Repository navigation
Release v0.8.19 - #324
Merged
Merged
Release v0.8.19#324
Conversation
generate_image hardcoded gemini-2.5-flash-image, so newer Gemini image models never reached an operator, and a local change was overwritten by the next /aios:update. NANO_BANANA_MODEL (env) or a per-call `model` argument now chooses it; the default is unchanged and the result names the model that ran. Closes #320. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015LaDkQ6Gpkcy5B6C5M4wFV
…316) A sandboxed Bash tool may write the vault and $TMPDIR only. Step 1's clone runs outside the sandbox (it needs the network), but four later steps wrote into /tmp: Step 2's file list, the .gitignore merge's temp files, and two CRLF probes. Each stopped the run on a sandbox denial or a permission prompt. - The file list is .aios-run/update-files; Step 1 deletes any old one so a failed Step 2 can never hand Step 3 a stale list, and Step 7 removes it. Step 2 creates .aios-run/ (a vault that never ran the floor has none). - Probes and the .gitignore merge use .aios-run/ too. - tests/update-sandbox-writes.test.sh: a lint over every block after Step 1, plus Step 2 run live against a read-only clone. All four checks fail on main. Closes #316 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
When ~/aios is a symlink (setup makes one for a vault cloned elsewhere), `find ~/aios` lists only the link: no results, exit 0, no error. - /aios:update Step 3.9's leftover-directory sweep uses `find -H`; it had silently checked nothing on every symlinked vault. - CLAUDE.md § Identity step 2: search `~/aios/`, never a bare `~/aios`. A live Haiku session wrote its own `find ~/aios`, found no agent and wore none (budget +62, noted). - tests/find-symlink-root.test.sh: a lint for finds rooted at a bare ~/aios, and the sweep run through a symlinked vault. Both checks fail on main. Closes #319 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
…rompting (#323) Since Claude Code 2.1.326 each spawn or resume stopped on a trust prompt for its fresh ~/.aios/run/handoff-<hex> folder, so an unattended spawn waited on a question nobody saw. Trusting the parent once ends it (verified live: a new handoff folder under a trusted parent opens without a prompt), and each session still sees only its own folder. - hooks/run-trust-check.py: read-only. Silent when the folder is trusted or absent; otherwise one paragraph and the command (`cd ~/.aios/run && claude`, Yes, /exit). Honours CLAUDE_CONFIG_DIR and AIOS_RUN_DIR; never writes ~/.claude.json. - /aios:update Step 6.8 runs it every update; SETUP §9 runs it after the wrappers. - tests/run-trust-check.test.sh: 8 cases, including that ~/.claude.json is never changed. Windows: the terminal wrapper puts its handoffs in %TEMP%, which is not worth trusting whole; this covers ~/.aios/run, which the App and Glass use on every platform. Closes #323 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
#317) suggest-agent only ran when a session opened it by itself, and live probes showed that never happened mid-session: a request squarely in an agent's field got a good answer and no agent, so the App's suggestion card never fired. - hooks/suggest-agent-hint.py (UserPromptSubmit): matches the prompt against the agents index (names, aliases, match keywords; generic words count only beside a specific one). On one clear winner the session is not already wearing (last `Wearing agent:` line in the transcript), it adds one line of context pointing at the skill, which still decides. Silent for short prompts, slash commands and ties; never blocks; exits 0 on every path. - install-settings.py wires it as Hook G on both platforms; SETUP §8/§10 and hooks/_index.md. - agents/_index.md: content-writer also matches "blog post", "article", "write post". - tests/suggest-agent-hint.test.sh: 14 checks against the real index, including small talk, a slash command, the hat already worn, bad input and a missing index. Closes #317 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
Surfaces listed agents/**/*.md, so a venture's reference documents (laws, glossaries, examples, templates) counted as agents, and the App's agent check and its Ask panel disagreed. In one live vault 74 of 130 such files were documents. - The rule: a .md under agents/ with `name:` and `description:` frontmatter, not _index.md or README.md, not under a reference/ folder, not marked `agent: false`. Documented in agents/_index.md. - hooks/list-agents.py applies it: a list, `--json` for the App and Glass, `--unclear` for the lint (exit 1 when any file is neither). - /aios:housekeeping Bucket 37 proposes moving documents under reference/, marking one, or completing an agent's frontmatter; synced company folders are fixed at their source. - CI holds the bundled tree to the rule (tests/list-agents.test.sh). Closes #314 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
#274) Building a team meant reading agents/_index.md, picking agents and tiers one by one and writing each spawn request by hand. The first slice, without the App: - skills/aios/team-composer: reads the sentence (and a page the operator gave), names the stage (team-archetypes), and writes a `## Team` section into the project note: a charter, one row per member with agent, tier, kind (standing or crew), what it owns (single writer), who it reports to, and a first task. Re-running replaces the section. - hooks/team-plan.py: `check` refuses an unknown agent or tier, zero or two chiefs of staff, a member reporting to nobody, a task too long to type; `launch --approved` writes one spawn-inbox request per member (name, task with a report-back, tier, group named after the project). No live surface: nothing written, spawn lines printed. - tests/team-plan.test.sh: 11 checks, including that nothing is written without --approved. Approval cards and a live team view in the App and Glass are follow-ups in their repos. Closes #274 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
