Operator inbox gaps against Paperclip — 3 November 2026
This text is LLM generated.
For review#227: the operator wants to see and answer work here, rather than receive another list of links in chat. The evidence below supports three additional gaps, not five. The larger problems that more often force the operator to leave the app—missing concrete questions, incomplete coverage, unavailable writes and missing run context—are already assigned. Counting them again would create a longer backlog, not a better inbox. These are issue-ready suggestions for the controller, not filed issues or implemented features. Ranking estimates frequency and consequence from this snapshot and the reported workflow; it is not measured operator telemetry.
Evidence boundary and deployment identity
Review main and this assessment's source base were
f71d4e46814b4660b27714a1ab228e39b857a586; the GitHub main API confirmed that
SHA during inspection. Paperclip was inspected at
909d86b59a841d5f014cdb84909956d00625d9d0 (its remote HEAD when cloned).
All Paperclip links below pin that commit; this is not a claim about its cloud
deployment or any operator's Paperclip installation.
Review's Pages run
succeeded at the review SHA. Deployment 7528677052 reported success at
2026-11-03T19:35:45Z for https://cgwalters-forge.github.io/review/.
Publicly fetched app.js and style.css matched a clean local npm run build:
| Asset | SHA-256, both local and served |
|---|---|
| app.js | 716198de84f9caaf362f2411ed1271af2506aaee5e02b222fe7a76b9159ebe91 |
| style.css | fdc30d6fb3bf80e0b5d81725eee6b37464a36e2b9a120439745c17ff2c218754 |
That verifies the served assets at observation time, not signed-in behavior, private publication contents, cache state on the operator's device or future deploys. The legacy Pages-build endpoint returned 404; deployment statuses, the Actions run and asset comparison supplied the identity instead.
The fixture is the approved redacted publication at
2026-11-03T17:41:35.039037224Z.
Its provenance PR says
spend/overhead and one security-sensitive item were removed. Do not infer
absence of costs or of that item from this fixture. It contains 12 ranked
asks and 12 descriptors, no groups, no approval heads and no supported
descriptor actions. Six descriptors are missing-concrete-decision; the
first three P0 titles are only “labelled needs-human”. Items and digest are
partial, with 85 omitted asks. Twelve is therefore not the operator's total
backlog. The older top-level holds list is not an extra inbox to append.
Already known: excluded after re-reading bodies and comments
The tracker sub-issue API returned these six open children. This list includes this assessment itself; closed historical children are not new candidates. Issue comments are decision/status evidence, not instructions for this run.
| Open child | Excluded work, including latest context |
|---|---|
| review#221 | Honest exact group membership, no representative write, real-data replay. #224 shipped inspect-only groups. The latest plan acknowledges the real fixture has zero groups; constructed variants must not be called observed real groups. |
| review#222 | Per-request actions, promotion, receipts and partial outcomes. #225 shipped descriptor-aware refusals; latest decision authorizes promotion by commit-bound APPROVE, not bare /promote. Reply/close/merge/enqueue refusals are accepted until publication offers those actions. |
| agent-board#640 | One producer-owned inbox including decisions/drafts, counts, source completeness/clocks, recent controller decisions/run receipts, working/waiting/retry/WIP and relevant budget/pool holds. Includes publication/caller integration, not just a schema. Generic ask quality and coverage belong with this producer read-model work, not a new dashboard proposal here. |
| review#223 | Consume that publication, delete browser discovery/ranking, immediate outstanding counts, retained drafts, freshness and collapsed context; measure request/JS reduction. |
| agentic-job#536 | On-demand ACP session, runtime writes, reconnect and real tailnet acceptance. Latest triage says bridge/spec/model/probes exist; runtime writes await credential decisions, real end-to-end checks await wrapper/tailnet work. Not a new “add chat” gap. |
| review#227 | This comparison; no follow-up issue creation in this run. |
The parent tracker#976 also excludes rediscovering the general focused/live/actionable-in-place requirements. Every suggestion below stays in Needs you, its existing item/PR pane or the existing Status surface. Publication is producer-owned; no new tabs, scheduler, activity poller, browser interpretation of prose commands or new orchestration transport is proposed.
