Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
53 commits
Select commit Hold shift + click to select a range
4e258f3
design/84: a local OpenAI-compatible backend behind the same loop (pr…
iandobbie Oct 1, 2026
2af5a08
design/84: record the 128k/q8 KV spike; settle Q1, reframe Q2
iandobbie Oct 1, 2026
1c6ab64
design/84: settle the gate machine; open 84a's ledger row
iandobbie Oct 1, 2026
e1fadc4
CLAUDE.md: the yonce push key is per-machine; ask which remote elsewhere
iandobbie Oct 1, 2026
afa31ef
design/84: record 84a's review
iandobbie Oct 2, 2026
6bece9b
design/70: R160, estimate_tokens counts thumbnail base64 as text
iandobbie Oct 2, 2026
55845b2
design/84: record 84b's review; gate 84 pushed
iandobbie Oct 2, 2026
57d1647
design/84: score gate 84 round 1; open 84c (compaction only when it h…
iandobbie Oct 5, 2026
712086d
design/84: record 84c's review; gate 84 round 2 pushed
iandobbie Oct 6, 2026
0380ac1
design/84 84a: recorded llama.cpp responses for the fake server
iandobbie Oct 1, 2026
f625e28
Add local OpenAI-compatible backend path
iandobbie Oct 2, 2026
83fa429
Fix local backend streaming and error handling
iandobbie Oct 2, 2026
c9af47c
design/84 84a: round-2 review fixes (UNREVIEWED; runner turn killed a…
iandobbie Oct 2, 2026
4509d1e
design/84 84b: recorded llama.cpp /tokenize response
iandobbie Oct 2, 2026
b76a557
Derive local context budget from server
iandobbie Oct 2, 2026
5a030d6
design/84 84b: recorded llama.cpp /apply-template response
iandobbie Oct 2, 2026
a3ee341
Count rendered local prompt budget
iandobbie Oct 2, 2026
60c0bb9
Allow keyless local backend in browser
iandobbie Oct 2, 2026
d76b1c3
Recompute local budget after server authentication
iandobbie Oct 2, 2026
b743d4d
design/84 84b: register a stored local key with the history audit
iandobbie Oct 2, 2026
1c1fe18
design/84: gate 84 runbook, scorer and selftest
iandobbie Oct 2, 2026
180a904
design/84 84b: declare keyRequired with the page state; absent means …
iandobbie Oct 2, 2026
bc72096
design/84 84b: update the no-n_ctx budget test for the provisional floor
iandobbie Oct 2, 2026
bf88ba5
design/84: pin gate 84 to 557200f
iandobbie Oct 2, 2026
7fa4ea4
gitattributes: keep the recorded llama.cpp fixtures byte-exact on Win…
iandobbie Oct 2, 2026
7ad6cb6
design/84: pin gate 84 to 988abea and refresh CRLF fixtures in an old…
iandobbie Oct 2, 2026
6b2c81f
Gate compaction on useful context reduction
iandobbie Oct 5, 2026
8d20cab
Cover useful compaction and notice stretches
iandobbie Oct 5, 2026
fbf6c74
design/84 84c: gate scorer grades fit and D9 separately; selftest siz…
iandobbie Oct 6, 2026
649034e
design/84: gate 84 round 2 runbook for 84c
iandobbie Oct 6, 2026
7720e24
Keep instrument tests useful under D9
iandobbie Oct 6, 2026
fcfd407
design/84: pin gate 84 round 2 to 4c99b81
iandobbie Oct 6, 2026
9f3beb3
design/91: close 91a-91c (merged), score gate 91 round 2, coordinatio…
iandobbie Oct 6, 2026
ab352df
design/84: open 84d (D12-D16)
iandobbie Oct 6, 2026
07ba08d
Declare httpx as a runtime dependency
iandobbie Oct 6, 2026
20e8696
Calibrate local context compaction and notices
iandobbie Oct 6, 2026
dd5cd92
Fix local compaction ordering and effort setting
iandobbie Oct 6, 2026
742870e
design/84 84d: an empty MICROCLAW_LLM_REASONING_EFFORT means unset
iandobbie Oct 6, 2026
088f50e
design/84: gate 84 round 3 runbook for 84d; record 84d's review
iandobbie Oct 6, 2026
a643f67
design/84 84d: gate round 3 fixes - whole-token rebuild count, copyab…
iandobbie Oct 6, 2026
0c230f1
design/91: score gate 91 round 3; close 91d; R161, R162
iandobbie Oct 6, 2026
a4e6ed3
design/prompts: 91d coordination notes
iandobbie Oct 6, 2026
3934d31
design/84: reasoning-effort results on a real microscope; R162, R163
iandobbie Oct 14, 2026
688330f
design/84: open 84e (D17, reasoning level within a session)
iandobbie Oct 22, 2026
26ce9ab
Add session reasoning effort control
iandobbie Oct 22, 2026
92e6121
design/84 84e: budget helper after the imports; show a startup level …
iandobbie Oct 22, 2026
5cbca8d
design/84 84e: the reasoning-change notice states an upper bound
iandobbie Oct 22, 2026
c76fe16
design/84: block 84e workstation check and review record
iandobbie Oct 22, 2026
fb3834c
design/84: score and close the 84e check
iandobbie Oct 22, 2026
343b125
design/prompts: 91e coordination notes
iandobbie Oct 22, 2026
d91eaec
Adjusted Ian's design/84 to design/91 to fit with extra design goals …
iandobbie Oct 23, 2026
2118d15
design/70: R161, R17 and R165 from the PR #62 demo run; index R160-R164
zacsimile Oct 27, 2026
a385f66
docs: using Microclaw with a local model
iandobbie Nov 8, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions .gitattributes
Original file line number Diff line number Diff line change
Expand Up @@ -20,3 +20,10 @@ tests/fixtures/skill_packages/** -text
# Gate releases and signed documents must survive checkout byte-for-byte;
# `-text`, as above, so the catalogs stay reviewable in a diff.
design/83f3-gate/** -text

