Repository navigation
feat(context-cap): trim only on a cold cache, add a worker cap and tool-result pointers - #29388
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS No design concern blocks this: the cold-cache gate, the opt-in worker switch and the optional Notes
[DESIGN-REVIEWED] b961894 Description read[DESCRIPTION-READ] b1ef813e91564140cb9fff7eba83faca32eb3a3cd36b01aa2745ec4aaf9c865c That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. Both product-shape items, removing Rule A and adding the worker cap (off by default), are covered by base RFCs: Notes
What this change shipsInventory (7 items) — 6 justifiedIntent: keep a conductor's or worker's chat from being trimmed while its prompt cache is warm, and let a trimmed chat fetch its old tool output. ADDITION, with a fix to the trim note riding along.
[FIRST-PRINCIPLES-REVIEWED] b961894 Description read[DESCRIPTION-READ] 7e7dba5d178a180deb401a18ce30fb8c020c9be7ea303c11db87fb8978c20a44 That sha256 names the PR description this verdict read. Recompute it with |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
93f1f72 to
2f36a0d
Compare
2f36a0d to
8b0e214
Compare
…11-03) (#29401) Records the owner's 2026-11-03 ruling in both context-cap RFCs: a warm prompt cache is never trimmed, conductors included. The conductor cap loses Rule A (over 400K whatever the cache); Rule B (over 200K once the previous turn ended 300 s or more ago) is its only trim. The worker cap is accepted as amended: cold-only over 300K, its own switch session.worker_context_cap default off, and the conductor's seed with a tool-result pointer list. Both RFCs record the new risk (a warm chat can grow past 400K until its cache expires or KAS compacts at 80%) and the offline replay of 2026-11-01 to 2026-11-03. The conductor RFC moves to partial (phase 1 is on main), the worker RFC to accepted. The implementation is #29388. Co-authored-by: Joe Guo <zejiangg@amazon.com>
…flict) (#29627) * fix(desktop): open editor links from framed panes after a confirm (#28465) Co-authored-by: ejhong <ejhong@amazon.com> * fix(governance): judge cron turns by their session and fail closed on unattended turns (#27745) A cron turn now hands the PreToolUse gate its own session key, so a governance profile bound to that session's surface governs it the way it governs a chat turn. When the governance tier's own evaluation raises, a turn nobody is watching (cron, heartbeat, a monitor wake) now refuses the call instead of allowing it. An attended turn keeps "no governance opinion", so a transient fault never refuses an ordinary user action. Both paths log the error. Co-authored-by: w <w@k.local> * fix(deploy): show the newest profile check, lock Verify while it runs (#28845) * fix(imessage): keep inline code and URLs verbatim when flattening (#28829) * feat: starred mates show in the left dock, drag to reorder (#28812) A starred mate is now a row in a Mates group in the left dock, under the Main rows. The star is the only switch, so there is no second pin to keep in step with it. Rows use the rail's own NavItem with the face in the same 16px icon box as every other dock icon. The dock shows the first 3 starred mates and an "N more" row that opens the roster, because the dock's top block does not scroll. Dragging a row reorders the group; the order is a new crew-record field, star_order, written through PUT /api/agents/{name} next to starred, so it follows the person to every device. Unstarring clears it, so a re-star lands at the bottom. A failed write refetches the roster and shows a notice. The group sits outside the rail's fixed blocks and shares the leftover room with the app list: on a short window the apps shrink first, down to one row, then the group, down to one row, so neither the installed apps nor the footer leave the rail. The phone rail beside the chat and Mates drawers shows the same mates as tiles after the app tiles. While a docked mate's chat is open, its dock row is the lit one rather than the Mates page row. Closes #28789 Co-authored-by: CrysisDeu <crysisdeu@users.noreply.github.com> * feat(artifacts): open an image artifact in the zoomable image viewer (#28786) * fix(code-review-sage): pass counts into the catalog on glued count labels (#28700) * fix(apps): refuse manifest shell strings on native Windows instead of misreading the spawn failure (#28619) detectInstalled and openCommand run through /bin/sh -c, which native Windows does not have. The registry probe and the install precheck swallowed the FileNotFoundError as "not installed", and the open route answered "running remotely" before it reached the spawn. Each site now checks the platform first: the probe and precheck start no process and log the reason, and the open route answers 409 app_open_unsupported_platform. Fixes #27780 Fixes #27798 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> * docs(rfc): record the first-Mate welcome and the goal-saved Heartbeat offer (#28604) Adds two decision entries for Zezhen's owner decision on #25476, amends rfc-crewmate-guides-and-mate.md and rfc-crewmate-heartbeat.md to match, and turns the prompt-wording pins in the two #25476 tests into structural checks against crewmate_prompts.json. Fixes #25528 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> * fix(i18n): pass counts into the catalog on chat count labels (#28590) The comment-thread tally, the file-chip "Show N more" row, the queued count, the pending-approvals header and the subagent overflow row printed a number in front of a separately translated fragment. No catalog could place the number or pick a plural form. Each label is one key with a count, with plural forms per locale, and the fragment keys it replaces are dropped from every catalog. * ci: skip history blobs in main-ratchet-audit's checkout (#28459) Co-authored-by: sim <a@b.c> * fix(sidebar): explain what a sleeping session is on its expander (#26803) The "Show N sleeping sessions" row only carried a native title ("Not used in over 7d"), which shows after a delay, never on keyboard focus, and did not say what sleeping means. The row now opens an instant tooltip below it, also wired as its accessible description, saying the sessions are not closed and keep their history, were folded away only because they went unused past the threshold, and which menu row behind the filter button changes or turns this off. What is hidden, and when, is unchanged. * fix(i18n): translate the approval card sentence and code block names (#28833) The approval card appended a hard-coded English " wants to run" after the command, and the code block named its region "<lang> code" (or "code") and labelled an untyped block "code" in English, in every language. The approval sentence is one catalog phrase with the bold command as a <title/> component, so word order follows the language and the command stays verbatim. The code block label and region name are catalog keys with a {{lang}} placeholder. All twelve shipped catalogs carry them. * fix(i18n): give four count labels their singular forms (#28801) * docs(contributing): group tool-made changes into one pull request per directory (#28582) * docs(contributing): group tool-made changes into one pull request per directory * docs(contributing): keep a rename that must land atomically in one pull request A rename that crosses directories cannot be split by directory without breaking callers, so drop it from the group-per-directory examples and say it stays in one PR. --------- Co-authored-by: sim <a@b.c> * fix(acp): attach no image from a path written as literal text (#25534) build_prompt_blocks scanned the raw message, so a readable image path inside inline code, a fenced block, a 4-space-indented line or escaped \ markup was rewritten to [image: <name>] and attached. strip_image_refs masked only fenced and inline code, so the two sides disagreed on indented and escaped paths. mask_literal_spans in image_refs masks all four shapes (indented and escaped by the same predicate iter_local_refs uses), and both the builder and the history scrubber now scan _PATH_RE over that view. A fence that never closes is not masked on either side, so a dangling opener cannot swallow a channel's appended attachment lines. The builder masks each user segment of the prompt on its own (kiro_crew.prompt_segments, as build_message marks it), so injected text that ends inside an open fence (a cancelled turn's partial answer, a Slack thread root) cannot pair with the user's next fence line and hide the image between them. Injected segments are sent as written, as before; a plain string is one user segment, judged whole. Fixes #11713 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> * feat(dashboard): move MCP Management to Customize → Connections (#24933) MCP Management (stub MCP servers, share their backends, Sharing assessment) was one of the Settings → Developer internals pages, among logs, telemetry and storage, away from the MCP servers it governs. The share-refused notice on the MCP Servers list sent people to that separate page to approve a launch. Render McpManagement as a section below the servers list in Customize → Connections → MCP Servers (/capabilities?tab=mcp&view=mcp-servers); the component is reparented, not rewritten. Drop the mcp-pool page from Settings → Developer and redirect both old links, /settings/developer/mcp-pool and /developer?tab=mcp-pool, to the new section so bookmarks, docs and tips keep working. While the servers list holds unapplied edits, MCP Management's "Ask the agent" hand-offs are withheld so they cannot navigate away and drop them. Point MCP_MANAGEMENT_PATH, the doctor message, the curated tip, docs, the feature map and the GUI feature catalog at the new location, remove the now-unused tab label from every locale, and add tests for the redirects, the placement and the repeat-link scroll. Closes #24920 * fix(chat): route every queue drain through one derived queue-held gate (#22224) A queue held for one reason was drained anyway by a path never taught to consult the hold, burning the queued prompts into the failure they were