfile change is one line.

Test	Asserts
default_filter_enables_every_harness_target_root	every root in HARNESS_TARGET_ROOTS resolves to enabled at info; a sixth root added upstream fails here rather than silently going dark
harness_targets_are_enabled_at_info_not_debug	acp::wire/pool::prompt stay off at debug — guards against "fixing" quiet logs by turning on the frame dumps
response_chunks_are_off_by_default	acp::stream off at every level, siblings still on
an_empty_or_blank_rust_log_is_treated_as_unset	"" and whitespace
an_unrelated_rust_log_is_extended_not_replaced	hyper=warn survives, defaults appended
an_explicit_buzz_acp_filter_is_passed_through_untouched	operator override
an_operator_silenced_root_stays_silenced	pool=off is not re-enabled — the regression test for the append bug
an_operator_raised_root_is_not_downgraded	acp=debug reaches debug, including acp::stream
asking_for_the_stream_gets_the_stream_and_keeps_the_rest	a leaf directive suppresses only the leaf default
a_bare_level_names_no_target_and_suppresses_nothing	RUST_LOG=warn
a_span_field_directive_does_not_confuse_target_parsing	pool[work{id=7}]=off parses as target pool, not pool[work{id

Two traps worth naming, because both produce a green result that means nothing:

1. The desktop crate is excluded from the root workspace, so a bare cargo test
   at the repo root does not cover this. The --manifest-path is required.
2. cargo test exits 0 when a name filter matches nothing. Filtering on a
   function name that no test is called leaves you with a passing build that
   verified nothing — check the reported test count, not the exit code.

No UI change, so no screenshots.

🤖 Generated with Claude Code