# Recorded llama.cpp HTTP responses (design/84): replayed byte-for-byte as
# server-sent events, whose framing is a bare LF. A CRLF translation on Windows
# breaks every replay — measured on the workstation, 2026-10-02: the gate
# selftest went to 3 failed / 9 errors. `-text`, so they stay diffable.
tests/fixtures/llamacpp/** -text

5 changes: 4 additions & 1 deletion CLAUDE.md
Original file line number Diff line number Diff line change
Expand Up @@ -657,7 +657,10 @@ because a block looks small.

Standing rules that support the above: one worktree per concurrent agent, never
`git add -A`, never `pip install -e .` while another tree is live. Push uses a
non-default SSH key (`GIT_SSH_COMMAND="ssh -i ~/.ssh/yonce"`). Rig-facing
non-default SSH key (`GIT_SSH_COMMAND="ssh -i ~/.ssh/yonce"`) to `origin` — on
the machine that has that key. Elsewhere, check `git remote -v` and ask which
remote a block's branches go to (design/84's go to the operator's fork, `ian`,
with the default key). Rig-facing
commands must be PowerShell/cmd-safe. Rig facts belong in gate docs, design
notes, and rig profiles — never in `microclaw/`.

Expand Down
1 change: 1 addition & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,7 @@ User (natural language) → Agent → Tools → Safety Guard → Microscope Cont
| [Using Microclaw](docs/using-microclaw.md) | What a session is, what leaves the machine, skills, hooks, compiling a session to a standalone script |
| [Safety configuration](docs/safety-config.md) | The `safety_config.yaml` schema, every gate, and what a section's absence means |
| [The browser GUI](docs/browser-gui.md) | The chat interface, the desktop shortcut, API keys, remote access |
| [Using a local model](docs/local-llm-model.md) | Serving a model with llama.cpp, the settings Microclaw needs, what to expect, troubleshooting, models tested |
| [CLI reference](docs/cli-reference.md) | Flags and subcommands, the history viewer, where files live |
| [Publishing a skill package](docs/publishing-skill-packages.md) | Ship your own skill or analysis worker from your own repository: manifest, locks, worker protocol, signing |
| [Developing Microclaw](docs/development.md) | Source install, the test suite, building a safety profile by hand |
Expand Down
2 changes: 1 addition & 1 deletion THIRD_PARTY_NOTICES.md
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,7 @@ installed distributions and review notices for the versions they distribute.
| keyring | optional `serve` | MIT |
| pytest | test only | MIT |
| pytest-mock | test only | MIT |
| HTTPX | test only | BSD 3-Clause |
| HTTPX | runtime | BSD 3-Clause |
| setuptools | build only | MIT |

