docs(context): report snapshot retention benchmarks

Record timing ranges and memory savings, including zero retained replay bytes and the 79 to 145 microsecond text-capture regression. Replay counting still serializes JSON, and inspection still renders the full history per page.

test(context): cover payload release and benchmark snapshots

Check payload release, replay delivery and session cleanup through real snapshots and daemon sessions. Offline runs measure capture, inspection and replacement with one warm-up and five samples.

perf(context): drop omitted payloads from snapshot retention

History retention drops from O(n + v + p) to O(n + v), with n entries/images, v visible bytes and p discarded payload bytes; capture complexity is unchanged. The 64 KiB replay fixture releases 62.5 MiB, while 1k-text capture shifts from 79 to 145 microseconds.

test(tui): cover transcript buffer ownership

Cover retained snapshots, detached history, discard, chunk boundaries, and message deduplication. The full repository gate and TUI race suite pass.

perf(tui): assemble live chunks in amortized O(n)

Replace O(n²) copying per live chunk with amortized O(n) assembly, immutable snapshots, and detached history. For 256 KiB in 32-byte deltas, median assembly fell from 46.35 to 0.56 ms and allocated bytes from 288.83 to 1.43 MB (-benchmem -benchtime=300ms -count=3).

perf(daemon): bound preview tail memory to O(1)

Preview extraction stays O(n) time; extra space drops from O(n) to O(1), capped at 512 runes. Growth-sequence medians before -> after (-benchmem -benchtime=300ms -count=3): 16 KiB: 1.493 -> 0.307 ms, 273 -> 238 allocs, 838432 -> 180889 bytes 128 KiB: 157.109 -> 12.879 ms, 2178 -> 1830 allocs, 44233404 -> 9233466 bytes

test(daemon): cover streamed preview tails and benchmark argument growth

Pin rune offsets and sanitized tails at partial JSON boundaries that E2E cannot reliably pause at. Benchmark complete 16/128 KiB growth sequences in 1 KiB steps.

test(python): cover trace lifetime and daemon persistence

Cover late tasks, nested delivery gates, failed writes, and traces surviving capture eviction.

fix(python): seal and release finalized traces

Retained trace storage drops from O(B + A) to O(1), with B snapshot bytes and A activity bytes. Python 3.11 tracemalloc, 16 traces with sixteen 120 KiB files and 64 activities each: 30.3 MiB to 6.7 KiB after release.

test(tui): check survivor ownership after trimming

Cover row, byte, and complete eviction, including descriptor clearing and independent survivor storage.

perf(tui): reduce survivor allocations to O(1)

Copying survivors drops from O(n + B) to O(n), where n is surviving rows and B is their text bytes. Three 300ms benchmark runs cut row-cap allocations/op from 1,002 to 2 and bytes/op from 136,960 to 40,960; byte-cap allocations/op fell from 510 to 2 and bytes/op from about 314,369 to 21,760.

test(tui): benchmark trimming a full display

Measure append-and-trim allocations at the row and byte caps, with rendering outside the timer.

perf(python): release delivered cell images

Delivered-image archive space drops from O(R * B) to zero for R retained cells and B image bytes per cell; delivery remains O(B). Across three runs of 24 fresh 512x512 PNG cells, post-GC Python allocations fell from 12.03 MiB to 18.7 KiB, with zero archived image bytes and identical payloads.

fix(test): read complete owner frames without buffering

Keep frames visible to select and read each payload fully, including large image results.

fix(python): release obsolete job output watchers

Use weak callbacks and clear obsolete watchers while preserving early-read notifications and retention limits. Benchmark: retained memory falls 73%, from 5.02 to 1.35 MiB after 256 jobs.

make bursts clickable to see what agent did

fix(python): validate protocol messages and remote settings

fix(python): align tooling and Nix runtimes with Python 3.11

fix: preserve Python 3.11 compatibility in remote kernels

chore: remove audit cruft and exclude generated reports

chore: remove lint audit CSVs

refactor(gleam): clean up code and enforce focused lint checks

fix(python): enforce ruff and ty checks

fix(fmt): format test.sh

refactor(cli): remove redundant daemon startup wrapper

refactor(antigravity): move schema policy into Gleam

fix(daemon): return 503 when request actors disappear

Restart regression passed three runs.