…lease (#309) hooks/ had 47 loose files at its top level beside three domain folders, so nobody could tell what belonged together. Now: context/ vault/ pipeline/ agents/ media/ security/ install/ (+ claude-identity/, git/, custom/ unchanged) - Every moved file leaves a forwarding stub at its old path (python via runpy, bash with symlink resolution so ~/.local/bin/aios-commit still works, ps1 via $PSScriptRoot). The stubs go in the next release. - Every reference in commands, docs, skills, agents, tests and workflows uses the new paths; CHANGELOG history keeps the old ones. Paths a script computes from its own location (vault root, hooks/git, hooks/.venv, the pipeline log) are fixed by hand. - install-settings.py moves a hook registered at an old path to its folder, keeping the operator's interpreter, root and guard (~ move); it also recognises Hook G. - install-git-hooks.{sh,ps1} find hooks/ from install/ and link hooks/vault/aios-commit. - CI: tests/hooks-layout.test.sh (top level is folders + stubs only, each stub's target exists, no old path referenced outside stubs/history); the folder allowlist names the seven domains. A test fixture per suite follows the new layout; pinned pre-change refs keep the old one. CLAUDE.md budget +105 (longer paths only). Also: team-plan.py quotes the spawn lines it prints with shlex (a task holding $(...) or backticks stays literal when pasted) and refuses control characters; its liveness check uses claude-identity/_fs.pid_alive (signal 0 terminates a process on Windows). Closes #309 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
…nd the settings move From the release's security review (no high findings; two medium, five low): - team-plan.py: `check` prints every member's task verbatim and a plan hash; `launch` needs `--approved <hash>` and refuses a note edited after the yes. Printed spawn lines are quoted for PowerShell on Windows (curly quotes doubled), POSIX elsewhere. Inbox requests are written to an O_EXCL temp and renamed, never over a waiting request. A malformed surface file no longer crashes it. The skill shows tasks verbatim. - install-settings.py: the #309 move rewrites a path only when it resolves to this framework's hooks/ (an operator's own ~/.claude/hooks/provenance.py stays). - run-trust-check.py: a `projects` that is not an object reads as untrusted, not a crash. - Bash stubs clear CDPATH, use `cd -P --`, and stop after 40 symlink hops. Tests: team-plan (16), install-settings (48), hooks-layout (8); each new check fails on the code before it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
- hooks/pipeline-executor.py (stub) carries the `# /// script` block of its target: `uv run` builds the environment from the file it is given, so `/today` through the old path ran with none of its dependencies. hooks-layout checks every such stub. - install-settings.py: a quoted root with a space (C:/Users/John Smith/aios) is moved too; a path followed by ; | & ) < > is recognised. - team-plan.py: each task starts with the agent to wear and where the charter is, so a member wears its planned hat whatever its session name; resolve-tier runs through bash (Windows cannot execute it directly); a request that appears mid-launch is reported, never a traceback. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
- Python stubs are safe to import: imported, a stub loads the real module and takes on its names (bus_log.is_bus_prompt stays callable for provenance and an outward-action guard) instead of running its main() and exiting 0; run as a script, it puts the real file's folder first on sys.path. hooks-layout checks the import. - team-plan.py fingerprints each request in the bus log before writing it, so a member's first task reads as delivered by the inbox, never as typed by the operator. - Hook G stays silent outside the vault, as Hook E does (a client repo gets no hats). - team-composer and update Step 6.8: on Windows, a silent Store `python3` is never a pass. - CHANGELOG: an action item for permission allow-rules naming a moved hook's old path. - Two tests that could pass without checking now check (run-trust-check, sandbox writes). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
… changes Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
Copy detection shows a moved file as its changes. It must not let a file the audit leaves out (a proven vendored doc), or any non-code file, reach a script, hook, workflow or SKILL.md unread: such copies are shown in full, and the brief lists every copy. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7
…int rejects old hook paths) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
- hooks/pipeline-executor.py (stub) forwarded twice, so /today through the old path ran the whole pipeline twice; rewritten, and hooks-layout checks each python stub forwards once. - PowerShell stubs re-run their target with -File in the same edition, so named switches (-Decline, -Json) bind as parameters instead of arriving as strings. - team-plan.py: tiers are read from resolve-tier's RUNGS line, never by running bash (on Windows a bare `bash` can be WSL's launcher); printed spawn lines use -Tier on Windows; the note-path check is case-insensitive there. - list-agents.py reads the whole frontmatter, so a long description cannot hide an agent. - video-watch-guide.md installs into hooks/media/, beside ocr-image.swift. - ps_quote tested on all five PowerShell quote characters. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
… an article Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
- Hook G stays silent on a spawned worker's prompts until the worker wears its own hat (its name chooses it), and in headless sdk runs; after that it hints as before. - list-agents and Hook G read files with utf-8-sig (a BOM must not hide an agent) and walk each real folder once (a symlink cycle under agents/ ends). - Imported python stubs put the real module's folder on sys.path, as their docstring says. - release-audit: a copy into code from outside the framework's own folders is shown in full. - CHANGELOG: the allow-rule action covers any rule naming a moved path, not only wildcards. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