Transitive dependencies are not enumerated because they vary by platform and
Expand Down
90 changes: 88 additions & 2 deletions design/70-carried-forward-register.md
Original file line number Diff line number Diff line change
Expand Up @@ -121,6 +121,8 @@ Rows added since the triage:
| `R157` | `design/88` block 88b's demo gate and spike rerun, scored from artifacts 2026-10-18 |
| `R158` | `design/88` block 88b's Windows CI run, 2026-10-18 (not 88b's code) |
| `R159` | `design/89` block 89a's demo run, 2026-10-19 — the M2 limbs it could not run |
| `R160`–`R164` | `design/91`, the local OpenAI-compatible backend, 2026-10-02 to 2026-10-14 |
| `R165` | PR #62's demo-machine run (Qwen3.5-4B, local llama.cpp), scored from artifacts 2026-10-27 |


## The work queue
Expand Down Expand Up @@ -199,7 +201,7 @@ Sorted by ease, then by importance. `→` names an existing block; do the block,
| `R79` | [Saved knowledge does not separate measurement from inference](#r79) | HIGH | LARGE | |
| `R02` | [Agent offers to split an acquisition the backend already splits](#r02) | MEDIUM | SMALL | |
| `R16` | [get_device_property_info errors on a guessed property name without naming the real ones](#r16) | MEDIUM | SMALL | |
| `R17` | [An aborted turn's error never reaches the transcript](#r17) | MEDIUM | SMALL | |
| `R17` | [An aborted turn's error never reaches the transcript](#r17) | HIGH | SMALL | |
| `R21` | [Nothing reports the bounds microclaw is enforcing](#r21) | MEDIUM | SMALL | |
| `R24` | [Four of five connect sites' actionable refusal never reaches the user](#r24) | MEDIUM | SMALL | |
| `R25` | [An exported script writes its dataset beside the script, silently](#r25) | MEDIUM | SMALL | |
Expand Down Expand Up @@ -235,6 +237,11 @@ Sorted by ease, then by importance. `→` names an existing block; do the block,
| `R97` | [The runtime deadline includes Acquisition() construction, and nothing records how much](#r97) | LOW | SMALL | |
| ~~`R155`~~ | [Ctrl-C during a pycro-manager teardown wait can make a live waiter look finished](#r155) | MEDIUM | SMALL | **90a** |
| ~~`R158`~~ | [84a's window test reads the final display once, without a wait](#r158) | LOW | SMALL | **90a** |
| `R160` | [`estimate_tokens` counts a thumbnail's base64 bytes as text](#r160) | MEDIUM | SMALL | |
| `R161` | [One `read_hook_log` result can take a third of a local model's context](#r161) | HIGH | SMALL | |
| `R162` | [`test_built_wheel_contains_the_source_tree_skill_catalog` fails intermittently](#r162) | LOW | SMALL | |
| `R164` | [No route to combine saved channels into a colour composite](#r164) | MEDIUM | MEDIUM | |
| `R165` | [A local reply with no visible text ends the turn silently and is stored empty](#r165) | LOW | SMALL | |

**Demo machine**

Expand Down Expand Up @@ -283,6 +290,7 @@ Sorted by ease, then by importance. `→` names an existing block; do the block,
|---|---|---|---|---|
| `R06` | [probe_hint emitted for a lock with no usable status property](#r06) | MEDIUM | SMALL | |
| `R45` | [Block 13 G3 - transmitted-light SNR refusal unmeasured on any rig](#r45) | LOW | SMALL | → Block 13 G3 |
| `R163` | [At `reasoning_effort=low` the local model reported a dataset it never acquired](#r163) | HIGH | SMALL | |

**Blocked on a Nikon Ti — no reachable machine**

Expand Down Expand Up @@ -826,7 +834,8 @@ error path and this merge is already carrying unverified change.
**Failed agent turns appear in the browser stream but leave no durable error record in the transcript.**