Operator walk-through, not a feature scorecard
Decide what needs me. Review preserves producer Priority/age order and truthfully warns about incompleteness (card selection, renderer). The first three generic asks still require inspection; later asks contain real alternatives and recommendations. A fresh timestamp does not make the missing question concrete. This is primarily existing producer/action work, not an additional gap.
Paperclip's operator guide describes pending approvals, rejection and revision requests. Its approval detail actually loads comments and linked tasks and implements those mutations. Those are Paperclip governance approvals, not evidence that it safely promotes a forge PR or implements review's exact-head semantics. Its Inbox mixes approval/run/join-request attention with tasks. The useful comparison is context next to a decision, not copying agent hiring or CEO controls.
Inspect, then answer. Review already has guarded inline answers, document
choices, draft sending, head-checked approval and proposal execution; it is
incorrect to say “no inline answers”. Main-hook tests exercise
real handlers with mocked GitHub, including changed asks, wrong viewers and
uncertain receipts. This snapshot publishes no supported reply actions, but
that does not prove production answer forms are absent: the current
renderer checks the named descriptor capability for Approve/Do, while Answer
can still call its target-reading hook without checking the reply capability.
The diagnostic below omits that hook, which alone explains why it shows no forms.
Signed-in eligibility and any mismatch with the published reply policy need
verification under #222, not a new gap or a mobile rendering claim.
Paperclip has approval comments/actions and conditional inline
resolution (attention tests); neither proves universal
free-form answers on every inbox row. Finishing review's write coverage is
#222/#640 work.
Check what happened or who is working. Review's item pane already loads comments, gists and tracker sub-issues, but displays only the last five comments (loader, view, limit). Status displays published Markdown, not a per-agent control plane. Paperclip has agent run history, including selecting an older run when the recent list is empty (source, desktop/mobile test), and task history expansion (source, tests). Recent working, waiting, retry and outcome context is already #640/#223. Older task evidence is a narrower additional gap below; a new Agents tab is not justified.
Understand resource pressure and purpose. Paperclip's Costs source
has summary and agent/project/provider/biller breakdown queries and date
windows; accounting tests cover usage availability. The useful
shape is period → pool/provider → task/agent contribution → known/unknown,
not an amount or a public financial dashboard. Review intentionally removed
usage-dashboard polling; budget/pool hold reasons and reset/observation facts
are already #640. This redacted snapshot doesn't provide evidence of an additional
cost-visibility defect. Paperclip Goals and its API docs
provide organization-scoped goal hierarchy/project links. Review has Org and
tracker epics/sub-issues instead. There is no evidence this operator needs a
company hierarchy editor; only typed decision context is useful (rank 1).
Return later or from a phone. Review's From people fold handles human mentions/review requests and notification mark-read, with explicit token limitations (documentation, tests). It is not change tracking for bot decision cards. Paperclip's Inbox has per-item read/unread state and dismissals (source, tests); this does not establish OS push delivery or notification reliability. Review's CSS wraps titles and identifiers, uses 16px touch fields and retains focused drafts (CSS, tests). Paperclip has mobile-specific run handling, but no signed-in phone comparison was performed. The narrow replay below supports a density observation, not a separate proven mobile bug.
Shipped, gated, proposed, unverified
The review behaviors above are in source on the pinned main, not claims based on its older design proposal. Action refusals are shipped; commit-bound promotion completion and unified publication are open work, not already accepted features. The replay is not a session-chat integration test.
Paperclip's streamlined shell is default-on, even during settings loading/failure. Its combined Inbox + Tasks navigation is explicitly opt-in, off by default; it is not the default comparison and is not a reason to add tabs here. Task thread presentation also switches between classic and newer components in the inspected source. Source/tests support the capabilities cited, not production availability of every branch. Paperclip tests were inspected, not executed, and its server was not started.