held back from. * ci: blobless full-history checkout and 20-minute budget for the stable gate and internal content scan (#29301) Two jobs need every commit (the stable gate resolves the release tag and reads CHANGELOG at it; the internal content scan diffs base..head), but on the runner fleet a full fetch-depth: 0 clone takes 8.6-10+ minutes and their 10-minute job timeouts cancel them inside the checkout step. Add filter: blob:none to both checkouts (all commits and tags arrive, the working tree is still populated, older blobs are fetched only on demand) and raise both timeout-minutes to 20 as headroom. Neither job pairs a full clone with a tight budget any more, so both leave the bounded-checkout ratchet's allowlist; new tests pin the filter and the budget. * fix(memory): write no history line when the consolidation entry is empty (#29271) The history consolidation prompt always asked for a 2-5 sentence paragraph, so cron, subagent and automation-only sessions produced a dated "No substantive conversation occurred" entry every night. The prompt now lets the model answer an empty string when a span holds no decision, outcome or fact, and says a scheduled run that did real work still gets an entry. The V1 writer appends only a non-blank string entry, as the member store's apply_consolidation already did. The marker still advances. Closes #20448 * docs(rfc): let the warm welcome link the pull requests it checked live (#29286) Refs #29285 Co-authored-by: Joe Guo <zejiangg@amazon.com> * fix(update): install the release's pinned KAS engine on every update (#26386) Every update path installed only the new Python code. A release that moves KAS_PIN_VERSION then starts with no engine under the pinned name, so on the bundle route every KAS spawn is refused, and doctor says "nothing installed" while the old engine sits on disk. Add platform/release_apply.apply_release(), the one function every update path goes through (CLI git and managed venv, dashboard git and wheel, auto-update git and wheel). It installs the new code, then runs the NEW release's `kirocrew kas install --if-in-use` on the new interpreter, and switches over only when both succeeded: the wheel engine runs the step before promote, the git callers restart only on success. The new flag installs when agent.kas_bundle is on (a failure fails the update), and refreshes an engine already on disk when the route is off (a failure prints the retry and the update still switches), so a host that cannot fetch the engine keeps updating. Older engine versions stay on disk for rollback. Doctor now names the installed and the needed version when the pinned one is missing, and the hint lines sending the operator to run `kirocrew kas install` after an update are gone. Co-authored-by: CrysisDeu <crysisdeu@users.noreply.github.com> * docs(rfc): the follow-up card's Start sends its prompt (#26380) * fix(security): accept Salesforce client ids at operator-listed OAuth endpoints (#29262) A Salesforce MCP sign-in URL is refused even after its My Domain endpoint is added to oauth_endpoints.json. At an approved endpoint the gate skips the bare-secret scan for client_id only when it is exactly 43 base64url characters, and a Connected App consumer key is about 85 characters, so most real keys are read as a secret. At an endpoint listed in oauth_endpoints.json, and only there, client_id is now also accepted without the bare-secret scan when it has the Salesforce consumer-key shape: 3MVG9 followed by 70 to 100 of [A-Za-z0-9._]. The operator chose that endpoint, so the gate trusts it for this one shape. At builtin and MCP-discovered endpoints client_id keeps the full scan, as on main, and every other parameter at every endpoint is unchanged. Closes #28554 Co-authored-by: Matt Briones <mdbrio@amazon.com> * fix(mcp): keep the control plane unmarked when no admin registry can apply (#29017) Co-authored-by: w <w@k.local> * feat(acp): report kiro-cli's own model fallback (#29310) * feat(acp): report kiro-cli's own model fallback kiro-cli can move a turn onto another model inside one session/prompt, under both engines (KAS and V2). Crew ignored the notice, so the transcript named the model that refused, a pinned session drifted back onto it, and a V2 turn the engine had recovered was relabelled a refusal by Crew's own metadata fold. Decode both wire shapes into one notice (acp/engine_fallback.py), keep the session's model in step with the move, and end each turn with one line in kiro-cli's wording on the dashboard, cron, heartbeat and sub-agent results. A V2 refusal hop drops the refusal it moved off, so every consumer reads a true terminal. Surfaces that show the model (the sub-agent chip, channel /model pickers) follow the engine, and the model chips now cut long ids from the left as intended. Co-authored-by: cathzhao <265661517+cathzhao-test@users.noreply.github.com> * fix(acp): stamp an engine promotion as a runtime model write A promotion raised the slot's pick generation without marking it as the runtime's own write, so an Issue Radar crew read it as a person's pick and held its triage/work model switch. Stamp _model_runtime_pick_gen the way config_watch does for its runtime clear. Also let the business-counters source pin match the annotate call now that it passes stop_reason over several lines. --------- Co-authored-by: cathzhao <265661517+cathzhao-test@users.noreply.github.com> * fix(mcp-gateway): route crewmate files through the broker overlay pass (#28798) A crewmate's worker on KAS got no tools from a gateway-routed stdio server that needs env credentials: the overlay pass read ~/.kiro/agents only, so the worker had no stub and KAS declared the server itself without its env. The same pass now also rewrites each crewmate file in ~/.kiro/mates/agents into <overlay>/mates/<id>.json under the stub identity crewmate:<id>, with the same reader, launch approvals, sidecar ledger, fingerprint, cache and prune. A linked mates/ folder is refused, and a failed overlay write removes its sidecars. On KAS only, a worker session finds its stubs by its own execution record, in one lookup whose entry names KAS also subtracts. An app callback from a crewmate stub is governed under the kirocrew-mate-worker profile and its own identity, and runs only when both permit. Unverifiable listings, records, files, collisions or overlays yield no stub. Fixes #26261 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> * test(pi): bump the lane's pi to 1.4.0 and re-capture its corpus (#25718) pi 1.4.0 is the latest @earendil-works/pi-coding-agent release; the Real Adapter Contract lane pinned 1.3.0. pi-acp stays at 0.0.34, its latest. - Pin @earendil-works/pi-coding-agent 1.4.0 in test/real_adapters (lockfile regenerated with npm install --package-lock-only). - The lane's pi tests pass on 1.4.0 unchanged; only the G3 pin gate reds, on all ten pi live fixtures. Re-capture all ten on pi-acp 0.0.34 over pi 1.4.0 with the same drivers and reductions as the 1.3.0 corpus. The steering and saved-spec drivers were not kept from the last round, so they were rebuilt from their fixtures' notes; each reproduces the old frame sequence and facts (steer consumed and answered; saved spec confirmed, an edited one refused with capability_runtime_unverified, an unreferenced pooled stub not handed to pi). Snapshots regenerated. - test_pi_saved_spec reads the header version from the lane pin instead of a literal, so the next bump does not edit it. - Comments and READMEs that cite these captures now name pi 1.4.0, and the compaction numbers quote the new capture (13484 -> ... -> 49649, then 13597). Checked before raising the pin: pi's engines.node (>=22.19.0), its thinking levels (off..max), its built-in tool set and the agent loop's steering/finishTurn path are the same in 1.3.0 and 1.4.0. Crew's gate hooks every tool_call, so 1.4.0's change to which extension tools are active on resume does not open a path around it. Part of #22063 Co-authored-by: Joe Guo <zejiangg@amazon.com> * feat(crew): a crewmate's chat shows its sessions' approvals and questions, and badges its roster row (#29276) Co-authored-by: Joe Guo <zejiangg@amazon.com> * feat(dashboard): add a link row for the pull requests the warm welcome checked live (#29295) Refs #29285 Co-authored-by: Joe Guo <zejiangg@amazon.com> * fix(kas): say when a spec hook command is too long to run (#28923) Co-authored-by: zejiangg <zejiangg@amazon.com> * fix(agents): read linked global member templates (#28702) * fix(agents): read linked global member templates Read linked global templates by declared name for member context, execution prompts, session MCP servers, and side-turn read-only specs. Fence targets outside the agents directory against managed state, including resolved and opened paths. Preserve regular-template precedence and reject divergent linked declarations. * fix(agents): screen Windows link targets and keep linked read-only bases Before resolving an agent spec, _read_agent_spec now runs the hooks Windows link screen, so a link whose target names an untrusted UNC share is refused and audited without the SMB probe. The read-only spec reaper counts a base that only a linked entry declares, matching what publication reads, so it no longer deletes a live derived spec. The session MCP fallback imports member context inside the function. Continues #14851. Co-authored-by: Fabio Fernandes <7543296+fabiohof@users.noreply.github.com> --------- Co-authored-by: Fabio Fernandes <fabiohof@amazon.com> Co-authored-by: Fabio Fernandes <7543296+fabiohof@users.noreply.github.com> * test(acp): put the not-found spawn's workspace beside the data home (#29248) * fix(notifications): unack a cron note by the entry its ack added (#29173) * fix: route three docstring-only chokepoint bypasses through their helpers (#28529) Three sites did the job of a shared helper without it, against the rule the helper's own docstring states: - auto_research subquestion_queue.save_queue staged through a fixed `subquestion_queue.json.tmp` and renamed it; it now goes through `atomic_write` (unique temp