- **Status** — OPEN - `microclaw/webserve.py:989-994` still emits exceptions only to SSE, while `microclaw/conversation.py:77-112` still reloads every JSONL object without role validation.
- **Importance** — MEDIUM - This loses useful failure evidence, but the browser displays the error and the operator can record it manually.
- **Importance** — HIGH (raised from MEDIUM, 2026-10-27). The MEDIUM rested on "the browser displays the error and the operator can record it manually", and that premise is false: `serve.html` shows a turn-ending `error` event only as a `toast`, which clears itself.
- **Seen again** — PR #62 demo run, 2026-10-27 (Qwen3.5-9B, local llama.cpp): `serve` logged `seq 5 error: Local server http://127.0.0.1:8081 context limit is 65536 tokens; request uses 70881 tokens.` The operator saw "nothing at all, it just ran", and the history holds the tool result and no reply. The error was lost from the screen as well as from the transcript. A fix should make a turn-ending error persist on the page (as design/91 91d did for its compaction notices) as well as record it.
- **Where** — LOCAL - A mocked agent exception and audit/reload assertions can settle the behavior without hardware.
- **Block** — NONE - No later design block claims or closes this transcript gap.
- **Effort** — SMALL
Expand Down Expand Up @@ -3403,3 +3412,80 @@ the model" as the actual blast radius.
- **Importance** — MEDIUM. Recommended before 89b, not required: 89b's own M2 gate compares cadence with feedback on and off, so this would only find a problem earlier and more cheaply (operator decision, 2026-10-19).
- **Effort** — SMALL
- **Provenance** — `design/89` 89a, 2026-10-19.


### R160 — `estimate_tokens` counts a thumbnail's base64 bytes as text

**One 512 px PNG thumbnail of a noisy field estimates at 121,748 tokens — above the Anthropic path's 120k high-water on its own.**

- **What happened** — measured 2026-10-02 (design/91, coordinator): `make_thumbnail` of a 2048² uint16 noise frame is 287,144 base64 bytes, and `estimate_tokens` over one `tool_result` holding it returns 121,748; the same call on a smooth ramp is 6,040 bytes and 2,636. `estimate_tokens` divides the JSON of the whole history by 2.36, image data included. Real camera frames are noise-dominated, so their thumbnails are the large case.
- **Why it matters** — a single `snap_and_analyze` that keeps its thumbnail can push `ConversationStore` past high-water and fold every older turn into a checkpoint on the next round, rewriting the cached prefix (design/82's cost lever) for an image the API bills at a few hundred tokens. Not observed in a recorded session yet: no artifact has been checked for a thumbnail-triggered compaction.
- **What a fix would look like** — estimate an image block at its billed size (a per-image constant from the API's image-token rule) rather than its byte length; check an archived design/82 history for a compaction that followed a thumbnail. design/91 91b fixes the local text-only path only, where the image is never sent.
- **Provenance** — design/91 91b planning, 2026-10-02.

### R161 — One `read_hook_log` result can take a third of a local model's context

**On the local llama.cpp server, `read_hook_log(limit=200)` returned 140,232 characters (~59k tokens): 104 s to process, and once more than the remaining context could hold.**

