Observed during WordsApp memory maintenance on 2026-09-13: seven relationcreate calls using kind=supersedes from memory 34 to memories 1, 2, 6, 14, 17, 18 and 19 returned 'Relation saved.' Subsequent relationlist calls…
The intended model is that any GitHub source can reserve a repository and let agents work on it without adding users — valuable for open source, and a prompt-injection surface. Project instructions are delivered by memo…
Relations 287 connect tasks and memories only; the page says "Documentation pages are not part of the relation graph yet". A page cannot point at the task that caused it or the decision it explains, so its history and c…
Observed 2026-09-26 in a Claude Code session outside of any project D:\tmp, no .agents/conderis.json. Steps 1. itemstorekind=memory, fields={...}, visibility=private without projectId - published as personal memory memo…
As title suggests, there should be a recent call, so agents know what happened recently when working with multiple agents.
Corrected scope All tasks remain project tasks. Specify a read-access matrix for private, organization and public visibility; no standalone user/organization task scopes. Work Define private-task ownership and access by…
Work Add visibility and any necessary ownership fields to the existing project task model. Preserve ProjectId, task numbers, canonical references, parents and existing identities. Validate allowed visibility values and…
Work Apply the approved access matrix to task lists, reads, creation, patch/update, assignment, claim and completion. Visibility changes require explicit supplied fields, suitable authority and expected version. Omitted…
Work Keep projectId + canonical number contracts. Add documented visibility inputs/outputs and optional list filters without granting access through a filter. Explain private defaults, subtask inheritance, organization…
Work Add a compact Private / Organization / Public selector to project task creation and authorized editing. Explain who can read the task. Show organization/public options only when valid for that project, with useful…
Work Test private/organization/public visibility using two users, multiple organizations, public/private projects, external members and agents with differing grants. Cover UI/API/MCP reads and writes, direct references,…
Problem Hosted memorysearch takes only 30 results, KnowledgeService.List takes 100 before projection, and neither exposes a cursor or hasMore. The 2026-09-13 freshness audit needed multiple topic searches to inspect 62…
Outcome An authenticated agent can determine what it is allowed to do in a project without attempting a mutation and reading the resulting error. Problem The server derives effective permissions from AgentKey GranularPe…
A single-operator service on shared hosting is a weak answer to "should I put my team's project knowledge here" — but concealing it is worse. Visitors will assume the least favourable version of anything left unsaid. Sa…
See 290, section 4. Compute a 0–100 score per project from structure and timestamps only, without an LLM, and always return the findings that name the items. Work - Start with the findings that need no new data: crowded…
See 290, section 1. Title up to 80 characters for people, summary up to 300 characters as a terse gist for recall, detail for everything else. Work - New limits in memorystore, the web editor and the data model; tool de…
See 290, section 2. Work - Weight on memories critical, normal default, low, settable through memorystore, the web editor and batch store; changing weight of shared or public memories follows the usual review rules. - R…
See 290, section 3. Work - Optional parent memory in the same project and scope. A memory with active children is a bag; there is no bag flag. - Rules enforced in the service: exactly one level, no critical children, no…
See 290, section 3 recall. Depends on the bags subtask. Work - Grouped recall lists top-level memories only. A bag line shows its summary, the number of visible children and their titles, shortened to fit the budget: 48…
See 290, section 5. Work - memorystore responses carry deterministic hints: an existing memory with an equal or strongly overlapping title Similar: 48 – update instead?, and a bag that has just become stale because this…
See 290, sections 4 and 5. Do last, once score, weight and bags exist. Work - System guide and tool descriptions: when to update instead of adding, when to merge true duplicates archive the rest with a supersedes link,…
A task is usually implemented over several commits, plus fixes and reviews. Record them so people in the web app and agents without a checkout can go from a task to its code: what changed, whether a half-done task has p…
Part of 216's evidence metadata. A memory or documentation page describes current behaviour; what matters later is which state of the code it was verified against, not every commit that ever touched it. Commit lists sta…
See 302. One contract for every list tool, described once in the system guide. Contract - Request: limit small default, documented maximum and cursor from the previous response. - Response: entries plus total, hasMore a…
See 302. itemget returns whole items today; a page can hold 100,000 characters. Contract - offset and take in characters not bytes, which split UTF-8 characters, take with a high default such as 20,000 so ordinary items…
Data model change from 309 Decision 6, split out of 315 so that the v2 tool surface 310 did not wait for a migration. v2 is live and maps version onto revisions internally; this task removes the revision tables. Start w…
Domain change from 309 Decision 1, split out of 315 so that the v2 tool surface 310 did not wait for it. Start with 309, section "Continue here". State today The itemstore description already lists visibility private, s…
Anfrage: Prüfen, ob "Channels" Sinn ergeben – ein Chat, in den Agents schreiben können und gerade verbundene andere Agents die Nachricht im selben Moment bekommen. Alternative: ein MCP watch-Befehl. Erste Einschätzung K…
Problem itemaction claim, complete, block and unblock require the item version, so every action needs a preceding itemget. In a coordinator workflow one agent reviewing and closing many subtasks done by other agents thi…
Problem Task content is capped at 4000 characters. Self-contained task briefs for autonomous agents goal, plan, edge cases, done-when, checklist with build commands and project rules regularly exceed that; a planner age…
Problem Agent-ready tasks repeat the same project rules and build/compile checklist. When a rule changes e.g. '.meta files are git-ignored' became 'tracked', 41 task texts are outdated at once; the only fix is overridin…
Problem In about half of the Codex exec runs of one session the conderis MCP was either not loaded or answered 401 memoryrecall on an old project-memory endpoint. The agents then guessed: tried other endpoints or browse…
Observed 2026-10-03 by Claude Code: during a window of roughly 15-20 minutes, itemstore failed 6 times with only 'An error occurred invoking itemstore' no code, detail or request ID. Five failures were creates with 400-…
Design deliverable Describe a session-end proposal with kind new/update/duplicate/conflict, target canonical reference, base revision, source evidence, rationale, summary/detail, scope and visibility. Default to a few d…
Design deliverable Create a small non-sensitive scenario set: resume a task, recall a decision, detect superseded guidance, avoid duplicates and answer a project-specific question. Compare current recall to proposed lex…
Deliverable A decision matrix for: hosted proprietary service + public integrations; open core + paid hosted/team features; source-available server with explicit terms. Compare adoption, trust, contribution potential, s…
Working hypothesis Free individual use can lower adoption friction; paid teams may value shared projects, permissions/review, managed operation and support. Treat feature boundaries and prices as testable proposals, not…
Deliverable A staged launch plan around a concrete demo: one agent records a decision and unfinished task, another harness resumes with evidence and selects ready work. Prepare proposed README structure, quickstart, arc…
Why this is separate 252 removed the lock upgrade that made concurrent creation deadlock. A retry policy was deliberately left out there, because it is a wider decision than that fix. What was measured Turning on Enable…
Decision 3 of 309: forcing a proposal works per call propose: true in itemstore, done in 315 and as a setting on the agent key in the web app. The key setting was split off from 315 on 2026-09-19 and runs after the swit…
Problem In a batch store, tags must be a comma-separated string; passing a JSON array fails the whole batch. Agents naturally send arrays. Proposal Accept both a string and an array of strings for tags single and batch,…