…t script is optional The two-audit gate (#267) asked every part of a release to PASS from non-deterministic reviewers, which never converged: each fix bought a new run with new findings. Reviewers read the diff, findings are fixed, a second round runs only after large fixes, and the maintainer tests. release-audit.py stays as an optional second opinion. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
- team-plan.py: the approved hash covers the charter too, `check` prints it, and the note is parsed once so the rows hashed are the rows sent; invisible characters (Cc, Cf, Zl, Zp) are refused; a request is linked into place, never replacing one; the note path sent to workers names the real root. - suggest-agent-hint.py: silent only on a spawned worker's first prompt (before it has replied and picked its hat), so named primary sessions and long sessions get hints; a worn hat is read from the session's own replies, never from tool results. - install-settings.py: a hook path is recognised inside nested quotes or after a stray apostrophe. - release-audit.py: renames get the same full-content rule as copies; paths are never C-quoted; a copy that cannot be shown in full stops the audit; a copy from outside the framework's own folders is always shown in full. - PowerShell stubs start PowerShell from $PSHOME. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
…s the tasks Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
…on, not ~/aios The self-locate rule in CI: ~/aios is only a symlink an operator may not have. team-plan sends workers the note's real path. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
…ommits Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
…elease - issue-scrub: **bold** and linked names in USER.md are loaded and redacted; home paths keep their folders only under ~/aios, ~/.aios and ~/.claude (Closes #325). - video-watch: the headless reader runs in its frames folder, may Read only there, and loads no user or local settings (an injected "read ~/.config/..." succeeded before, is refused now); a URL must be http(s) and follows `--`; local paths are absolute (Closes #326). - bus_log: inbox paths compare normalised, so Windows requests are fingerprinted (Closes #327). - openrouter: Gemini serves text only when chosen in the secrets file or by --provider; falling back to it needs AIOS_FALLBACK=gemini (Closes #328). tests/security-fixes-0819.test.sh; openrouter-fallback gains the two opt-in checks. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
- video-watch: the reader's allow rule is `Read(//abs/**)` (one leading slash means relative, so the scoping had come from cwd alone); a relative --workdir is made absolute, since the reader runs inside it. - issue-scrub: a private home path is redacted to the end of the line even with spaces in it (~/Google Drive/Clients/…, C:\Users\x\My Documents\…), `aios/..` no longer counts as the framework, and an escaped alias wikilink `[[acme-corp\|Acme Corp]]` loads both names. - openrouter: --check shows an environment-only Gemini key as unused for text and whether the fallback is on; the docstring and hooks index describe the new rule; the CHANGELOG step never handles the key (the operator puts it in gemini.env themselves). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
…lder names Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
Superseded runs kept the macOS runners busy: the bash 3.2 lane queued 11-12 minutes for a 2-minute job. Runs on main are never cancelled. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
…ommits Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
…s to moved hooks USER.md personalizations, custom/ files, routines and launch agents keep working this release through the forwarding files and would break silently when they go. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
Merged into the release branch so it runs the full CI with v0.8.19 (a fork PR skips the required checks). Closes #320. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ri8aqK7pwGNSj7zCjcvetU
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Release v0.8.19. Merge with a merge commit (never squash or rebase): the CHANGELOG entry cites this branch's first and last commits.
In this release
suggest-agentstill decides, so mid-sessionSuggested agent:lines appear (the App's card renders them).team-composerskill +hooks/agents/team-plan.py: one sentence becomes a## Teamplan in the project note (charter, standing/crew, owners, one chief of staff, a tier each), launched through the spawn inbox only with--approved.hooks/agents/list-agents.py,--jsonfor surfaces,--unclearfor housekeeping Bucket 37; CI holds the bundled tree to it).context/ vault/ pipeline/ agents/ media/ security/ install/. Forwarding stubs at every old path for one release;install-settings.pymoves registered hooks to their new paths (~ move); a layout lint keeps old paths from returning.~/.aios/runis trusted (hooks/install/run-trust-check.py), in SETUP §9 and/aios:updateStep 6.8. Confirmed live by two sessions, see aios-app#182./aios:updatescratch writes go to the vault's.aios-run/, never/tmp(the sandbox blocked them).find -Hfor a symlinked~/aios; CLAUDE.md step 2 says to search~/aios/.Closes #317
Closes #274
Closes #314
Closes #309
Closes #323
Closes #316
Closes #319
Closes #325
Closes #326
Closes #327
Closes #328
--), bus_log (Windows fingerprints), openrouter (Gemini text and fallback are opt-in).Checks
Every new test fails on main (or on the commit before its fix) and passes here; two subagent review rounds (correctness and security) plus a final one; every
validate.ymlstep was run locally with stdin closed (202 pass; the rest are the bash 3.2, Windows and CI-summary lanes, plus a gitignored local file).🤖 Generated with Claude Code
https://claude.ai/code/session_011iqyiHcMbsucZrrjMVyWQ7