name, Windows rename retry). Bytes on disk are unchanged. - `kirocrew knowledge stats` padded its SOURCE column with len()/ljust, so a CJK source name pushed the count columns right; it now pads by `messaging.tables.display_width`, and the column is never narrower than its header. - GET /api/knowledge/export built `attachment; filename=<ns>.knowledge` by hand, which sends a non-ASCII namespace as raw bytes; it now uses `session_export.content_disposition` (RFC 5987 filename*=UTF-8''). test/test_docstring_chokepoint_sites.py pins those three sites plus the three grapheme-safe cut sites to their helpers, each with a spy on the module's own binding and, where the helper changes an observable result, a behavioural assertion that fails on the hand-rolled form. Part of #10614 * fix(knowledge): decode a UTF-16 BOM .kiroignore and warn on NUL text (#28644) Windows PowerShell 5.1 Out-File and > write UTF-16 with a BOM; decoded as UTF-8 every rule was NUL-interleaved and matched nothing, so the scan ingested everything the file excluded. Sniff the UTF-16 LE/BE BOM, and log one WARNING when decoded text contains NUL (BOM-less UTF-16). Closes #27843 * feat(cli): add kirocrew app update (#29309) `kirocrew app update <name>` sends the same request as the Update button on an app's page (POST /api/apps/<name>/update) through the running gateway's owner-only socket. With no gateway answering it refuses and changes nothing: an update stops the app's backend, swaps its tree and re-registers its crons with the running scheduler, which a separate process cannot do. Split out of #25512 so the CLI verb can land ahead of the automatic-update loop. * ci: blobless full-history checkout for the security scope lanes (#29347) The Security Scope Review jobs and the fork content scan check out with fetch-depth: 0, and on the runner fleet that clone alone now outlasts their 15- and 30-minute budgets: on 2026-11-03 the candidates job and the publish job were cancelled inside actions/checkout after 30 and 15 minutes. Apply the pattern from #29301 to every full-history checkout in security-scope-review.yml, fork-security-scope-review.yml and fork-internal-content-scan.yml: add filter: blob:none (all commits and trees arrive, the checked-out tree is complete, older blobs are fetched on demand) and raise the 15-minute job budgets to 20. The bounded-checkout test now pins the filter and the budget on each job. * fix(dashboard): fade the bottom of the MCP and Skill browser lists when more rows sit below (#29357) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * refactor(members): move the Schedules exit-guard state into its owner hook (#29274) * fix(mcp): ask for a one-line reply before ending a monitor or spawn turn (#29270) monitor_start, monitor_watch, spawn_run and spawn_continue told the model to end the turn at once. A turn that stops right after a tool with no text breaks the turn-end contract (docs/decisions/2026-10-03-every-agent-turn-ends-with- text-or-nothing-to-do.md), so the empty-response recovery injected a synthetic continue turn after a successful arm or spawn. Each ack, the applied monitor outcome and the monitor_start description now tell the model to reply with one line saying what it started, then end the turn. The recovery ladder is unchanged. Closes #9324 * fix(apps): read the official catalog caches through read_bytes_with_retry (#28641) On Windows a cache read that overlaps the store's own atomic_write replace raised PermissionError, which _read_cache and cached_entries folded into 'no cache'. Read the four cache sites through read_bytes_with_retry, the helper the other apps/ JSON readers already use. Closes #27831 * fix(mcp): leave env-blocked servers out of the share confirm while sharing is off (#29278) * fix(overview): fade the bottom of the Prompts, Steering and Skills list panes (#29358) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * refactor(acp): unchain the session-token ternary in kas_agents (#29259) hoist_servers_for_session stamped the session token onto hoisted servers through a list comprehension whose result was a chained ternary. It is now an if/elif loop in the same order: a managed server goes through _tokened_managed_element, otherwise an app server goes through tokened_app_element, and every other element is kept as projected. The same helpers get the same arguments, and the list is still freshly built in the same order. The comment above it said only managed entries get the token; app entries do too, so it now names both. * fix(subagent): a finished answer a reap got to first still reaches the parent (#28670) When a reap begins before a run's answer finishes, the reap owns the ending and the report. The completion then said only the stop's error (a deadline reap) or labelled the whole answer a partial (a user stop), so the parent never received a result the run had finished. The completion now carries the answer, named as finished, while keeping the stop's outcome and error. * fix(dashboard): show command completion as unavailable on Windows (#28981) On a Windows gateway both completion tiers answer an empty list: the PATH filter refuses every directory and the path tier has no shell cwd. That reply is the same one the off setting gets, so the Command completion switch looked on while no popup ever appeared. Settings > Display > Terminal now reads the gateway platform with useGatewayPlatform(). On Windows the switch renders off and disabled, with a line saying completion is not available there and that the shell's own Tab completion still works. The stored value, the completion route's reply and every non-Windows gateway are unchanged; a non-owner, who sees the platform as unknown, keeps the plain switch. The new string is in every shipped catalog, in each locale's informal register. Fixes #27881 Co-authored-by: Matt Briones <mdbrio@amazon.com> * test(hook): give weak tests real outcome checks (#26614) Co-authored-by: zejiangg <zejiangg@amazon.com> * feat(security): let a person override a command deny rule for one call or the session (#27574) Co-authored-by: w <w@k.local> * feat: show the feature tip as a one-line pill that opens into a card (#29324) The tip above the composer filled the strip with accent purple, sat three dismiss controls side by side and spread its body across the full composer width, so it read as busy and pulled focus from the composer. It now starts as a neutral-glass pill (lightbulb, title, chevron, close). Opening it shows a small card: an accent tile, the title, the body capped at a readable width, then the action button, Learn more and Turn off tips under the text, with collapse and close at the top right. The settings gear is dropped; Turn off tips is the same opt-out and the toggle stays in Settings > Chat > Discovery. Refs #29323 Co-authored-by: CrysisDeu <crysisdeu@users.noreply.github.com> * feat: draw the left nav rail icons at 20px (#29338) Rail icons were 16px and read small, most of all on the collapsed rail where the icon is all a row shows. Every rail icon now draws at 20px. Row height stays 36px expanded and 32px collapsed: the collapsed row's vertical padding drops from 8px to 6px. The row's left inset is fixed so the icon does not move while the rail expands or collapses; it drops from 12px to 10px so the icon centre stays 29px into the 58px rail and the collapsed icon remains centred. Refs #29336 Co-authored-by: CrysisDeu <crysisdeu@users.noreply.github.com> * fix(crew): leave out an absolute glob's link that leaves home (#29319) A crewmate whose agent template declares an absolute file:// glob under a folder that is a link to another volume (a cloud desktop layout) fails every turn with "Essential source <home>/<folder>: linked document or directory", because the declared-glob walk refuses any link it meets whose target is outside the admitted root. kiro-cli follows the link and loads the same files natively. For an absolute declaration, the walk now leaves such a link out, unread and not descended into, and the snapshot names it in one core essential-context#outside-skipped:<template> note. The KAS document reader skips it the same way. Nothing outside the root is opened or listed. A target that is managed memory or member state still refuses, and so do a link the path screen refuses, a link that stays inside the root, any link a relative declaration meets, and the launch-record size check. * fix(mcp): apply the 30s initialize cold-start floor on every platform (#29391) Right after a gateway daemon restart or upgrade, every session re-requests its backends at once, and several cold-start together on a host busy with the restart. A handshake that normally takes about 3s then takes 13-15s, overruns the 10s POSIX deadline, and the backend is reaped while it is about to answer, so every turn fails with "KAS managed MCP initialization failed ... initialize did not complete within 10s". The 30s floor added for Windows cold starts now applies on all platforms. A configured mcp_gateway.initialize_timeout_secs above the floor still wins. The spawn gate's initialize watcher and slow-init fraction already read the backend's effective deadline, so they move with it. Fixes #26379 * fix(acp): keep an agent-started turn's frames through a request wait (#29269) * fix(acp): keep an agent-started turn's frames through a request wait On a claude client with an agent-started turn handed off but not yet read, a request wait, wait_for_compaction or _drain_post_compaction_metadata that ran in that window filed the turn's start session/update in _mcp_notifications, so stream_unsolicited() read the turn without its first chunk. The stdout claim the claude adapter binds now keeps those frames for the turn: while a hand-off is pending, the claimed reader holds every frame of a kind that starts a turn, requests from the agent included, and puts them back at the front of _buffer in arrival order when it returns. Requests share the list because a request wait puts back the requests it defers on its own, which would reorder them against the held chunks. Other frames keep that reader's handling. A frame that would not fit in _buffer beside the held ones is not held, and a kill or reset drops the hold. Nothing registers on_agent_turn yet, so users see no change, and a Kiro client never binds the claim, so its read path is unchanged. Fixes #22262 * fix(acp): hold an agent-started turn's closing usage_update with its frames --------- Co-authored-by: wenlin <34065934+rinbn@users.noreply.github.com> * feat: add the live browser harness to chat (#26950) Drive the agent's browser through Browser Harness and a CDP relay in the gateway, and show the page live in the transcript. The person can take over at any time, and quick asks plus a sign-in sheet let the gateway type a secret into the page without the agent seeing it; nothing typed is saved. Fields inside iframes bind to quick asks, phones can type and swipe-scroll during a takeover, decision-model help goes through one provider registry, and the playwright-cli stack is retired. * fix(pptx-maker): keep the old engine when a Windows finalize rollback cannot move the new tree (#28737) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * fix(pipeline-conductor): print banned n/a when fleet_probe has no /proc to scan (#28642) Closes #27785 * fix(config): read config.local.json in three raw config readers (#29196) config.local.json is an overlay whose keys win over config.json, and the loader and Settings report the merged value. Three readers parse config.json themselves and ignored the overlay: the knowledge LLM pool (agent and knowledge), the dashboard terminal settings, and the Slack voice fallback used when no loaded config is passed. A setting placed only in the overlay was shown but not used. Add read_merged_config_files, which merges the whole documents in order as the loader does before any section is read, and use it in all three; each caller passes its own path helpers so their test seams still reach the read. * test(dashboard): golden for a replayed chat turn and slot-list pass-work budgets (#29304) * fix(memory): refuse memory forget on native Windows up front (#28966) * fix(memory): refuse memory forget on native Windows up front On a Python with no socket.AF_UNIX (native Windows), `kirocrew memory forget` showed the key's value and asked for confirmation, then failed with "the gateway's dashboard socket is not reachable from here", which reads as a transient fault. The verb's transport needs the dashboard's Unix socket, so it can never work there. The owner-terminal check, which runs before any store opens, now refuses first when socket.AF_UNIX is missing (the same condition the transport fails on), naming native Windows and the dashboard Memory page. The forget help and its --help description name the limit. POSIX behaviour, including the stopped-gateway refusal, is unchanged. Fixes #27903 * test(memory): give Windows runners AF_UNIX in forget verb tests The forget verb now refuses first when socket.AF_UNIX is missing, so on native Windows runners every test that drives the verb past that check hit the platform refusal instead. A unix_socket_platform fixture adds socket.AF_UNIX for the owner-terminal and refused-before-open tests; the gateway call is faked or refused in each, so no socket is opened. The native-Windows test still removes it again. Part of #27903 --------- Co-authored-by: Matt Briones <mdbrio@amazon.com> * feat(linux): add AppImage desktop integration (#9931) Install the bundled freedesktop launcher and hicolor icons when a Linux AppImage starts, and repair only files whose private manifest provenance still matches. Run synchronous reconciliation in a worker thread so packaged-Python metadata checks never block Electron main. Give AppImage launchers desktop IDs distinct from deb and rpm, and scope spaced display names plus the seven-size nightly icon tree to AppImage packaging so package-installed behavior stays unchanged. Keep the hash, hard-link, POSIX metadata, xattr, ACL, durability, and concurrent-writer protections. Cover worker failure containment, main-thread responsiveness, package target naming, ownership, and recovery behavior. Fixes #9930 Co-authored-by: Josh Kiepert <jkiepert@amazon.com> * fix(desktop): do not respawn into a port netstat lists no owner for while a bind is refused (#29050) On Windows, wedge recovery and Force-stop & Retry no longer read an empty netstat answer as "the port is free". They check it with a real loopback bind first. A refused bind is retried every second for up to the Windows gateway start budget (WINDOWS_LOCAL_GATEWAY_WAIT_MS, 120 s) while the loading screen shows, because a gateway that served connections leaves its port refusing binds for a while after it exits. Recovery respawns as soon as the bind succeeds; only a port still refused at the end gets the existing restart-required dialog. A port netstat lists an owner for is handled as before. Closes #28268 * fix(dashboard-verify): autouse fixture in shared conftest drains verify slot (#29244) When asyncio.wait_for times out on asyncio.to_thread(verify_package_page), the OS thread keeps running and removes the browser profile in _drive's finally block before _SLOT.release(). The test session's KIROCREW_TMP_RESIDUE_STRICT check fires at session teardown and can see the profile dir if the thread has not yet reached its finally. Fix: add one autouse sync fixture _drain_dashboard_verify_slot in test/conftest.py. After yield, the fixture checks sys.modules for kiro_crew.dashboard_verify (free when never imported), and if _SLOT is locked waits with wait_until() until it unlocks. This covers every test in the repo that exercises the save/snapshot/verify paths, including test_dashboard_save_tool*.py, test_dashboard_snapshot.py, test_dashboard_verify.py, test_agent_panel_routes_dashboard.py, test_agent_panel_routes_page.py, test_dashboard_preview_keep.py, test_dashboard_gallery_approvals.py, test_session_dashboard.py and test_change_cards_settings.py. The ~34k tests that never import dashboard_verify pay no cost. No production code change. Co-authored-by: CrysisDeu <crysisdeu@users.noreply.github.com> * fix(onboarding): import each Hermes profile's cron jobs in that profile's timezone (#29409) The Hermes adapter read every profile's config.yaml but used the first timezone it found for all cron jobs. A profile's job with no zone of its own then ran on the root profile's clock, or borrowed a sibling profile's zone instead of being reported as needing one. Each Hermes profile is its own home with its own config. The scan now keeps each profile's parsed config beside it and fills a job's missing timezone only from its own profile's config. Closes #29408 * fix(mcp-gateway): withhold a crewmate stub once its entry is registry (#29419) The crewmate overlay is the last pass's. When a crewmate's server is retyped `type: registry` after that pass, a KAS worker that starts before the next pass still got the old local stub: worker_overlay_spec and withdrawn_worker_stubs checked only for a dropped or muted entry. The KAS custom-agent projection already withholds a registry-marked entry, so the session-level injection launched a local backend the projection refused. Both checks now also read mcp_entry_is_registry_governed on the governed merged entry, so such a stub is neither injected nor kept. * docs: add a user-visible behavior bucket to the host contract (#29280) Add bucket 10 to agent-host-contract.md: one row per behavior a person sees in Crew, a column per harness (ok / gap / n/a / unknown) and the test that pins each row. Seeded from the fixes first made on KAS, so each harness can be compared against the others without reading the plumbing. Part of #29275. * fix(instances): name the unreachable pane port when a remote crew pane times out (#29349) Co-authored-by: w <w@k.local> * fix(prepare-pr): let pre_push_gate.py run git status on Windows (#29421) os.devnull is the nul device on Windows, and git refuses a device as core.excludesFile, so every git status exited 128 before any gate ran. Override with the empty value instead: git skips an excludes path it cannot open on every OS, and -c still beats a hostile excludes file. Fixes #29344 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> * fix(web): give the workflow library load failure a retry and a hand-off (#29164) A failed workflow-definitions list read rendered a bare text-danger paragraph: no role=alert (assistive tech never announced it), no retry, and no agent hand-off. Render it through ErrorNotice (role=alert + the hand-off, safe because the list read holds no draft) with a Retry button that refetches the query. Adds the workflowLibrary.retry string and its pseudolocale entry. The retry test asserts the list re-renders after a successful refetch by the list ROW (a button), not a bare text match: once a definition loads the detail pane auto-selects it and renders its name again in a heading, so a plain findByText would match two nodes. Before/after screenshots committed under temp-screenshots/29162/. * test(linux): skip lease-held publish checks where the kernel refuses write leases (#29444) * test(linux): skip lease-held publish checks where a hard-linked inode cannot be leased Two real-filesystem tests for the AppImage desktop integration have failed on the CodeBuild Linux runner on every main run since the integration landed. Both stop at the first assertion that needs a published update, with the module result {"skipped":true,"reason":"launcher customized"}. Diagnostic runs on that runner showed no fingerprint mismatch: the manifest's recorded hash, dev, ino, mode, uid, gid, mtime and xattr digest all matched the live launcher. The packaged helper returned "busy" from publish, because F_SETLEASE F_WRLCK on the launcher inode failed with EAGAIN. A scan of /proc/*/fd found no process holding the inode, and the lease was still refused after 0.5 s and 2.5 s, on both the public name and its anchor hard link. A freshly written file with one name was granted the lease. The runner has kernel 4.14.355-287.756.amzn2 with an overlayfs root. The module treats a refused lease as "keep the old file", so the update is preserved and reported as customized, as designed. The shared probe now repeats the module's own sequence on a scratch file: Node writes and closes it, hard-links it to a second name, and a separate interpreter reopens each name O_RDWR and asks for the lease. When the host refuses it, the publish part of each test is skipped with the errno in the reason; the ENOTSUP fingerprint assertions before it still run. Where the lease is granted, both tests run unchanged. Replacing the helper's lease call with a refusal turns both tests red on such a host. The two publish assertions now print the integration result, so a failure names the reason it got. Fixes #29439 * test(linux): assert the busy deferral where the host refuses write leases The previous commit skipped the publish half of two real-filesystem tests on hosts that refuse F_SETLEASE F_WRLCK on a hard-linked inode, such as the CodeBuild Linux runner. A skip leaves those hosts checking nothing, so the tests now assert the module's fail-safe there instead. On a lease-refusing host each managed update must report the launcher as preserved ("launcher customized" with the "changed during reconciliation; preserving it" note), every lease-held publish run must answer "busy", the launcher and its anchor hard link must keep the same bytes, inode, link count, mode and mtime, and no displaced name may appear beside them. The ENOTSUP test still runs its fingerprint assertions first. The open-descriptor test keeps its descriptor-held deferral check on every host, then checks two more launches with the descriptor closed and confirms no recovery record is ever written. Where the lease is granted, both tests run their full publish, recovery and cleanup assertions as before. Verified locally by simulating the runner: the probe reports a refusal and the packaged helper refuses any lease on an inode with more than one link. All 91 linux-desktop-integration tests pass that way and on the unmodified host. Making the module overwrite the launcher, or replace its inode with identical bytes, after a busy publish (only once no descriptor holds it) turns both tests red at the snapshot comparison. Refs #29439 * fix(workflows): count only calls that did not fail as preserved results (#25972) When a workflow run fails or is cancelled, its completion message and workflow_status say how many agent results were preserved. RunHandle.snapshot counted every entry in agent_results, but a failed call also keeps an entry there (None, kept for the resume cache) beside its reason in agent_errors. A run whose 3 calls all failed was reported as "3 agent result(s) ... preserved" right above "3 agent call(s) failed", and all 3 "results" were null. The preserved count and partial_results now cover only the calls with no recorded failure. The stored agent_results, which reruns replay, and a finished run's per-call record are unchanged. Closes #25969 * docs(decisions): record that kas sign-ins stay stored when off (#28914) A new decision entry supersedes the 2026-11-02 KAS sign-in entry, which is immutable. Turning agent.kas_persist_mcp_sign_in off stops KAS from using the stored grants but does not erase them, because silently wiping logins would confuse users; erasing stays a manual step. The entry also corrects where a stored false comes from: the Settings page and kirocrew config set write only the key being set, and only kirocrew config edit on a home with no config.json writes every default. * test(agent): give weak runtime and session tests real outcome checks (#29263) ACP, session, KAS, subagent, stop, spec, cloud and AI spine tests that only called a mock now read a real result: a stored audit row, a session handle field, a conversation log, a lease or record on disk, or a return value. Each counted fix fails against a bug planted in src/. No source code changes. Refs #25195 Co-authored-by: zejiangg <zejiangg@amazon.com> * test(cron): give weak cron, dashboard and memory tests real outcome checks (#29438) Cron, dashboard, chat, history, lesson, knowledge, review, memory and notification tests that only called a mock now read a real result: a stored job or store row, a file the code wrote, a real lock, a real chat slot, or a return value. Each counted fix fails against a bug planted in src/. No source code changes. Refs #25195 Co-authored-by: zejiangg <zejiangg@amazon.com> * feat(remote-dispatch): retire a binding whose crew stays down, fence it, redispatch elsewhere (wave 8a) (#26956) Co-authored-by: Joe Guo <zejiangg@amazon.com> * docs(dashboard): shorten long comments and docstrings (#29022) Co-authored-by: zejiangg <zejiangg@amazon.com> * feat(dashboard): fade the Agent Templates list bottom while rows sit below (#29405) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * docs(rfc): context caps trim only on a cold cache (owner ruling 2026-11-03) (#29401) Records the owner's 2026-11-03 ruling in both context-cap RFCs: a warm prompt cache is never trimmed, conductors included. The conductor cap loses Rule A (over 400K whatever the cache); Rule B (over 200K once the previous turn ended 300 s or more ago) is its only trim. The worker cap is accepted as amended: cold-only over 300K, its own switch session.worker_context_cap default off, and the conductor's seed with a tool-result pointer list. Both RFCs record the new risk (a warm chat can grow past 400K until its cache expires or KAS compacts at 80%) and the offline replay of 2026-11-01 to 2026-11-03. The conductor RFC moves to partial (phase 1 is on main), the worker RFC to accepted. The implementation is #29388. Co-authored-by: Joe Guo <zejiangg@amazon.com> * docs(harness): name pi 1.4.0 in the specs and set-model in the steering fixture row (#29399) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * perf(models): check kiro sign-in only right before the list-models spawn (#29389) Co-authored-by: Joe Guo <zejiangg@amazon.com> * perf(artifacts): cache only changed paths for the session-docs scan (#29382) The session-docs list walks every session through read_file_change_messages. Its cache kept each file change whole (before/after bodies) and held 256 sessions, so on a larger store the walk evicted every entry before reuse and each call re-read every session file. Keep only the string path of each change and size the cache by _METADATA_CACHE_MAX so it covers the store. Output is unchanged. Closes #29379 Co-authored-by: Joe Guo <zejiangg@amazon.com> * fix(acp): defer Crew tools under kiro-cli 3.x's single-binary install (#29375) From 3.0.0 kiro-cli ships no launcher: the POSIX install is one kiro-cli, the chat binary itself, with no kiro-cli-chat beside it. The deferral gate read that shape as the old launcher that execs a PATH kiro-cli-chat, so every Crew MCP server stayed resident on every request. A sibling-less kiro-cli now defers when the pinned binary itself reports 3.0.0 or later. A 2.x or unknown version keeps the servers resident as before. Co-authored-by: Joe Guo <zejiangg@amazon.com> * fix(hooks): admit Unicode letters and digits in Windows hook paths (#29374) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * fix(cron): refuse a command cron on Windows when it is created (#29366) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * feat(dashboard): draw the warm welcome's link row as the chat's PR pills (#29364) Refs #29285 Co-authored-by: Joe Guo <zejiangg@amazon.com> * perf(crew-board): redact each distinct string once per board reply (#29333) Co-authored-by: Joe Guo <zejiangg@amazon.com> * fix(crew): say why a quiet task waits instead of a bare "stalled" (#29292) The Ultra Mode Board and the goal-board page called every quiet open task "stalled". Now each row says why: Stuck (red) only when nothing is left to move it (its chat closed, or no main chat reads its reports); Quiet (amber) when the ledger's stale flag holds on an open chat, so nothing is scheduled to wake it but it can be resumed; Waiting names the PR it reported, else its own words, else "Paused, no reason given". A blocked report on goal-board reads "is blocked", so "stuck" means one thing. All locales updated. Co-authored-by: Joe Guo <zejiangg@amazon.com> * feat(enterprise): KAS is the default agent harness and the agent_backend floor (#29284) * feat(enterprise): KAS is the default agent harness and the agent_backend floor RFC rfc-crew-enterprise-edition.md D6. In an enterprise build: - the agent_backend floor is KAS (governance_floor_backend), so no policy removes it; the oss floor stays the Kiro CLI harness; - the built-in baseline allows only kas, so the Kiro CLI harness needs a remote rule; - the gateway refuses to start when KAS is the harness and is proven unable to start on either route (enterprise_boot.require_kas_harness), naming the fix, instead of falling back to the Kiro CLI harness. Closes #28616 * docs(harness-parity): allow refusing a KAS session KAS cannot start (D6) In an enterprise build the gateway refuses a session whose resolved backend is KAS when its boot proved KAS cannot start on the route it would take (kiro-cli unable to run KAS, or no runnable engine bundle), wherever that backend came from: agent.acp_backend, agent.member_acp_backend, a crewmate pin or a spawn_run(backend=) override. Record it in the harness-parity H1/H3 rule and the H1/H3 spec rows. Anything that refuses or alters a session whose resolved backend is Kiro CLI is still flagged. Update governance.md from the pending note to the shipped behaviour. Part of #28616 --------- Co-authored-by: Joe Guo <zejiangg@amazon.com> * feat(crew): keep the crewmate conductor in step with the root conductor (#27976) The crewmate's conductor copied the root conductor's patrol paragraph into crewmate_prompts.json, so a later change to the root charter would reach Crew Mode chats and silently skip every crewmate. The rules both conductors need (dispatch steps 1, 3 and 4, bind before seed, the work-ledger patrol, the interval band and renewal, acceptance, act-by-status with the verdict, claimed PRs, close, and reading the goal-conductor skill first) are now written once, in spec_prompts.py _LEDGER_CONDUCTOR_RULES. _CONDUCTOR_SYSTEM_PROMPT is assembled from _CONDUCTOR_PROMPT_SECTIONS, byte-identical to before, and conductor.patrol names each rule as a placeholder that member.ledger_rules_block fills. test_mate_conductor_parity.py fails when a root charter section is neither a shared rule nor allowlisted with a reason, when a crewmate piece copies root charter text, when the effective root conductor has a server, tool, grant or mapped skill the effective crewmate conductor lacks and the allowlist does not name, or when an allowlist entry is stale. Effective means the installed spec put through the per-session projection (kas_agents.to_client_custom_agent), which adds a crewmate thread's session-level dashboard and panel servers and their grants. crew-mode.md states the rule. Also carries the two fixes this branch needs because it stacks the ledger mount on the ledger prompt: three mate conductor tests that assumed no mount, and the re-pinned mate conductor spec digests. Part of #27930 Co-authored-by: cixuuz <26782336+cixuuz@users.noreply.github.com> * fix(issue-radar): keep a crew's failure stand-down from being revived (#21919) * fix(issue-radar): keep a crew's failure stand-down from being revived An Issue Radar crew's autonudge loop stands down after five cycles in a row that fail (a spent usage limit, a backend error after retries) or never get a session. The crew watchdog then switched it back on during its next pass: update(active=True) is a revival, which zeroes the failure streak, so the loop failed five more turns, stood down again, and was revived again for as long as the gateway ran. The watchdog now leaves a stood-down loop alone. It pauses the crew with a reason the crew page shows, which also revokes its unattended grant through the existing re-read, and removes the spent loop so that Resume relaunches the crew with a fresh streak. Any other inactive loop is still switched back on. Ported onto the crew_runtime split (#21107): the change now lives in crew_runtime_supervision.py, and the crew_runtime facade's manifest names the new owner symbols and the two reason constants it serves. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit a74cb9bd2198ce938693b79d31f0000eec5e7166) * test(issue-radar): drive the stand-down tick through the timer seam The stood-down-loop tests patched asyncio.sleep to run the stand-down tick at once, a stdlib clock patch the flake ratchet rules out. They now re-arm the loop's own timer at delay 0 and await it, the seam the autonudge suite uses. Their teardown also left work that writes under the temporary store running when the directory was removed, which Windows refused with WinError 32. Teardown now awaits the detached persists, stops the service, and awaits the cancelled timers and reconciler first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 00af3c81c3d6a668389c93b73a4b2a3cba5fc284) --------- Co-authored-by: Leon <leon.lai.yc@gmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * fix(mcp): never run the registry reachability whoami on an event loop thread (#29431) On a thread running an event loop, registry_governance_known_unreachable now answers from the cached verdict for the current sign-in fingerprint, or doubt (keep the registry marker), and warms the cache on a daemon thread. It never runs kiro-cli whoami there and never waits for the probe's lock. Off the loop the probe is unchanged. Fixes #29321 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> * docs(rfc): record Phase 0 verdicts for session-keyed scratch (#29422) * fix(docker): move kiro-cli into /usr/local/bin when the installer picks ~/.local/bin (#29341) * ci: stop the GPT lane's blocked-push count at the threshold it needs (#29305) The push-budget step in codex-review.yml ran _review_pushes.py to completion on every GPT 6.1 Review run. That reads the PR timeline and then the check runs of every head the PR has ever had, so a long-lived PR paid dozens of REST calls per review from the repository's shared GITHUB_TOKEN quota (61 for one PR with 36 blocked pushes). The section only needs to know whether the count reached PUSH_BUDGET_MIN. _review_pushes.py's entry point now reads REVIEW_PUSHES_STOP_AT and passes it as stop_at, which blocked_push_count already supported, and the lane sets it to PUSH_BUDGET_MIN. The prompt says "at least N". An environment variable rather than a flag, because the lane stages the script from the PR's base: an older base ignores the variable and counts in full. pr_findings.py --rounds and sweep_check.py are unchanged and still count in full. Part of #28761 Co-authored-by: sim <a@b.c> * fix(dashboard): keep the Services search and staged MCP changes across a trip to a chat (#29273) * fix(dashboard): keep the Services search and staged MCP changes across a trip to a chat Opening a chat unmounts Customize, so the Services search text and the MCP Servers changes staged for Apply were dropped on the way back. Both now live in sessionStorage for the browser tab. Closes #25962 * fix(dashboard): clear only the submitted MCP draft on Apply and drop no-op staged entries An Apply that finished after Customize unmounted left the applied changes in sessionStorage, so they came back as staged. Apply now empties the stored copy when it still holds the draft it submitted, so a change staged on a later visit survives; Discard always empties it. Staged entries the saved config already matches, or whose server is gone, are dropped once the list loads. The hook moves to website/src/hooks. * fix(scratch): fall back when the scratch root override cannot be stat'ed (#28636) * fix(dashboard): retire an orphan unread watermark once the gateway confirms it (#28709) * fix(dashboard): an idle read retires an orphan shared unread watermark A shared unread watermark newer than every saved row can never be covered by any read. Builds before #15014 minted such watermarks: they named a turn's `done` row that a gateway restart then dropped, and the `mc-unread-shared` record they wrote outlives the upgrade. Every local read kept the record (markSlotRead ignored the surviving watermark), every relayed read re-lit the badge (remoteSlotRead adopts the survivor), and every reload restored it (restoreUnreadBadges), so the dot came back forever. When a slot is idle (`running: false`) its last_ts is the newest row that exists, so a read that cannot cover the persisted watermark proves the watermark names a row the gateway no longer serves: retire the shared record entry, and the badge dies with the row it references. A running (or unknown) slot keeps the survivor: the lag-window protection is untouched, because mid-turn a last_ts legitimately trails rows the window has not been pushed yet. Fixes #19688 * fix(dashboard): confirm an orphan unread watermark with the gateway before retiring it The first commit retired an uncovered shared watermark when this window's own slot row looked idle and live. A window can look live and still lag a sibling that just watermarked a real reply, so it could delete that sibling's badge. Now a read that leaves an uncovered watermark asks the gateway for a fresh slot list, issued after the watermark is read. A real reply is a saved row the gateway already holds, so its last_ts covers the watermark and the record stays. Only a watermark newer than every saved row of an idle session is removed, and only if the record still holds that same value. The slotsLive flag is no longer needed and is gone. Co-authored-by: akushonkamen <akushonkamen@163.com> --------- Co-authored-by: akushonkamen <akushonkamen@163.com> * docs: document Windows proxy workaround for missing Claude models (#24473) Added troubleshooting steps for missing Claude models on Windows when using a proxy, including environment variable configuration and a workaround. * fix(i18n): route usage, size and turn-cost labels through shared formatters (#28778) Several dashboard labels built numbers by hand: token counts as an English K/M ladder, the API time as "125.0s", the Overview Today line glued from fragments ("1 sessions"), file sizes as 1024-based "KB", and the chat turn-stats line with an English "credits" word and a hand-built "$" cost. They read wrong in English and cannot be translated. Each label goes through the shared helpers instead: fmtCompact for token counts, fmtElapsed for the API time, fmtBytes for sizes, and formatCost for dollar costs. The Today line is one catalog phrase with plural count keys, and the turn credits word is a catalog key, in all twelve shipped languages. * feat(dashboard): draw Explore as one panel with a header that owns its close (#28527) * fix(notes): guard note saves with a content hash as well as mtime (#22790) * fix(notes): guard note saves with a content hash as well as mtime The Notes save guard compared only modification times, so an external write that kept the mtime (os.utime, rsync -t, coarse-timestamp filesystems, two writes in one millisecond) passed the check and was silently overwritten by a stale buffer. GET /api/note now returns contentHash, the SHA-256 of the exact bytes the returned content came from, and PUT /api/note accepts it back as baseHash. When sent, the hash alone decides freshness, everywhere the mtime check ran, including before each contended-rename retry. Saves without baseHash keep the mtime-only behaviour; tokenless saves sample a hash too. Successful saves and 409 conflicts return contentHash, and the editor stores and sends it, including from the conflict banner. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(notes): report a note deleted mid hash check as the deleted conflict When a baseHash save's check saw the note on its stat but the note was deleted before the read, the save answered 404 no_such_note instead of the 409 ESTALE with mtime -1 that a deletion gets everywhere else. The editor recovered on the next save, but showed a generic error first. A read that finds nothing now re-checks whether the note exists and reports the deletion as the usual conflict. A hash mismatch also hands the bytes it compared to the 409 instead of reading the note again. The server facade's forwarded-names manifest now lists the new server_paths helpers, and the test patches read_note_snapshot through the facade, as the facade mirror guards require. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * fix(apps): read rc and v-prefixed versions when checking for updates (#29062) _version_newer kept its own split("-") parser, so an insider build (0.4.0rc3, PEP 440 with no separator) or a v-prefixed version failed to parse and was never "older": the update badge and notice never showed for them. It compares through apps.version.parse_version, the parser check_min_version already uses, and still treats a string that is not a whole version (1.0.0 <b>, latest) as never newer. * fix(whatsapp): count a mention or quote on a media caption as addressed (#28872) In a group in mention mode, _is_addressed read contextInfo only from extendedTextMessage, so an @-mention of the agent or a quote-reply to its message typed as the caption of a photo, video or file read as not addressed and the agent stayed silent. It now reads the contextInfo of the image, video and document submessages too, checking each, since a protobuf returns an empty default for an absent submessage. * fix(i18n): keep the provider, author and actor names inside their sentence (#29187) * fix(i18n): pass counts into the catalog on glued count labels (#28284) PR check group headings, the skipped-check tail, the knowledge graph and picker totals, the JSON and JSONL preview counts and row cap, the resolved-comments toggle and the logs match count glued a number in front of a translated fragment, or used English outright. No catalog could place the number or pick a plural form. Each label is one key with a count, with plural forms per locale, and the fragment keys it replaces are dropped from every catalog. * fix(context): headless task, planner and workflow turns are non-interactive and window-sized (#25084) * feat(remote-dispatch): Feature Previews switch for work_ledger.remote_dispatch; hide work-item routes while off (#27564) Co-authored-by: Joe Guo <zejiangg@amazon.com> * docs(slack): shorten long comments and docstrings (#29489) Co-authored-by: zejiangg <zejiangg@amazon.com> * fix(crew): start a Mate whose absolute resource resolves outside home (#28675) Co-authored-by: Joe Guo <zejiangg@amazon.com> * fix(apps): show a built-in app skipped at boot as unavailable (#25565) When boot skips a built-in app whose files cannot be read, record the skip on the application. GET /api/apps and GET /api/apps/{name} stamp unavailable with the redacted reason on that built-in record, and the Library tile and the detail page show it as unavailable instead of as a working app. Fixes #25322 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> * fix(crew): goal-board shows a task stuck when its main chat closed (#29339) (#29484) The Crew panel calls an open task "Stuck · the main chat closed" when its conductor's chat is gone (the ledger is orphaned). The goal-board page showed the same task as quiet or waiting, because `with_worker_facts` stamped only `worker` and `stale`. `page_safe_for` now takes the slot the page's folds are read from, and `with_worker_facts` stamps `orphaned` with `work_ledger.is_orphaned` over whether that slot is still open, the Crew panel's own read. The template's `view()` makes an orphaned open report (progress / blocked / question) stuck after needs-you and working, as the Crew panel does, and its pill reuses the panel's wording. Without a slot nothing is stamped, so the page keeps its old rule. Fixes #29339 Co-authored-by: Joe Guo <zejiangg@amazon.com> * docs(goal-conductor): how remote dispatch works under work_ledger.remote_dispatch (wave 9) (#29479) Co-authored-by: Joe Guo <zejiangg@amazon.com> * docs(goal-conductor): ask when 'discuss' is ambiguous; wrap up a debate when only small fixes remain (#29475) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * docs(rfc): a crewmate's harness follows Settings -> Agent Harness when unset (#29442) Co-authored-by: Joe Guo <zejiangg@amazon.com> * fix(kas): let an agent whose allowedTools names the shell run past denyByDefault (#29440) Co-authored-by: Zejiang Guo <zejiangg@amazon.com> * fix(security): retry a denylist read that meets a concurrent replace (#28017) On Windows a read of denied_commands.json during _write_denied_state's replace fails with a sharing violation, and the tolerant reader folded it into {}. The user-rule toggle and delete pre-checks then answered 404 unknown user rule for a rule that exists. Both denylist readers now read through read_bytes_with_retry. Fixes #27875 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> * fix(decisions): bound Plumb and Laya CPU use like the Strands launcher (#24679) * fix(decisions): bound Plumb and Laya CPU use like the Strands launcher The Plumb-4B and Laya launchers let torch use every core and queued every decision behind the one in progress, so a burst of decisions kept the machine at full load computing answers whose callers had already given up. Both launchers now cap torch at half the CPUs the process may use (at most 16, one inter-op thread), set before torch loads, and admit one decision at a time: a decision that cannot start within 1 s is answered 503, which the gate already treats as a provider error and falls back from. Other routes (health, attestation) are not gated. This matches the Strands Decider launcher's bounds from #17618. Fixes #24577 * test(decisions): drive the Plumb late-body test by a stepped clock The test moved past the body deadline with fixed sleeps, so its result depended on when the server thread started reading. It now advances the handler's clock past the deadline once the handler has set it, then sends the last byte. --------- Co-authored-by: Matt Briones <mdbrio@amazon.com> * fix(secrets): refuse authorize on Windows so no unusable grant is signed (#27994) * fix(secrets): refuse authorize on Windows so no unusable grant is signed On native Windows `kirocrew secrets authorize NAME URL` signed a custom-secret policy and the Settings Secrets panel advertised that command, but `call_api_with_secret` refuses every me…
8b0e214 to
6259820
Compare
|
Dispositions for the review findings on 8b0e214, carried to head 6259820 (rebased on main after #29401 merged). GPT 6.1 F2 (upheld): the 299.9 s case races the clock. Fixed. GPT 6.1 F1 (downgraded): lookups skip size-rotation archives. Not changed. The pointer list and the lookup read the same live chain ( Design note 1: watch compactions once Rule A is gone. Agreed, and recorded as a risk in Design note 2: the note names a list that a failed pointer read leaves out. Accepted as is. The read fails only if the transcript read raises, which the seed already logs; the model then gets "No tool call" for any guess and falls back to re-reading, as the note also says. Not worth another CI round on its own. No override is requested. |
6259820 to
2a528a2
Compare
2a528a2 to
683a144
Compare
|
Opus 5.5 blocking finding on 2a528a2 (
Context: on this host the ACP layer already cuts stored outputs at 8,000 (181,590 transcript tool rows checked, max 8,000), but storage itself allows |
|
self-added: no
|
|
self-added: no
|
c69facd to
afd4de7
Compare
…ol-result pointers A warm prompt cache is never trimmed, for conductors and workers. The conductor cap loses Rule A (over 400K whatever the cache); Rule B (over 200K once the previous turn ended 300 s or more ago) is its only trim. A dispatched worker (kirocrew-worker) gets the same cold-only check over 300K behind a new session.worker_context_cap setting, default off. A trimmed session's seed now lists each stored tool result as one line (tool_call_id | tool | args | size), and get_chat_session takes an optional tool_call_id that returns that one call's stored output. The trim note no longer says get_chat_session reads tool output back without it; it says to re-read files and fetch other output by tool_call_id. The decision is recorded in rfc-conductor-context-cap.md and rfc-worker-context-cap.md by #29401.
afd4de7 to
b961894
Compare
|
Backend Tests (3.12, 3) on afd4de7 was ours: test_transcript_derivation_seam.py flagged a plain read_messages_chained_full in context.py. Fixed in b961894: build_tool_pointers now reads through the guarded derive_messages_chained_full_with_keys (rotated archives included), and a private or busy transcript lists no pointers. New test: test_a_private_transcript_lists_no_pointers. |
|
@bolichen97 Review-ready at b961894 (one commit on main 5c747a0). All 98 checks are green and readiness is passed. GPT, Opus, Design and First Principles have no blocking findings on this head. Since your review: the tool_call_id regex is now a fixed-character-set custom_validator. Two new reviewer findings are fixed: the lookup and the pointer list read rotated archives, and an approved call is listed once. The pointer list reads through the guarded derivation seam, so a private transcript lists none. Each fix has a test. Could you re-review? Your CHANGES_REQUESTED is the last thing open. |
|
A maintainer review earlier on this PR signed off the design, so the PR now waits for a maintainer approval. The one requested change is fixed on this head. The The review found no blocking issues on b961894. What it checked:
Done well: the body backs the default change with an offline replay over real transcripts. The Reader lines also name the exact tests that pin the Rule A removal. Optional:
|
bolichen97
left a comment
There was a problem hiding this comment.
Approving b961894.
The one change I requested at 683a1447dd is done. src/kiro_crew/validation_parts/messaging_session_schemas.py:363 no longer carries a pattern=re.compile(...) for tool_call_id. The field keeps only max_len, and _validate_get_chat_session (:348-:353) checks each character against the fixed _TOOL_CALL_ID_CHARS frozenset at :345. An empty id is refused too. The regexes left in that file were all there before this PR, and the diff adds no other new regex.
What else changed since 683a1447dd, per git range-diff, and why it looks right:
src/kiro_crew/context.py:1314andsrc/kiro_crew/mcp_tools/sessions.py:529now read throughderive_messages_chained_full_with_keys. The pointer list and thetool_call_idfetch therefore read the same guarded rows, rotated archives included. Every listed id can be fetched, and a transcript that is private or busy lists nothing and is refused.test_a_rotated_tool_result_is_still_listed_and_fetchedandtest_a_private_transcript_lists_no_pointerspin both.src/kiro_crew/context_assembly/replay.py:528-:541keys pointer rows bytool_call_id, so an approved call's two rows are listed once, at the newer position.test_an_approved_call_is_listed_oncecovers it.- The
tool_call_idbranch atmcp_tools/sessions.py:447still runs after the key, incognito and workspace gates, and:393refuses it withcrew.
Ran test/test_conductor_context_cap.py, test/test_search_chat_history.py and test/test_validation_parts_facade.py on this head: 134 passed. All CI checks on this head are green.
Non-blocking: as the Captain notes, tool_call_id can read a stored tool output from any session the caller can already open with get_chat_session, not only its own trimmed chat. That matches what the message read already allows. A line in the description saying cross-session reads are intended would settle it.
AI-assisted, sent by Bolin Review bot
Problem / Motivation
Goal: A conductor's or worker's chat is trimmed only once its prompt cache has already expired. A warm cache is never trimmed, and a trimmed chat can still fetch any old tool result by its id.
The owner ruled on 2026-11-03 that a warm cache is never trimmed, for conductors and workers. Today the conductor cap still trims a warm chat over 400K (Rule A,
CONTEXT_CAP_HIGH_TOKENSinsrc/kiro_crew/dashboard/chat_turn/context_cap.py). Dispatched workers have no trim at all, so a worker chat grows until KAS compacts it at 80% of its window. The trim note also tells the model thatget_chat_sessionreads tool output back, but that tool returns only user, assistant and inject rows, so a trimmed chat has no tool path to its old tool output.Why it matters
76% of worker turn starts already land on an expired cache (the 15-minute conductor patrol is longer than the 5-minute cache). A trim at those turn starts costs no cache. A trim on a warm cache throws away a cache the next request would have used. The wrong note sends the model to a tool that cannot answer.
Not a goal
rfc-worker-context-cap.mdsection 5.3 (work brief, last report, edited-file list).What changed (motivation → approach → change)
Owner ruling → one cold-only check for both agents → three changes.
D1, a cold-cache gate (
context_cap.py):_context_cap_rule(tokens, idle_secs, line)returns"B"only whenidle_secs >= CONTEXT_CAP_CACHE_SECS(300 s, the existing constant) andtokens > line. Rule A andCONTEXT_CAP_HIGH_TOKENSare gone. A warm or unknown cache never trims, at any size._Session.last_used, the previous turn's end. That comes after the previous turn's last request started, so it never calls a warm cache cold.WAKE_RESET_AGENTS(conductors) usesession.context_capwith the 200K line.kirocrew-workeruses a newsession.worker_context_capwith a 300K line (CONTEXT_CAP_WORKER_TOKENS). Every hold and the damping guard apply unchanged.How each default behaves:
session.context_captrue(unchanged)session.worker_context_cap(new)falsetrue: a worker chat over 300K is trimmed at a turn start whose previous turn ended 300 s or more ago, never on a warm cache.D2, pointers:
tool_call_id(an approved call's two rows count once):tool_call_id | tool | args | output size. The header names the chat'ssession_key. Built bybuild_tool_pointersinsrc/kiro_crew/context.py(the chained transcript with its rotated archives, redacted likebuild_session_replay) andtool_pointer_textinsrc/kiro_crew/context_assembly/replay.py.get_chat_sessiontakes an optionaltool_call_id. It returns that transcript tool row'smeta.output(redacted when stored), cut toTOOL_CALL_OUTPUT_MAX_CHARS(8,000) with a marker for the rest, and redacted again. It runs the same key check, incognito refusal, workspace gate and SEL log, and reads through the derivation seam with rotated archives (derive_messages_chained_full_with_keys), so an id listed before a size rotation still resolves. It is refused together withcrew.mcp_tools/sessions.pyand the field inmcp_tools/catalog/sessions.py.validation_parts/messaging_session_schemas.pychecks the id with a fixed character set, served byvalidation.py.session.worker_context_capis inconfig/sections_session.pyandconfig/loader_builders.py.D3, the note:
CONTEXT_CAP_TRIM_NOTEnow says to re-read a file you need, and to fetch any other output withget_chat_sessionplus this chat'ssession_keyand the call'stool_call_id.Docs:
configuration.md, the module spec row andconfig-baseline.jsonare updated. The RFC amendments are in their own docs PR, #29401.Product-shape note: this changes what the conductor cap does by default (no Rule A). The decision is recorded in
rfc-conductor-context-cap.md(topartial) andrfc-worker-context-cap.md(toaccepted) by docs PR #29401, merged to main on 2026-11-04. This PR is rebased onto it.Backwards compatibility
Compatible:
get_chat_sessionwithouttool_call_idreturns the same output (testtest_a_call_without_tool_call_id_is_unchanged). The new argument is optional. The new setting defaults off.session.context_capkeeps its key and default.Reader: src/kiro_crew/dashboard/chat_turn/context_cap.py:_context_cap_trim -- entry: chat -- Rule A's removal leaves warm conductor sessions in place;
test_a_warm_conductor_over_400k_is_never_trimmed,test_a_cold_conductor_over_400k_is_trimmedReader: test/test_conductor_context_cap.py:_context_cap_rule -- entry: chat -- the only caller of
CONTEXT_CAP_HIGH_TOKENS/ Rule A besides the check; updated totest_only_a_cold_cache_over_the_line_trimsTests
Proof: PROVEN test/test_conductor_context_cap.py::test_only_a_cold_cache_over_the_line_trims[400001-0.0-None], test/test_conductor_context_cap.py::test_only_a_cold_cache_over_the_line_trims[900000-299.9-None], test/test_conductor_context_cap.py::test_only_a_cold_cache_over_the_line_trims[900000-None-None], test/test_conductor_context_cap.py::test_only_a_cold_cache_over_the_line_trims[900000-300.0-B], test/test_conductor_context_cap.py::test_a_warm_conductor_over_400k_is_never_trimmed[412000] (head 683a144)
test/test_conductor_context_cap.py:test_a_warm_conductor_over_400k_is_never_trimmed[412000, 900000],test_a_worker_trims_only_on_a_cold_cache_over_300k[warm-over, just-warm].test_a_cold_conductor_over_400k_is_trimmed,test_a_worker_trims_only_on_a_cold_cache_over_300k[cold-over].test_a_worker_trims_only_on_a_cold_cache_over_300k[cold-at-line, cold-under], and the rule tabletest_only_a_cold_cache_over_the_line_trims.test_the_worker_cap_is_its_own_switch,test_the_worker_cap_is_off_by_default_and_config_json_turns_it_on.test_the_trimmed_seed_lists_each_stored_tool_result,test_the_pointer_list_keeps_the_newest_100.test_the_trim_note_names_the_tools_that_really_return_tool_output._Sessionsto a cold cache, because the session-over-400K fixtures they use no longer trim warm._context_cap_idle_secs), so the 299.9 s case cannot cross 300 s on a slow runner. The clock itself is covered bytest_the_clock_is_the_previous_turns_release_not_its_acquire.test/test_search_chat_history.py:test_a_trim_pointer_round_trips_to_its_tool_output(pointer id →get_chat_session(tool_call_id)→ that output, redacted).test_an_unknown_tool_call_id_says_so.test_a_large_tool_output_is_cut_to_the_promised_length(a 20,004-char output comes back as 8,000 chars plus a marker).test_tool_call_id_refuses_an_incognito_session.test_a_call_without_tool_call_id_is_unchanged. This control passes with and without the fix, by design.With main's
src/and this PR's tests, 19 of the new or changed tests fail. Only the unchanged-call control passes on both sides.Manual verification
Offline replay of the gate over real KAS transcripts on one operator's host, 2026-11-01 05:10 to 2026-11-03 23:41 UTC (2.77 days). Sizes and timestamps only, priced with the P-cross-session-prefix-cache price model (0.108 + 0.372 cr/100K per request, + 1.141 cr/100K when cold). Each chat's KAS sessions are joined into one no-trim timeline, so earlier real trims do not hide the size a chat would reach.
Related Issues
Refs #22082, #23037
Checklist
Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license.