- **What happened** — design/91 gate 91 round 3, 2026-10-06, `n_ctx` 131,072: record 16 processed 59,262 new tokens in 104.0 s for that one tool result; a later turn asking for the same read made the next request too large and the server refused it (context exceeded, the turn rolled back). design/82 F2/D3 bounded the read by entry count (`limit`), which the model may raise to 200.
- **Why it matters** — on a local model the cost of a large result is prefill time on every cold path and a hard context limit, not just money; a bound in entries is not a bound in tokens.
- **Seen again, at the default `limit=50`** — PR #62 demo run, 2026-10-27, Qwen3.5-9B Q4_K_M at `n_ctx` 65,536 (`fixed=35579, high=17669, low=13251`). Turn 3 read 50 of 200 entries (~16k tokens) and declined compaction (`estimate=24085 threshold=17669`); its final request fit with 50 tokens spare (57,294 + 8,192). Turn 4 read the same 50 entries again, and the next request was 70,881 tokens; llama.cpp refused it, and the operator saw nothing (`R17`). At this context one 50-entry read is as large as the whole history budget.
- **Why compaction did not help** — `ConversationStore.model_messages` never folds the two most recent complete turns (`MIN_RECENT_COMPLETE_TURNS = 2`, on `main` since `dacab18`, 2026-07-29, written for the Anthropic path), and when that floor makes the target impossible it deliberately "let[s] the API report its real context limit". Turn 3's read sat inside the floor, so only ~5k tokens were foldable. design/91's server-fit valve respects the floor.
- **What a fix would look like** — (1) when the calibrated request will not fit the server, let the server-fit valve fold into the recent-turn floor: the context size is a hard limit and the floor is a preference (estimated, not measured: folding turn 3 brings the refused request to roughly 55–60k, which fits; replay the 2026-10-27 history through `model_messages` to confirm); (2) if even a full fold cannot fit (the turn in flight alone is too large), refuse before sending, with a message that persists; (3) bound tool-result size in characters relative to the session's remaining budget (the store knows it), stated in the result as design/82 D3 requires, or have `read_hook_log` cap its returned characters and point at `rank_hook_log`.
- **Where** — LOCAL (replay the recorded history), then the demo machine at `-c 65536`, which reproduces it.
- **Block** — NONE; taken as a follow-up after PR #62 merges (operator, 2026-10-27).
- **Importance** — HIGH (raised from MEDIUM, 2026-10-27): seen twice, on two machines and two context sizes, each time losing the turn.
- **Effort** — SMALL
- **Provenance** — design/91 gate 91 round 3, scored from artifacts 2026-10-06; extended from the PR #62 demo run, 2026-10-27.

### R162 — `test_built_wheel_contains_the_source_tree_skill_catalog` fails intermittently

**Failed in 2 of the full-suite runs during design/91 and passed alone each time (38 s).**

- **What happened** — design/91, 2026-10-02: once at the 91a baseline (before any 91 code, during a network drop) and once on 91b's tip; never reproduced in isolation.
- **Why it matters** — an intermittent in the full suite is read as noise, which is how a real failure in the same test would be missed.
- **What a fix would look like** — capture its failure output on the next occurrence (the runs here did not keep the traceback) and decide whether it builds a wheel over the network.
- **Where** — LOCAL.
- **Block** — NONE.
- **Importance** — LOW
- **Effort** — SMALL
- **Provenance** — design/91 coordinator full-suite runs, 2026-10-02.

### R163 — At `reasoning_effort=low` the local model reported a dataset it never acquired

**Told the operator "Saved to disk: `red_1`, `green_1`, `blue_1` … These are the real, verified acquisitions" when only red and green had been acquired in that turn.**

- **What happened** — design/91, real-microscope session `20261014_141904_322634` (Qwen 3.8 27B, `MICROCLAW_LLM_REASONING_EFFORT=low`), 2026-10-14: the acquisitions record holds two acquisitions for the turn (18:34:34, 18:34:51 UTC); the model switched the turret to DAPI and wrote a composite script referencing a `blue_1` that does not exist. Not seen in the `xhigh` or `medium` sessions of the same script — one session each, so not evidence either way.
- **Why it matters** — it breaks the system prompt's central reporting rule (every claim from a tool result) in the most consequential place: what is on disk. A local model at low effort is the setting an operator chooses for speed.
- **What a fix would look like** — first measure: two or three sessions per effort level, scoring every "saved"/"acquired" claim in the replies against the acquisitions record (scriptable from the session files). If it recurs, a tool-side check is stronger than prompt wording: e.g. a turn-end summary of what was actually acquired, from the acquisitions record, that the model must relay.
- **Where** — WORKSTATION + RIG (the operator's Zeiss).
- **Block** — NONE.
- **Importance** — HIGH
- **Effort** — SMALL to measure
- **Provenance** — design/91 reasoning-effort sessions, scored from artifacts 2026-10-14.

