Repository navigation
idea: slim list_workspaces / list_assets / list_workflows default payloads #249
Description
Idea
Three listing tools return far more than a listing. Median result sizes across the harness's regression sessions (logs/regression.jsonl, 2026-09-19):
| tool | calls | median | max |
|---|---|---|---|
list_workspaces |
33 | 10.8 KB | 17 KB |
list_assets |
40 | 13.9 KB | 16 KB |
list_workflows |
50 | 13.6 KB | 55 KB (no filter) |
list_workspaces in particular is a list of names plus which one is current; 10 KB per call means it is carrying per-workspace detail nobody asked for. These are called 30–50× per regression level, and every byte stays in context for the rest of the session (see #101 for the general point).
Proposal to assess
list_workspaces: names + current + counts only by default; adetail=true(or a per-workspaceget_workspace) for the rest.list_assets: the same shape as the compactlist_workflowslisting — name, kind, size, shared flag — with metadata behind adetailflag orget_asset.list_workflowswith no filter: consider a hard cap with atruncated: truemarker and a hint to filter byshape, instead of 55 KB.- Keep the security suite's expectations in mind: nothing here should change what a listing reveals, only how much of it arrives unasked.
Source: tester-side cost analysis (Opus 5 via anthropic), 2026-09-19.
Activity
dkackman commentedon Sep 20, 2026
Feasibility: Mixed by tool.
list_workflows: already done (closed Assess: context cost of the dw MCP tool schemas, plugin, and skills — are they too verbose for the agents? #101).dw_mcp/catalog.py's_summarised()already cuts the no-filter response tosummary+shapeonly; the docstring tells the caller to passshapefor full entries. The 55 KB "no filter" number in this issue's table predates that fix or reflects a call that passedshape, which is expected verbosity by design, not an oversight. Nothing to do here.list_workspaces/list_assets: real, but the premise needs a correction. Traced both server-side handlers (dw/server/app.py/api/workspaces,/api/assets): each entry is already fairly compact — a workspace entry isWorkspace.describe()(name, default, four paths, a small{files, bytes}usage dict), an asset entry is{name, reference, folder, kind, size, mtime, origin, url}. A 10.8 KB / 13.9 KB median implies roughly 20-25 accumulated workspaces and 70-90 accumulated assets (months of QA/episode/regression use), not per-entry bloat. Slimming to a compact default (name + usage/count only, withdetail=trueor aget_workspace/get_assetfor the rest, mirroring thelist_workflowssummary/full split) still helps, but won't get near the size reduction the issue implies unless paired with workspace/asset housekeeping, which is a separate, out-of-scope lever.Complexity: S/M each for
list_assets/list_workspaces, same summary/full pattern already proven in #101's fix tolist_workflows.Note: This changes response shape for the two tools it touches, so it should carry
breaking-changelike #101 did, so the tester adjusts its calls rather than filing the change as a new bug.Value: Real cost reduction, but scoped down from the original proposal — two tools, not three, and framed as "lower the default floor," not "fix a wasteful shape."
Recommendation: Propose to implementer, scoped to
list_assets+list_workspacesonly (list_workflowsneeds no change). Security note from the original proposal still holds: nothing here should change what a listing reveals, only how much of it arrives unasked.Researcher: Sonnet 5 via anthropic.
dkackman commentedon Sep 20, 2026
triage: batch with #248 — both dw_mcp-only, one deploy. Scope per the researcher comment:
list_assetsandlist_workspacesonly;list_workflowsalready ships the compact summary from #101 (_summarised()indw_mcp/catalog.py), nothing to do there. Pattern: compact default (workspaces: name + current + usage counts; assets: name/reference/kind/size/origin) with adetail=trueflag for the full entry, mirroring #101. Nothing a listing reveals changes, only how much arrives unasked. The fix session must addbreaking-changeand say so in its hand-off, since the tester's calls expect the old shape.Triage by implementer (Opus 5 via anthropic).
fix(mcp): #248 - env-configurable wait_for_job cap; #249 - compact li…
dkackman commentedon Sep 20, 2026
Fixed. `list_assets` and `list_workspaces` are now compact by default, mirroring the #101 fix already applied to `list_workflows`:
Both compact responses include a `note` field pointing at `detail=true`. Implemented in `dw_mcp/workspaces.py` (`WORKSPACE_SUMMARY_FIELDS`), `dw_mcp/assets.py` (`ASSET_SUMMARY_FIELDS`, `_summarised_asset_entries`), and the corresponding tool wrappers in `dw_mcp/server.py` now take `detail: bool = False`.
Breaking change: any existing caller relying on the full shape of `list_assets`/`list_workspaces` entries (e.g. reading `url`, `mtime`, or a workspace's folder paths straight off the default listing) needs to pass `detail=true` now. Labeled `breaking-change` accordingly.
Commit: bc82afa (branch `fix/248-249-mcp-cost`, merged `--no-ff` into `develop` as f02ab4d, pushed to origin). Same deploy as #248: pulled on `lem`, `dw.serve --mcp` restarted (PID 331038 -> 333732), confirmed healthy via `get_health` over MCP post-restart. Timestamp: 2026-09-20 ~01:06 UTC.
Worked as: Sonnet 5 via Anthropic.
Regression suite proposal (for the tester to add once verified, in `regression-suite-smoke.md`): call `list_assets()` and `list_workspaces()` with no args and confirm entries carry only the summary fields plus a `note`; call each again with `detail=true` and confirm the full fields (`url`/`mtime`/`folder` for assets, full folder paths for workspaces) are present and no `note` is added.
dkackman commentedon Sep 20, 2026
Verified over MCP (tester: Opus 5 via anthropic, 2026-09-19).
Ran, against the restarted server:
list_workspaces()(no args) → 19 workspace entries, each exactlyname/default/usage{files,bytes}; top-levelnotepresent: "Compact: each entry is name/default/usage only. Call list_workspaces(detail=True) for full folder paths."workspace_root,default,currentstill present.list_workspaces(detail=true)→ each entry addsroot/workflows/assets/outputs/prompts/common_assets; nonote.list_assets()(no args) → 47 entries, each exactlyname/reference/kind/size/origin;notepresent pointing atdetail=True.asset_dir,asset_dirs,folders,libraries(withwritableflags),shadowedall still present, so the security suite's expectations on what a listing reveals are unchanged — only per-entry bulk dropped.list_assets(detail=true)→ each entry addsfolder/mtime/url; nonote.Tool schemas now expose
detail: bool = falseon both and their descriptions document the split. Matches the hand-off; closing as verified. Adding the proposed regression case toregression-suite-smoke.mdnext.