GetDCFStores walks every mounted filesystem root for known media extensions (pkg/dcf/dcf.go). That finds any media anywhere, not a DCF store. Per CP-3461: a store is a volume with a root DCIM/; inside are directories NNNABCDE (100–999 + 5 free chars); objects are ABCD + 4-digit file number. Parse that structure — directory numbers, basenames, file numbers — and expose it on the types instead of only path/size/extension. Non-conforming files under DCIM/ are surfaced as such, not silently mixed in.
Done: a fixture card tree parses into stores/directories/objects with their DCF numbers; media outside DCIM/ is not part of a store; extension lists remain only for classifying object type.
`GetDCFStores` walks every mounted filesystem root for known media extensions (`pkg/dcf/dcf.go`). That finds any media anywhere, not a DCF store. Per CP-3461: a store is a volume with a root `DCIM/`; inside are directories `NNNABCDE` (100–999 + 5 free chars); objects are `ABCD` + 4-digit file number. Parse that structure — directory numbers, basenames, file numbers — and expose it on the types instead of only path/size/extension. Non-conforming files under `DCIM/` are surfaced as such, not silently mixed in.
Done: a fixture card tree parses into stores/directories/objects with their DCF numbers; media outside `DCIM/` is not part of a store; extension lists remain only for classifying object type.
Parse one card from an fs.FS rooted at its mount point, in an unexported function; tests call it with fstest.MapFS, and GetDCFStores calls it with os.DirFS(mountpoint). The public entry point does not change here: its shape waits on sneak's answer on #4.
Under DCIM: DCF directories (directory number 100 to 999, then five characters) and in them DCF files (four characters, then file number 0001 to 9999, then the extension), with the character rules the spec gives. Each object carries its directory number, file number and path; each directory its number and name. The extension lists stay, only to say whether an object is an image or a video. Counts and total sizes keep working.
A file or directory under DCIM that breaks the naming rules is listed separately on the store as not conforming, never mixed in with DCF objects.
PRIVATE/M4ROOT stays part of the card: the README names it as where cameras put video, and the current code finds those clips. Media files there are listed as today (path, size, extension, no DCF numbers). Media anywhere else on the volume is no longer part of the store.
Grouping files that share a number is #3, not this.
Test: one fstest.MapFS card with two DCF directories holding objects, a misnamed file and a misnamed directory under DCIM, a clip under PRIVATE/M4ROOT, and a JPEG outside both; assert where each lands.
Depends on #1. Done: as the issue, with PRIVATE/M4ROOT clips kept.
Model: opus-5-5
Plan. Follow the naming rules in the spec itself: CIPA DC-009-2010 is the English text of DCF 2.0 (CP-3461), https://www.cipa.jp/std/documents/e/DC-009-2010_E.pdf.
- Parse one card from an `fs.FS` rooted at its mount point, in an unexported function; tests call it with `fstest.MapFS`, and `GetDCFStores` calls it with `os.DirFS(mountpoint)`. The public entry point does not change here: its shape waits on sneak's answer on https://git.eeqj.de/sneak/dcf/issues/4.
- Under `DCIM`: DCF directories (directory number 100 to 999, then five characters) and in them DCF files (four characters, then file number 0001 to 9999, then the extension), with the character rules the spec gives. Each object carries its directory number, file number and path; each directory its number and name. The extension lists stay, only to say whether an object is an image or a video. Counts and total sizes keep working.
- A file or directory under `DCIM` that breaks the naming rules is listed separately on the store as not conforming, never mixed in with DCF objects.
- `PRIVATE/M4ROOT` stays part of the card: the README names it as where cameras put video, and the current code finds those clips. Media files there are listed as today (path, size, extension, no DCF numbers). Media anywhere else on the volume is no longer part of the store.
- Grouping files that share a number is https://git.eeqj.de/sneak/dcf/issues/3, not this.
Test: one `fstest.MapFS` card with two DCF directories holding objects, a misnamed file and a misnamed directory under `DCIM`, a clip under `PRIVATE/M4ROOT`, and a JPEG outside both; assert where each lands.
Depends on https://git.eeqj.de/sneak/dcf/issues/1. Done: as the issue, with `PRIVATE/M4ROOT` clips kept.
Model: opus-5-5
State, 2026-10-03: queued behind #1, not dispatched; can run alongside #6. The plan above is the worker's brief.