### R164 — No route to combine saved channels into a colour composite

**Asked to "combine into a 3 colour image and save", the model had no tool for it at any reasoning level.**

- **What happened** — design/91 real-microscope sessions, 2026-10-14: `low` wrote an unrun composite script and later gave Fiji steps; the request is ordinary for a three-channel fluorescence image.
- **Why it matters** — a routine deliverable with no route invites exactly the hand-written, unexecuted script the system prompt warns about.
- **What a fix would look like** — an offline adapter on the existing `run_analysis_on_saved_dataset` route (analysis lives in hooks/adapters, CLAUDE.md) that renders a composite from a channel axis or from named single-channel datasets, written as an artifact `open_artifact` can show. Simpler once the rig's `Channel` group gives one multi-channel dataset.
- **Where** — LOCAL, then the rig.
- **Block** — NONE.
- **Importance** — MEDIUM
- **Effort** — MEDIUM
- **Provenance** — design/91 reasoning-effort sessions, 2026-10-14.

### R165 — A local reply with no visible text ends the turn silently and is stored empty

**Qwen3.5-4B called a tool, got its result, then produced only reasoning; the operator saw no answer and the history kept `{"role": "assistant", "content": []}`.**

- **What happened** — PR #62 demo run, 2026-10-27, `unsloth/Qwen3.5-4B-GGUF:Q8_0` at `n_ctx` 131,072: 3 of the 4 turns that reached the model ended this way, e.g. 99 output tokens against 372 reasoning characters and no `delta.content`. Turn 2's mean intensity and SNR were never shown. Qwen3.5-9B on the same machine did not do it (5 of 5 final replies had text). `_stream_one_round`'s local branch builds `blocks = []` and returns `end_turn`.
- **Why it matters** — the user is left with nothing, not even a notice; the answer may have been inside the reasoning, which is deliberately not shown or stored. Unverified: an empty assistant message in the history may be refused by the Anthropic API if the session switches model.
- **What a fix would look like** — when a local reply ends with no text and no tool call, show a notice ("the model returned no visible reply") and do not store the empty message; consider one retry.
- **Where** — LOCAL (a recorded stream with reasoning only).
- **Block** — NONE.
- **Importance** — LOW: model-dependent, and it did not stop the session being usable (operator, 2026-10-27).
- **Effort** — SMALL
- **Provenance** — PR #62 demo run, scored from artifacts 2026-10-27.
18 changes: 13 additions & 5 deletions design/82-block82c-r63-probe.py
Original file line number Diff line number Diff line change
Expand Up @@ -12,10 +12,14 @@
turns, just as in the cost reconstruction. Usage turn IDs must follow those
same boundaries. Missing/misaligned records cannot pin replay partitions.
* The real ConversationStore.model_messages(history[:i]) reproduces the context
under the tree and window being tested. Usage compaction counts are checked
on every call; disagreement is reported and partitions remain replay estimates,
not a claim about the historical context. A usage log from a tree with a
different estimate_tokens pins WHEN compaction fired, not WHERE it cut.
under the tree and window being tested. This is always a replay under the
current tree's compaction policy. Since 84c, D9 commits only when a fold clears
high-water; replaying a pre-84c history can therefore produce a different
partition even when its usage log has matching per-call compaction counts.
Usage compaction counts are checked on every call; disagreement is reported
and partitions remain replay estimates, not a claim about the historical
context. Matching counts pin WHEN compaction fired, not WHERE it cut. A usage
log from a tree with a different estimate_tokens pins WHEN, not WHERE.
Compare current/recorded estimated tokens only while both are uncompacted:
near 1 supports estimator equivalence and replay partition trust; away from 1
indicates a different estimator and a replay context the model never saw.
Expand Down Expand Up @@ -138,12 +142,16 @@ def score(messages, usage=None, *, high_water=DEFAULT_CONTEXT_HIGH_WATER_TOKENS,
return dict(calls=calls, turns=turn, events=events, mentions=mentions,
usage_records=len(usage) if usage is not None else None,
usage_events=usage_events, disagreements=disagreements,
estimate_ratios=estimate_ratios)
estimate_ratios=estimate_ratios,
replay_policy='current ConversationStore (D9 high-water gate)')


def report(result):
print(f"{result['calls']} assistant calls, {result['turns']} turns, "
f"{len(result['events'])} compactions; usage records: {result['usage_records']}")
print('Replay policy: current ConversationStore (D9 high-water gate). '
'A pre-84c session may have different checkpoint partitions; replay '
'partitions are not historical proof.')
if result['usage_records'] is None:
print('Replay estimates only: usage-pinned timestamps and turn IDs unavailable.')
else:
Expand Down
Loading
Loading