…enderers, never a raw dump (#2162) (#2173) * docs(requirements): reports render on three surfaces, not two (#2162) FR-5 enumerated Agent Detail and the Operations fleet tab while a third surface — the Workspace agent page — shipped a raw JSON dump to external clients. Enumerating the surfaces is load-bearing, so the list is corrected first: the omission is the hole the bug fell through. Also records the two things the surfaces deliberately do NOT share: the fallback is a bounded, humanised, redacted key-value summary (never a raw dump), and the Workspace passes allow-raw=false so no raw payload is reachable behind it. FR-5a records the client-facing row window as two optional query params on the EXISTING portal detail route rather than a second route a portal principal could not reach anyway (#2128 class). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(workspace): window a client's tabular report instead of shipping it whole (#2162) Two optional query params on the EXISTING portal detail route, not a second route. The operator row reader (#1537) is Depends(get_current_user), which a portal principal — a verified email with no users row — structurally cannot satisfy; cloning it on a client-facing prefix would mean a second hand-written gate beside the uniform-404 one, and that is how those two answers drift apart. The server decides tabularity from the REAL payload, so the client never guesses: display_hint is agent-authored and can disagree with what was filed, and a client-side prediction would need a 400 plus a recovery re-fetch. A non-tabular payload with rows_limit set comes back whole with no row_meta. rows_limit absent is byte-identical to today. Rate-limited after the roster gate: paging re-reads the whole (<=5 MiB) blob per request, so the route that cuts transfer raises reads — acceptable behind a JWT, an amplification primitive on a prefix a client can loop. Also fixes two stale "256 KB" comments; the cap has been 5 MiB since #1537. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(workspace): render reports through the shared renderers, never dump them (#2162) The Reports tab shipped JSON.stringify(payload) to external clients. That is a disclosure defect, not only an ugly one: agent_page.py already refuses to expose an ask's `context` because free-form agent JSON has been a credential-leak surface (canary G-04), and `payload` is the same category. A typed renderer reads only the keys its hint declares, so routing the tab through the shared components/reports/ set strictly narrows what crosses. The fallback could not be reused as-is, because the shared fallback IS the raw JSON viewer. ReportSummary replaces it on EVERY surface — bounded (<=40 entries, truncated values, depth 1), humanised, with credential-shaped tokens redacted at value level anywhere in the string. Raw JSON moves one click behind a disclosure for operators; `allow-raw="false"` removes even that here. The prop names the POLICY rather than a mechanism, so it covers both a shape mismatch and an agent that deliberately set display_hint: "json" — a valid value in the MCP enum that a "fallback override" prop or slot would have silently left able to dump. shapeOk and all five payload.X keys stay in ReportRenderer.vue: extracting them is the natural refactor and it empties the set test_1535 regexes out of that exact file, breaking the cross-surface drift guard. Report state moves into the store (contract #21), and every await is generation-guarded. A reset cannot cancel an in-flight promise, so adding the reportsLoaded flag contract #15 requires would have turned today's TRANSIENT wrong-render into a permanent one — the new agent marked loaded-with-the-old- agent's-data for the life of the mount. The component additionally gates every read on the state belonging to the agent on screen, since the store outlives it and a fresh mount fires no props watcher. AC #4 was not free: three reused sites carried bare gray-500 meta text with no dark override, which the contract calls the floor and never meta text. Fixed, plus the set's only raw non-gray (bg-blue-500 -> bg-status-info-500), so the per-file non-gray count only shrinks. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs: record the third renderer surface and the portal row window (#2162) architecture.md: the renderer set is a THREE-surface contract, which widens the blast radius of the test_1535 pin — so the note that shapeOk must stay in ReportRenderer.vue lives beside it. The portal row window goes in the existing "Large payloads" bullet as two query params, not a second route. workspace-agent-page.md: a new "The Reports tab" section, placed above the #2161 prose. It records that rendering is PRESENTATION — exclusion stays projection-in-the-service, so this does not contradict the doc's own thesis — that the payload remains agent-authored untrusted content of the same class as asks.context, the honest residual a key-value summary leaves, the read amplification the window buys, and why the state moved into the store. agent-reports.md: three panel consumers, not two. Enumerating them is load-bearing — the list said "two" while the third dumped raw JSON at clients. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(workspace): generation-guard the in-flight bookkeeping too (#2162) The finally block cleared its own reportId from _reportInFlight unconditionally, but resetAgentReports already empties that map — so a late finally from the previous agent's request deleted the marker a NEW request had just written, and the next expand issued a duplicate. The guard leaking through its own bookkeeping. Narrow (report ids are global uuids, so the same id under two agents is close to impossible) but the fix is two lines and the failure mode is exactly what the guard exists to prevent. Also pairs LoadFailed's title with a message that says what to DO rather than restating the title (contract #25). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(workspace): scope the summary fallback to the portal, operator keeps JSON (#2162) AC #2 asks for a client fallback "deliberately stricter than the operator side, because the audience is an external client" — that is a request for a SPLIT, not for a stricter default everywhere. Making ReportSummary the fallback on all three surfaces erased exactly the distinction the criterion exists to create, and put an operator-visible behaviour change on a P1 client-facing bug fix. ReportRenderer now takes a `fallbackComponent` prop defaulting to ReportJson, so both operator panels render byte-identically to before; the Workspace passes ReportSummary. A raw dump is a FEATURE when you are debugging an agent's own output and a defect when the reader is that agent's customer. The override catches an agent-chosen `display_hint: "json"` as well as a shape mismatch, which is what the earlier allow-raw boolean was reaching for: `json` is a valid value in the MCP enum, so replacing only the mismatch path would leave an agent able to request a dump in front of a client. ReportSummary therefore loses its raw-JSON disclosure entirely — it is portal-only now, so there is nothing legitimate to disclose behind it, and a source guard pins that it never imports ReportJson. The AC #4 dark-ink fixes to ReportTable/ReportKpiTiles/ReportTimeline stay: theme correctness, no behaviour change, and they ratcheted fleet raw_nongray 810 -> 809. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(review): pin the redaction ORDER, and correct two docs that describe the superseded fallback design (#2162) Two review findings, both in the "the artefact says something the code does not" class. **1. `reportSummary.spec.js` claimed to pin an ordering it did not pin.** `safeText` redacts before it truncates, and the reason is real: truncating first can slice a credential-shaped token so the surviving head no longer matches, leaving a recognisable fragment on screen under a value that looks handled. The test asserting it used a `ghp_` token — and `\bghp_[A-Za-z0-9]+` needs exactly ONE character after the prefix, so the truncated head still matches and is still replaced. Verified by breaking `safeText` to truncate-then-redact during review: the suite stayed green. Replaced with a case that discriminates. `\bAKIA[A-Z0-9]{4}` needs FOUR characters, so a cut mid-token leaves a head the pattern no longer recognises — here `AKIAIO`, two characters of a real AWS key id. Re-broken against the new assertion: it fails, and passes again on restore. The original `ghp_` case is kept (it is still a valid behavioural assertion) but renamed to say what it actually proves. No source change: the ordering was already correct, only the proof was hollow. **2. Two docs still described the `allow-raw` design that review superseded.** The shipped shape is a `fallbackComponent` prop defaulting to `ReportJson`, with only the client-facing Workspace overriding it. `agent-reports.md` said the fallback IS `ReportSummary` "with ReportJson retained behind its disclosure" — describing a `<details>Show raw JSON</details>` that does not exist, inverting which surface gets which fallback, and contradicting the correct explanation four paragraphs below it in the same file. The architecture bullet was titled "one fallback" over a body whose whole point is that the fallback is per-surface. Both now say what the code does. This matters beyond tidiness: the split IS AC #2 (a client fallback deliberately stricter than the operator side), so a doc asserting a shared fallback is an invitation to "unify" it and re-open the disclosure. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: trinity-ability <309458136+trinity-ability@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Memory Folding #21

Description

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions