November 20: making persistent agents usable
This is a public engineering report, not a transcript. It covers November 20, 2026 in UTC through 10:08, including work continued in older development threads. Commit history below includes changes made on the evening of November 19 in Pacific time. Repository activity is not a claim that one person authored every change, that every change is deployed, or that tests prove product value.
Client releases and release readiness
Languid gained pinned, checksummed releases and explicit updates, followed by client 1.1.0. Managed clients can check for updates without installing them automatically. Developer checkouts are protected. The first deployment exposed a missing release manifest in the service image; the packaging fix added it and a regression test. The public release manifest was read back during this report and announces 1.1.0. This does not mean existing installations have updated. Unix installation was exercised in isolation; Windows runtime remains unverified.
The December 1 release-readiness board records six gates: Spaces migration, security, recovery, fresh-user installation, billing, and operations. Creating a checklist is not passing its gates. Review also identified billing reconciliation and old-client tool compatibility as areas needing further validation; this report does not publish exploit details or claim those issues resolved.
Images, tools, and computer skills
Both Languid chat and this website chat gained clipboard-image support. The website supports previews, removal, image-only messages, and history rendering. A separate cold-history hydration fix loads private images before returning agent context. A browser regression checks the actual optimistic-message-to-history replacement, rather than merely rendering an image fixture. The hydration change was included in a later backend rollout; it should not still be described as awaiting its first deployment.
Pi tool-image persistence carries image references through context reloads instead of losing them after the immediate tool response. The implementation was validated with five targeted regressions. Computer-scoped installed-skill discovery supplies a bounded index of installed Agent Skills and reads their instructions on demand. It passed seven integration and six filesystem tests. Discovery does not grant execution permission or install skills.
The website corrected collapsed tool-group counts: one visible call plus the number actually hidden. Eight browser tests were reported for that change. Live tool animation also landed, but a fresh report of invisible activity is still being investigated. The CLI hides administrative commands from general help; startup command layout and header hierarchy were refined separately.
Feedback and diagnostics
Private maintainer feedback triage adds stable references and an audit trail for resolving or reopening shared reports; its focused validation covered 38 tests. Triage notes are not replies to reporting agents. A subsequent CLI feedback-reply change now exists in public main, superseding the earlier statement that replies were wholly unsupported. Its deployment and end-to-end delivery were not independently verified for this report.
Opt-in provider stream diagnostics record content-free event information to help investigate Inkling responses. Twenty-three targeted tests passed. Better diagnostics are not a fix for the reported provider failure. Static-site publication support landed, but the hosting DNS issue remains unresolved in this report. The Android composer-position report was investigated without a reliable reproduction or a claimed fix. An Android installed-web-app footer report also remains investigation rather than a verified repair.
Design directions
The working design direction is a feedback-to-product loop: investigate, choose an outcome, execute, verify, ship, and observe. Background Machines should be able to do bounded work, not only reflect on it. Execution and evaluation should be separately specified for a work item. Engineering evaluation asks whether a change works; product evaluation asks whether it solves the problem and supports a sustainable product. Passing tests alone cannot answer the second question. These are design directions, not claims that the complete workflow already ships. The public product-development loop and agent authority model provide context.
Jazz and ATProto Spaces were considered as infrastructure, not replacements for the agent experience. Private synchronization is not the same property as local-first operation. PDS portability and agent-identity creation were also discussed: support should be demonstrated against external PDS implementations, rather than inferred from a successful hosted-account path. See ATProto and Spaces as application infrastructure.
Other OSS discussions covered discovering and sharing Agent Skills, including the vit project, while keeping discovery, vetting, and installation separate. Rex integration highlighted that terminal-session environment variables should not be copied into a permanent global configuration; a Pi color-spillover symptom remains unexplained. Rocksky and Multi-Scrobbler were considered as an integration path, not reported as configured or deployed. A Jazz-backed PDS was an architectural possibility, not an implemented replacement for signed ATProto repositories. Tool-schema consistency, prevention of repeated invalid calls, and accurate feedback-sharing language were also reviewed; the public commit inventory records the Discord routing/schema work.
Prism exploration kept the ring constraint but moved away from thin, one-line-per-frequency drawings. The material directions were a heavy pressure-deformed ring, viscous glass, and ceramic carved by negative space. A separate experiment made the ring perceptible through a surrounding flow around an empty center. These are design summaries, not claims of a publicly released source package. The photographic ring entry supplies the earlier visual context. The Languid snail remained the terminal identity; a small illustration skill explored the same quiet visual language.
What is live, and what remains open
The public client-release manifest confirms 1.1.0, and the private-image hydration change had reached a backend deployment. Later skills, feedback-triage, Pi-image, provider-diagnostics, and startup commits are public source changes; their presence on main is not deployment evidence. This report performs no Languid backend deployment. Its website publication is a separate operation.
Open work includes release-gate evidence, billing reconciliation, tool-version negotiation, provider diagnosis, site-host DNS, Android reproduction, and verified feedback-reply delivery. The source inventory below also records parallel repository changes so that work outside the directly discussed fixes is not silently omitted.
Public commit inventory
The following inventory uses commit timestamps in the UTC window above, capped at the report's source snapshot. It includes repository-wide changes, including parallel work; subjects describe the commit, not an independent deployment verdict.
Languid
- a2ad464 — Call agent memory blocks simply blocks in Pi header
- c644ad2 — Show selected agent identity and resources in Pi startup
- 9e27c8a — Chat: hold a quiet working line in the log while a run goes
- 416ab12 — Reply to agent feedback from the CLI; the agent hears back on its next turn
- fc139ab — Onboarding: allow PDS form-action on /signup and smooth first-agent flow
- 775dda1 — Refine Pi startup header hierarchy and hints
- e4af90e — AGENTS.md: always use a git worktree, never the main worktree
- a8e4bf9 — Server test: skills naming their computer aren't the session's computer section
- 7605b6e — Chat test catches up with the page; the copy button builds its icon without innerHTML
- 1e03cc0 — Show Pi startup slash commands on separate lines
- c69f9d0 — Agent pages with a trailing slash redirect to the page without one
- feec9a9 — Admin: control room at /admin/control
- 7189fb8 — Agent URLs use the agent's full handle: /agents/scout.languid.ai/...
- 05c97d7 — Persist private tool images across Pi context reloads
- fb9739b — Share usage with Languid: daily visits and browser errors, off until owners turn it on
- cea3241 — Admin: honest funnel, more steps, last seen, and a page per owner
- 6d60701 — Idle compaction counts from the last request's start, after 3 minutes; the system prompt puts stable parts first
- 7e534bd — Add private maintainer feedback triage with stable report references
- b5ba7ec — Discover computer-scoped installed Agent Skills on demand
- 74931b9 — Run history: an ai.languid.run record per run in the agent's private space, the Activity tab, and ai.languid.run.listRuns
- 383f170 — Docs: cloud sandbox computer and volume; link from Billing
- 76b6159 — Add opt-in content-free Inkling provider SSE diagnostics
- a62a41d — README header: meadow on both sides, open sky
- cb2f60f — pi-languid: the snail mark in the header
- e5dd728 — Track six December 1 release-readiness gates
- 15ea413 — Sidebar: flat and still. The agent's rows line up with Home, its threads alone scroll
- c4af4e0 — Include client release manifest in service image
- f10f8bb — Compactions go to the ops log, and languid admin summaries lists them
- 79b5b5d — Hide admin commands from general client help
- 576e633 — Tasks run in the background: task returns an id, the result wakes the thread
- 1f0ed60 — Release client 1.1.0 from pinned updater-capable commit
- 034e7c1 — Add pinned client releases and explicit managed updates
- 28631f1 — Billing: drop repeated sandbox heading and volume line
- 79d801c — Machines: agents set up, change and delete their own, after the owner approves
- 5baa898 — What's new: each entry links to its docs section or app page, no list dashes; docs gain The chat and Machines sections, and the queue and restart text match what ships
- ce2035a — CLI: add marshell alias and snail banner on usage
- 2cc190a — Inference XRPC: balanced and flex speeds on Sail Research (poll with getCompletion); asap stays on Together
- 2bae3a9 — Hydrate private images before returning agent context
- 00f6f25 — README header: widen the floor's break around the g
- d90f344 — Memory keeper: watermarks, change summary UI, block_read, consolidator template, and evals
- 113df47 — Inference XRPC: ai.languid.model.complete runs on Together at list price plus 5% (Sail removed for now)
- ae685eb — Chat: accept clipboard image items and test private image delivery
- b739d0c — Merge branch 'queue-messages'
- dc80bae — Chat: queue messages sent while a run goes; edit or remove them until the next step takes them
- f2c9c52 — XRPC: ai.languid.voice.transcribe for outside clients, and ai.languid.model.complete / getCompletion on Sail Research (asap, balanced, flex) at list price plus 5%
- 4f136f4 — Helpers wait in line for a run slot instead of failing
- 2a231ec — Sidebar: one agent at a time, with a switcher to pick another or make one; remembers the last agent opened
- 3949360 — Helpers: task runs the agent on a child thread until it finishes
- 9686ed2 — Fly host: serve agents' sites at sites.languid.ai (LANGUID_SITES_HOST)
- 8d2ae70 — Sites: agents publish static sites as wisp.place records, served at sites.<domain>
- 764a8e6 — README header: split the floor around the g's stem
- d8525dc — README: new header, the landing's line drawing on ink with a low-hanging g
- dda9045 — Chat: a restart resumes the chat runs it cut off. interruptRuns notes running chat runs in state.pickups and the next boot resumes them (resumeRun) after a short wait for computers; an open page resumes too. The leftover manual case is a quiet line with a text Continue, no pill
- b1bb244 — Inbox: messages from Languid. Send product updates to every owner or one from /admin/messages; they come from the ops account, live privately in the ops space (ai.languid.ops.message, ai.languid.ops.messageRead) and show in the Inbox until dismissed
- 957043a — What's new page: recent changes newest first, linked from the sidebar with a dot until you've seen the latest (seen state is a cookie in your browser)
- 766ffe2 — Chat: center the conversation, float the sidebar to its left
- 11ccf1e — test/xrpc: accept strict did:web host rejection from C3 fix
- f8483bb — Security hardening: C1, H1, M9, M11, L1-L6 in server core
- c21de72 — security: fix C3, H3, M1, M11, L1, L6, L7, L8 in core modules
- a9bf102 — Harden CLI, pi-languid, and macOS app security (H10, M13, M17, L7)
- e7f49c3 — Harden deploy, Modal, and install scripts (H5, M17, L7, L8)
- d59d7f8 — Bind Discord reply routes before runs and align tool schemas
- ede6ae2 — Permissions: Auto, a classifier approves reversible calls the owner asked for
- 22e1907 — Sidebar: drop the computer count from the agent's Settings row
cameron.stream
- b11174b — Show accurate +N more counts for collapsed tool groups
- 208dbd3 — Test image survives actual optimistic-to-history replacement
- ae49e39 — Make new tool calls visibly animate on /chat
- ec5639f — Support pasted images in agent chat
- 549fd49 — Searchable font picker
- 9554672 — Invert user bubble on /chat
- e9e7965 — Group tool calls into one bubble on /chat
- 39cf7d3 — Add a font picker to Settings
Return to NOW.