Additional ranked entries for the controller
1. Publish typed related evidence beside a decision
Problem: the operator must copy a reference or open the question issue and hunt again for the actual work. The fixture's tracker#12 names upstream PRs #408/#409, tracker#308 names a forge PR, tracker#7 names alternative stack members, and tracker#430 names a blocked child. These are references inside display text, not explicit navigable context in the card. The card title links to the ask itself, while the consumed contract has one action target and optional PR-group membership, not general related evidence or an epic/purpose breadcrumb. Existing item sub-issues are not a replacement for links on the question card. Paperclip's linked approval tasks illustrate the value without requiring its org model.
Fix / owner: agent-board publication data, with review consumption
only after the contract exists. Publish a bounded, typed related set with
relationship labels (reviewed PR, prerequisite PR, blocked child, parent epic),
safe reference parts and an observation time. Include only relationships the
producer can establish; unknown stays unknown. Review reuses its existing
item/PR deep links in one collapsed card section. Do not turn URLs or commands
in question text into authority, infer a group or add actions. This is distinct
from #221's exact group members and #222's per-member action eligibility, and
from #640's recent status/outcome context.
Size: M—one producer contract extension plus a small existing-card reader; relationship attribution/validation is the work, not new navigation.
Acceptance: on a derived, clearly labelled replay of these four asks, reach each published related item in the existing pane and return to the same unsent decision. No arbitrary published URL is followed; malformed/foreign references stay visibly unavailable; no extra discovery request occurs on initial load or unchanged poll. Cap the list (for example eight references per ask), disclose omissions and do not alter producer order/counts.
2. Distinguish an unseen decision revision from a familiar pending ask
Problem: a return visit presents old and changed bot questions alike. The snapshot includes several multi-day waits; repeatedly rereading them is work, while overlooking a changed premise is consequential. The publication has per-request identity/revision but review keeps only quality/forge/actions. Card signatures retain DOM during a session, not “seen this revision” across visits. Global changed/validated clocks and human notification read-state do not answer which bot decision changed. Paperclip has read-state integration and unread-row tests; copy the distinction, not its tabs or archive behavior.
Fix / owner: review. Retain a validated opaque per-request revision from the existing publication, and record locally which revision the viewer explicitly inspected. Display “new/changed since inspected” on the existing card; use the same session/remember/sign-out boundaries as the app's cache. Do not mark a decision done, hide it, reorder it, release a hold or mark a GitHub notification read. If identity/revision is missing, show unknown. No server notification system or new poll is necessary.
Size: M—small parser/view change, but identity changes, partial reads, sign-out and persistence need careful tests. This is recognition state, not #223's outstanding-count or #640's outcome-publication work.
Acceptance: inspect one request, revisit unchanged, then replay a new revision under the same target: only the last is “changed”. A refreshed validation clock alone never marks all cards changed; a partial/carried source cannot mark unseen work read. Counts/order and retained text remain unchanged; sign-out clears viewer state. Test session-only and remembered modes.
3. Let the existing item pane reveal older task comments
Problem: after seeing a current ask or receipt, the operator cannot inspect an earlier answer or rationale here once it falls outside the last five comments. loadContext reads comments but slices them; contextView offers “Latest comments”, not an older-history disclosure. This is visible source behavior, not proof that any particular omitted comment in the redacted snapshot is decisive. It is less frequent than identifying the work itself. Paperclip's task history source and older-history tests provide the relevant contrast; its agent roster and live-run controls do not.
Fix / owner: review. Add an explicit, bounded “Earlier comments” disclosure to the existing on-demand item context, reusing the comment renderer and GitHub cache. Keep chronological order, attribution, links and a visible history limit/error. No activity feed, run discovery, per-agent polling or historical reconstruction from bot prose. Recent run receipts/agent state remain #640/#223; this entry is only older forge conversation evidence.
Size: S—existing loader, cache and renderer; selection/disclosure and tests rather than a new publication or service.
Acceptance: a mocked item with seven comments initially shows the last five, reveals the two earlier ones only on demand, and preserves an unsent answer. A longer thread has a stated bound and explicit incomplete/error state, never an unbounded history fetch. No new cold-inbox or poll requests.