Model: opus-5-5
State, 2026-10-03: queued behind https://git.eeqj.de/sneak/dcf/issues/1, not dispatched; can run alongside https://git.eeqj.de/sneak/dcf/issues/6. The plan above is the worker's brief.
Model: opus-5-5
Two points the plan above leaves open, settled so the work can start:
A file under DCIM with a valid DCF name whose extension is on neither list, such as a .THM, is still a DCF object. It counts as neither an image nor a video. #3 groups it with its video, so it must not land among the files that break the naming rules.
A card with no DCIM, or no PRIVATE/M4ROOT, reads normally. The missing directory is not an error.
Model: opus-5-5
Two points the plan above leaves open, settled so the work can start:
- A file under `DCIM` with a valid DCF name whose extension is on neither list, such as a `.THM`, is still a DCF object. It counts as neither an image nor a video. https://git.eeqj.de/sneak/dcf/issues/3 groups it with its video, so it must not land among the files that break the naming rules.
- A card with no `DCIM`, or no `PRIVATE/M4ROOT`, reads normally. The missing directory is not an error.
Model: opus-5-5
Built in #11. Its description lists the spec rules the plan did not spell out, and two judgement calls.
Model: opus-5-5
Built in https://git.eeqj.de/sneak/dcf/pulls/11. Its description lists the spec rules the plan did not spell out, and two judgement calls.
Model: opus-5-5
State for the next manager of this repo, 2026-11-03. sneak stopped work on all repos except project-management, prompts and homoicon, so all work here is stopped: no worker or reviewer is running and no container is left.
#14 (#5): needs-review, assigned to clawbot, head d2728b1, no failed review. Its first review was stopped partway and posted nothing, so it needs a fresh review from the start.
#9 (next into master): merge-ready, assigned to sneak, head acb108a, no failed review. Its description is current. It is with him because it is ready for him to merge.
What is left, in order
A fresh review of #14, any rework, then a squash-merge. After the merge, run script/cibuild on next, add the import to the description of #9 and keep its label true.
One audit of next, filing and fixing every bug it finds. Start with these, which earlier briefs kept out of scope:
The README's opening sentence credits the DCF specification for PRIVATE/M4ROOT, and the specification never mentions it. sneak wrote that sentence.
The README's line saying dcf does not work yet.
isImageFile lists .raw and .arw but not other RAW formats such as NEF and CR2, so those count as neither images nor videos.
One short state comment on #9 saying it is finished with no known open bugs, with links.
The ruling is #4 (comment) and the plan is #5 (comment). The plan's names and its three judgement calls are settled. Ask the reviewer to try hard to lose data: overwriting, a differing copy kept, a partial copy left behind, a swallowed sync or close error, two files with one name and one day. The three points listed for the audit are not findings on this PR, and neither is the later file renaming.
Rule on failed reviews (sneak/project-management#43 (comment)): on a PR's third failed review, its manager steps into the PR with very specific instructions that end the loop: what the rework changes, what it leaves alone, and what the next review checks. No PR is paused for sneak or assigned to him over failed reviews.
For every worker brief
script/bootstrap replaces ~/.local/go when the Go there is another version, and that folder is shared on this host. Run script/cibuild and make targets with a private HOME, and run chmod -R u+w on it before deleting it.
Hold the build behind a lock: flock -w 1800 -E 75 /srv/code/tmp/.dcf-gate.lock script/cibuild. Exit 75 means it timed out and did not run.
Set the commit identity in each clone: clawbot, sneak+clawbot@sneak.cloud.
No gitea MCP tool was available in these sessions. Use the API with curl and $GITEA_TOKEN.
The wrappers are ~/bin/claude3 and ~/bin/claude2, which are not on the PATH.
A stopped session's row disappears from agents --json, so a loop waiting on a session must treat a missing row as finished.
Model: opus-5-5
State for the next manager of this repo, 2026-11-03. sneak stopped work on all repos except project-management, prompts and homoicon, so all work here is stopped: no worker or reviewer is running and no container is left.
**Done, on `next` at `acb108a`**
- https://git.eeqj.de/sneak/dcf/issues/1 through https://git.eeqj.de/sneak/dcf/pulls/8. Renaming `master` still waits on https://git.eeqj.de/sneak/project-management/issues/10.
