Repair gonol authority and canonicalize ai.sh - #101
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8f3d61631c
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| ORG_DISTRIBUTION.md | ||
| skills.json | ||
| tests/test_gonol_build_skill.py | ||
| tests/test_char_compress_authority.py |
There was a problem hiding this comment.
tests/test_char_compress_authority.py intentionally contains self.assertNotIn("EDCM owns text-domain gonol construction", text), so including that file in active_files guarantees the rejection loop finds the forbidden literal in the test itself. Consequently, the new Gonol authority gate step in .github/workflows/ci.yml always exits nonzero; bash tools/check_gonol_authority.sh currently reports this file as stale even though the assertion is guarding against that doctrine.
Useful? React with 👍 / 👎.
| ## Candidate boundary | ||
|
|
||
| An unresolved constructor is permission to construct a named, bounded candidate; it does not block declared experimentation. It blocks promotion beyond the evidence, not construction or testing. | ||
| An unresolved constructor or geometry operation is permission to construct a named, bounded candidate only where the declared evidence permits it. It blocks promotion beyond the evidence, not honest experimentation. |
There was a problem hiding this comment.
Replacing “does not block declared experimentation” here leaves the unchanged test_ucns_option_selection_skill.py::test_unresolved_authorizes_candidate_work_not_promotion assertion unsatisfied. Because .github/workflows/ci.yml runs full unit-test discovery before the other gates, this commit makes every CI run fail; retain the asserted contract wording or update the dependent regression test with the doctrinal change.
Useful? React with 👍 / 👎.
| BIN_DIR="$HOME/.local/bin" | ||
| fi | ||
|
|
||
| TARGET="$BIN_DIR/ai.sh" |
There was a problem hiding this comment.
ai executable
The installer always creates $BIN_DIR/ai.sh, while the root README, CLAUDE.md, and module declaration promise ~/.local/bin/ai and demonstrate commands such as ai status. On a clean installation, those documented commands remain unavailable unless an unrelated ai command already exists; install the symlink under ai or consistently document ai.sh.
AGENTS.md reference: AGENTS.md:L196-L197
Useful? React with 👍 / 👎.
| @@ -0,0 +1,303 @@ | |||
| # ratios: loc_comments=hmmm imports_exports=hmmm calls_definitions=hmmm | |||
There was a problem hiding this comment.
The executable launcher has no shebang even though the installer exposes it as a PATH command. Direct execution therefore either returns ENOEXEC or falls back to /bin/sh; with dash, it immediately fails on set -o pipefail—reproduced by invoking the installed symlink from env. Add a Bash shebang before the ratios seal, which the repository's ratios convention explicitly permits.
AGENTS.md reference: AGENTS.md:L196-L197
Useful? React with 👍 / 👎.
| if [[ "$existed" == 0 || "$dead" == 1 ]]; then | ||
| respawn "$name" "$command_line" |
There was a problem hiding this comment.
If an agent executable is absent on the first start, ensure_window_shell has already created a live login-shell pane before respawn returns 127. After the executable is installed, subsequent starts see existed=1 and pane_dead=0, so this condition never retries the agent and the named window remains an ordinary shell indefinitely; verify the current pane command or otherwise retain failed-launch state when deciding to respawn.
Useful? React with 👍 / 👎.
| command="cat >> \"\$HOME/$REMOTE_LOG_SUBDIR/$name.log\"" | ||
| remote "mkdir -p \"\$HOME/$REMOTE_LOG_SUBDIR\"; tmux pipe-pane -o -t $(q "$target") $(q "$command")" |
There was a problem hiding this comment.
On a VM with more than one local user, these persistent pane logs are created under the caller's normal umask, commonly yielding a traversable directory and world-readable log files. Coding-agent pane output can contain prompts, source, command output, or accidental credentials, so the launcher should set a restrictive umask and enforce private directory/file modes before attaching the pipe.
Useful? React with 👍 / 👎.
| # === MODULE_BUILD === | ||
| # id: skill_lib_ai_launcher | ||
| # module_name: ai | ||
| # module_kind: cli |
There was a problem hiding this comment.
The repository's meta-module-build/SKILL.md limits module_kind to its declared enum, which does not include cli; the other new blocks likewise use unsupported checker and installer values. Since these declarations are intended for machine parsing and coordination, introducing undeclared kinds makes consumers interpret or reject them inconsistently; map the modules to supported kinds or extend the authoritative schema first.
AGENTS.md reference: AGENTS.md:L120-L121
Useful? React with 👍 / 👎.
| --- | ||
| name: char-compress | ||
| description: Character-based context compression for agent handoff and skill writing, owned as a skill-lib procedure rather than current UCNS mathematics. Use this when compressing a long thread, document, repo audit, canon handoff, or agent working-memory state; when a context window is filling and operative facts must survive; when writing a SKILL.md that should be flesh-dense and bone-sparse; or when checking whether a compression deleted negation, order, quantifier, operator, named object, value, decision, or unresolved hmmm. Historical bone/flesh and text-stack terminology is local compression notation, not a UCNS theorem/status transfer, EDCM constructor, or edcmbone metric implementation. | ||
| description: Character-based context compression for agent handoff and skill writing, owned as a skill-lib procedure rather than current UCNS mathematics. Use this when compressing a long thread, document, repo audit, canon handoff, or agent working-memory state; when a context window is filling and operative facts must survive; when writing a SKILL.md that should be flesh-dense and bone-sparse; or when checking whether a compression deleted negation, order, quantifier, operator, named object, value, decision, or unresolved hmmm. Historical bone/flesh and text-stack terminology is local compression notation, not a UCNS theorem/status transfer, active language-construction authority, or edcmbone metric implementation. |
There was a problem hiding this comment.
Changing this canonical frontmatter without regenerating skills/char-compress/SKILL.md leaves the plugin-facing activation description on the previous text. The Codex plugin adapter drift step in .github/workflows/ci.yml runs python tools/build_codex_plugin_skills.py --check, which now exits 1 with drift: skills/char-compress/SKILL.md; regenerate and commit that adapter with the source change.
AGENTS.md reference: AGENTS.md:L68-L69
Useful? React with 👍 / 👎.
| | `a0p-instancing/` | procedural | — | Peer for a0-betatest (a0p): agents are per-user CRUD `AgentInstance` + `CharacterSheet`, each owning a trained native ZFAE weight bank (three 157-seed cores); no `sub_agent_spawn`/executor/`InstanceMerge` — only volatile `MemoryCore.spawn_sub/merge_sub`. Sequence: create→distill-train→readiness gate→mode inference→sentinel/pending-override→safetensors checkpoint. Canonical source is `a0-betatest`. | | ||
| | `plain-lens/` | procedural | — | Plain-language, multi-lens companion views of dense canonical text. Build easier on-ramps (domain/audience/role lens selectors, progressive disclosure) that never replace or talk down to the source, keep a static fallback under any dynamic layer, preserve operators/negations/quantifiers, and report an EDCM-style body-vs-footnote tension reading as an illustrative heuristic (not an edcmbone metric runtime). | | ||
| | `gonol-build/` | procedural | — | UCNS geometry / EDCM text construction discipline. Resolve current owning contracts and admissible scale options; preserve closure, atomic participation, occurrence identity, full-source receipts, and protocol-required replay. Refuses superseded omega/phi/psi, bone/flesh, and carrier-LCM language doctrine. | | ||
| | `gonol-build/` | procedural | — | UCNS gonol objects/constructors/geometry + Stack language-construction research discipline. Resolve the owning Stack workspace, preserve closure, atomic participation, occurrence identity, constitutive relations, provenance, and required replay; EDCM is measurement/evaluation only. | |
There was a problem hiding this comment.
This updated catalogue row says Stack owns language construction and EDCM is measurement-only, but the same agent guidance still says at CLAUDE.md:255 to resolve “EDCM admissible scale options” before gonol work. Agents following the operational rules therefore receive the superseded authority assignment despite this commit's stated repair; update that instruction to resolve the owning Stack workspace and current UCNS authority as well.
AGENTS.md reference: AGENTS.md:L148-L154
Useful? React with 👍 / 👎.
| if [[ "$agent" != shell && ( "$existed" == 0 || "$dead" == 1 ) ]]; then | ||
| respawn "$name" "$command_line" |
There was a problem hiding this comment.
Because remain-on-exit is enabled, an exited shell remains as a dead pane; this condition explicitly excludes shell, so ai shell selects and attaches to that dead pane instead of restoring a usable login shell. Handle a dead shell through the shell respawn path before attaching.
Useful? React with 👍 / 👎.
1 participant
Repairs two live skill-lib defects.
tools/ai.shlauncher for thea0VM (0:shell,1:grok,2:codex,3:deepcode), including actual pane health, explicit restart, remain-on-exit, persistent remote logs, and presence-only propagation of already-existing VM provider keys into tmux.tools/install_ai.shto symlink canonicalai.shinto caller PATH, preferring Termux$PREFIX/binand otherwise~/.local/bin.Historical EDCM constructor identities remain replay history only; no new language construction or ESS/ZFAE work is claimed here.
hmmm: third-party CLI executable names/authentication may change; command overrides remain explicit.