Replay and verification record
This patch is documentation-only: no product code, replay tests, fixtures, dependencies or configuration were changed. Existing checks were still run to support the assessment. No token was supplied to the UI; no live writes or private source reads were made.
Commands and results:
npm ci: exit 0; installed dependencies. Audit reported two existing vulnerabilities (one high, one critical); dependency remediation is outside this assessment.node --test test/holds.test.ts test/inboxview.test.ts test/maincards.test.ts test/style.test.ts test/peopleview.test.ts: exit 0, 133 passed, zero failures, test-runner duration 3.86 seconds.node /tmp/opencode/replay.mjs: temporary diagnostic using the existingtest/helpers.tsjsdom setup, real parser/card selector/renderer and frozen fixture time. First attempt exited 1 because the diagnostic looked for Show more inside the list; production inserts it as a sibling. Corrected diagnostic exited 0: 12 asks/cards in fixture order, zero refusals/groups, five initially drawn, 12 after two Show more clicks, no forms. It supplied no action/detail hooks and forbade fetch, so it checks display/selection, not full signed-in main orchestration or write eligibility.chromium --headless --no-sandbox --disable-gpu --no-proxy-server --screenshot=/tmp/opencode/replay-desktop.png --window-size=1280,1000 file:///tmp/opencode/replay-initial.htmland the same command withreplay-narrow.png,390,844: exit 0. Inspected both screenshots of the frozen rendered cards with the real stylesheet. Desktop shows all five initial cards; narrow view shows approximately three and the start of the fourth. Titles/identifiers wrap and the partial-source notice remains readable. Generic titles and repeated unsupported-action folds consume substantial phone space. No clipped title defect was observed. Chromium emitted desktop-service/component-update warnings, not a render failure. These are desktop Chromium viewports, not touch/Safari/keyboard emulation. Screenshots contain only approved redacted data and no usage amounts; binaries are deliberately not added to the patch.npm run build: exit 0.sha256sum dist/app.js dist/style.cssandcurl -fsS https://cgwalters-forge.github.io/review/app.js | sha256sum(alsostyle.css) gave the matching hashes recorded above.
To repeat the display diagnostic without a bespoke app/server: use
installDom() and fixture() from test/helpers.ts; pass the fixture's
needs_you to parseNeedsYou, then publishedCards, then fillInbox with
now = Date.parse(fixture.at), render = createRenderer(window), mayAct = true
and no action/detail hooks. Attach the list to a parent before filling it;
click the parent's .inbox-more twice. Save the initial DOM with the real
static/style.css and a viewport meta tag for the Chromium commands above.
This is the existing mocked DOM harness path, not a newly committed replay
test. The synthetic group replay in test/inboxview.test.ts remains synthetic.
Evidence links were checked for local file/anchor existence and pinned GitHub file/line existence, and the linked issues/deployment/run were read through GitHub. Deduplication used the complete open-child list above and latest issue comments, not just titles. New text reproduces no cost amounts, private publication payloads or credentials; the fixture was not copied or modified.
Uncertainties and recommended next slice
The assessment found two fewer additional gaps than the requested minimum of five. Notifications beyond in-app read state, per-task usage attribution, a new org model, push delivery and mobile interaction defects lack sufficient operator evidence to rank as separate new issues. Layout density is real in this replay, but its chief causes are the already-known generic asks and refusal coverage, not proof that review needs a mobile redesign.
Signed-in deployed acceptance, real request writes, tailnet sessions, touch keyboard/back navigation and Paperclip deployment behavior remain unverified. The partial snapshot cannot measure whole-backlog frequency or an earlier comment's decision value. Recheck open children before filing: parallel work may close issues or subsume these suggestions after this assessment.
Top next slice: finish #640/#223/#222's existing concrete-question and in-place-action acceptance before expanding scope. Of the genuinely additional entries, start with rank 1's bounded related-evidence publication, then its existing-pane reader: it removes the second navigation hunt on real multi-PR questions without restoring dashboards or browser orchestration. No new option or control is introduced by this documentation patch.