- https://git.eeqj.de/sneak/dcf/issues/6 through https://git.eeqj.de/sneak/dcf/pulls/10.
- https://git.eeqj.de/sneak/dcf/issues/2 through https://git.eeqj.de/sneak/dcf/pulls/11.
- https://git.eeqj.de/sneak/dcf/issues/3 through https://git.eeqj.de/sneak/dcf/pulls/12.
- https://git.eeqj.de/sneak/dcf/issues/4 through https://git.eeqj.de/sneak/dcf/pulls/13.
**Open PRs**
- https://git.eeqj.de/sneak/dcf/pulls/14 (https://git.eeqj.de/sneak/dcf/issues/5): `needs-review`, assigned to `clawbot`, head `d2728b1`, no failed review. Its first review was stopped partway and posted nothing, so it needs a fresh review from the start.
- https://git.eeqj.de/sneak/dcf/pulls/9 (`next` into `master`): `merge-ready`, assigned to `sneak`, head `acb108a`, no failed review. Its description is current. It is with him because it is ready for him to merge.
**What is left, in order**
1. A fresh review of https://git.eeqj.de/sneak/dcf/pulls/14, any rework, then a squash-merge. After the merge, run `script/cibuild` on `next`, add the import to the description of https://git.eeqj.de/sneak/dcf/pulls/9 and keep its label true.
2. One audit of `next`, filing and fixing every bug it finds. Start with these, which earlier briefs kept out of scope:
- The README's opening sentence credits the DCF specification for `PRIVATE/M4ROOT`, and the specification never mentions it. sneak wrote that sentence.
- The README's line saying `dcf` does not work yet.
- `isImageFile` lists `.raw` and `.arw` but not other RAW formats such as `NEF` and `CR2`, so those count as neither images nor videos.
3. One short state comment on https://git.eeqj.de/sneak/dcf/pulls/9 saying it is finished with no known open bugs, with links.
**For the review of https://git.eeqj.de/sneak/dcf/pulls/14**
The ruling is https://git.eeqj.de/sneak/dcf/issues/4#issuecomment-145367 and the plan is https://git.eeqj.de/sneak/dcf/issues/5#issuecomment-165158. The plan's names and its three judgement calls are settled. Ask the reviewer to try hard to lose data: overwriting, a differing copy kept, a partial copy left behind, a swallowed sync or close error, two files with one name and one day. The three points listed for the audit are not findings on this PR, and neither is the later file renaming.
**Rule on failed reviews** (https://git.eeqj.de/sneak/project-management/issues/43#issuecomment-162903): on a PR's third failed review, its manager steps into the PR with very specific instructions that end the loop: what the rework changes, what it leaves alone, and what the next review checks. No PR is paused for sneak or assigned to him over failed reviews.
**For every worker brief**
- `script/bootstrap` replaces `~/.local/go` when the Go there is another version, and that folder is shared on this host. Run `script/cibuild` and `make` targets with a private `HOME`, and run `chmod -R u+w` on it before deleting it.
- Hold the build behind a lock: `flock -w 1800 -E 75 /srv/code/tmp/.dcf-gate.lock script/cibuild`. Exit 75 means it timed out and did not run.
- Set the commit identity in each clone: `clawbot`, `sneak+clawbot@sneak.cloud`.
- No gitea MCP tool was available in these sessions. Use the API with curl and `$GITEA_TOKEN`.
- The specification is at https://www.cipa.jp/std/documents/download_e.html?DC-009-2010_E. The link in the older plan comments is dead.
**For the manager**
- The wrappers are `~/bin/claude3` and `~/bin/claude2`, which are not on the `PATH`.
- A stopped session's row disappears from `agents --json`, so a loop waiting on a session must treat a missing row as finished.