perf(retrieval): page O(n) rows and retain O(k) results

Sparse search: 2 pages, 2 decodes. 2200 matches: 20 retained; count remains exact. Node page: 7 of 45 summaries read. Text page: 128 of 300 rows decoded.

perf(lcm): cut below-trigger payload decodes from O(n) to O(1)

802-row fixture: 0 decodes below trigger or for stored prior. Forced compaction: 2 eligible rows decoded.

perf(history): cut image fitting from O(n*f) to O(n log(f+1))

Add indexed row classes and resumable backfill. Add snapshot-bounded source folds.

refactor(harness): remove unused projection and API wrappers

refactor(models): move provider cache schemas into Gleam

Keep Erlang adapters focused on transport, storage, and refresh coordination. Add parser, refresh, and picker regression coverage. Validated with gleam check, 170 Gleam tests, model-catalog E2E, and test.sh.

refactor(harness): move instruction policy into Gleam

Move selection, precedence, and rendering into Gleam. Keep discovery and validated reads in Erlang, with E2E coverage for ordering and file selection.

perf(cache-ttl): make warm reads O(1) and globs O(p*v)

Cache typed tables and failed revisions; reject duplicate layer IDs. 8,000-entry warm table: 34.4 ms -> 0.11 ms. 16-wildcard cold lookup: >2 s -> 13.4 ms.

perf(claude): cut required checks from O(r^2) to O(r log r)

Index branch requirements and skip overwritten fields. 800 names per branch: 46.8 ms -> 1.02 ms median normalization.

perf(models): replace O(p) provider scans with O(log p) lookups

Return typed results and trust catalog efforts; infer only for absent models. Use direct provider lookup without an endpoint and remove unused wrappers.

perf(capabilities): cut preference reads from O(n) to O(1)

Reuse one snapshot per selection and share the 1 MiB read/write cap. 128 checks on 198,914 bytes: 256.2 ms -> 2.8 ms.

perf(cache-ttl): cut merging from O(n^2) to O(n log n)

Move merge policy into Gleam and replace repeated list scans with an ID index. Preserve ordering and override behavior; file I/O and caching stay in Erlang. Three 8,000-entry layers: 7,329 ms in the original Erlang merge, 14 ms now. Median of three runs, merge only. Parity checks and test.sh pass.

refactor(models): move catalog selection into Gleam

Provider matching and conservative fallback limits are application policy. Move selection into catalog.gleam while keeping fetching, caching, and index persistence in Erlang. Preserve the JSON contract, persisted index format, and existing selection behavior. Replace duplicate unit tests with daemon coverage for shared capabilities, endpoint matching, and restart persistence. Verified 11,260 old/new selection and listing comparisons and the full test.sh gate.

refactor(claude): move schema normalization into Gleam

Claude schema normalization is pure provider policy, so it belongs beside the request encoder. Move the merge rules into schema.gleam and remove the Erlang module. Keep arbitrary schema metadata and nested schemas intact. Add a request regression test for property precedence, required fields, and descriptions. Verified 1,009 cases against the old normalizer, 159 Gleam tests, and the full test.sh gate.

refactor(capabilities): move preference rules into Gleam

Use typed decoding for overrides and validation while Erlang retains file persistence.

refactor(cli): share login state cleanup

Delete the temporary provider before restoring the saved config. Report cleanup errors and attempt restoration even when deletion fails.

chore(cli): drop unused snapshot cursor

The stream snapshot helper only reads events. Remove the unused cursor field from its response struct.

refactor(cli): share open mode classification

Use one helper to classify file modes for Python open calls and Path.open calls.

fix(cli): preserve cache fade in usage events

Decode cache fade steps so the footer can age cached token counts. Ignore malformed fade data while keeping the rest of the usage event.

extensions: add cdp browser tool

Signed-off-by: dawn <90008@klbr.net>

remote: run commands under login shell, discover agent socket, and raise timeouts

Signed-off-by: dawn <90008@klbr.net>

python: block daemon tokens and secret-shaped variables in clean_environment

Signed-off-by: dawn <90008@klbr.net>

fix(config): move provider validation beside its Gleam types

.

fix(tui): report terminal restoration failures during color detection

Stop TUI startup if restoring terminal settings fails, while keeping default colors for ordinary query failures.