Model: opus-5-5
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
GetDCFStoreswalks every mounted filesystem root for known media extensions (pkg/dcf/dcf.go). That finds any media anywhere, not a DCF store. Per CP-3461: a store is a volume with a rootDCIM/; inside are directoriesNNNABCDE(100–999 + 5 free chars); objects areABCD+ 4-digit file number. Parse that structure — directory numbers, basenames, file numbers — and expose it on the types instead of only path/size/extension. Non-conforming files underDCIM/are surfaced as such, not silently mixed in.Done: a fixture card tree parses into stores/directories/objects with their DCF numbers; media outside
DCIM/is not part of a store; extension lists remain only for classifying object type.Plan. Follow the naming rules in the spec itself: CIPA DC-009-2010 is the English text of DCF 2.0 (CP-3461), https://www.cipa.jp/std/documents/e/DC-009-2010_E.pdf.
fs.FSrooted at its mount point, in an unexported function; tests call it withfstest.MapFS, andGetDCFStorescalls it withos.DirFS(mountpoint). The public entry point does not change here: its shape waits on sneak's answer on #4.DCIM: DCF directories (directory number 100 to 999, then five characters) and in them DCF files (four characters, then file number 0001 to 9999, then the extension), with the character rules the spec gives. Each object carries its directory number, file number and path; each directory its number and name. The extension lists stay, only to say whether an object is an image or a video. Counts and total sizes keep working.DCIMthat breaks the naming rules is listed separately on the store as not conforming, never mixed in with DCF objects.PRIVATE/M4ROOTstays part of the card: the README names it as where cameras put video, and the current code finds those clips. Media files there are listed as today (path, size, extension, no DCF numbers). Media anywhere else on the volume is no longer part of the store.Test: one
fstest.MapFScard with two DCF directories holding objects, a misnamed file and a misnamed directory underDCIM, a clip underPRIVATE/M4ROOT, and a JPEG outside both; assert where each lands.Depends on #1. Done: as the issue, with
PRIVATE/M4ROOTclips kept.Model: opus-5-5
State, 2026-10-03: queued behind #1, not dispatched; can run alongside #6. The plan above is the worker's brief.
Model: opus-5-5
Two points the plan above leaves open, settled so the work can start:
DCIMwith a valid DCF name whose extension is on neither list, such as a.THM, is still a DCF object. It counts as neither an image nor a video. #3 groups it with its video, so it must not land among the files that break the naming rules.DCIM, or noPRIVATE/M4ROOT, reads normally. The missing directory is not an error.Model: opus-5-5
Built in #11. Its description lists the spec rules the plan did not spell out, and two judgement calls.
Model: opus-5-5
State for the next manager of this repo, 2026-11-03. sneak stopped work on all repos except project-management, prompts and homoicon, so all work here is stopped: no worker or reviewer is running and no container is left.
Done, on
nextatacb108amasterstill waits on sneak/project-management#10.Open PRs
needs-review, assigned toclawbot, headd2728b1, no failed review. Its first review was stopped partway and posted nothing, so it needs a fresh review from the start.nextintomaster):merge-ready, assigned tosneak, headacb108a, no failed review. Its description is current. It is with him because it is ready for him to merge.What is left, in order
script/cibuildonnext, add the import to the description of #9 and keep its label true.next, filing and fixing every bug it finds. Start with these, which earlier briefs kept out of scope:PRIVATE/M4ROOT, and the specification never mentions it. sneak wrote that sentence.dcfdoes not work yet.isImageFilelists.rawand.arwbut not other RAW formats such asNEFandCR2, so those count as neither images nor videos.For the review of #14
The ruling is #4 (comment) and the plan is #5 (comment). The plan's names and its three judgement calls are settled. Ask the reviewer to try hard to lose data: overwriting, a differing copy kept, a partial copy left behind, a swallowed sync or close error, two files with one name and one day. The three points listed for the audit are not findings on this PR, and neither is the later file renaming.
Rule on failed reviews (sneak/project-management#43 (comment)): on a PR's third failed review, its manager steps into the PR with very specific instructions that end the loop: what the rework changes, what it leaves alone, and what the next review checks. No PR is paused for sneak or assigned to him over failed reviews.
For every worker brief
script/bootstrapreplaces~/.local/gowhen the Go there is another version, and that folder is shared on this host. Runscript/cibuildandmaketargets with a privateHOME, and runchmod -R u+won it before deleting it.flock -w 1800 -E 75 /srv/code/tmp/.dcf-gate.lock script/cibuild. Exit 75 means it timed out and did not run.clawbot,sneak+clawbot@sneak.cloud.$GITEA_TOKEN.For the manager
~/bin/claude3and~/bin/claude2, which are not on thePATH.agents --json, so a loop waiting on a session must treat a missing row as finished.Model: opus-5-5