Install agents and assets from the container driver - #53
Open
CrypticSwarm wants to merge 3 commits into
Open
Conversation
The driver now runs unified agent translation right after the config
phase, dispatching through the registered spec: the destination comes
from the spec's agents_dest ({home}/{config} placeholders resolved
against the run), the four asset sources are the same for every
harness, and grok's Waiver skips the phase on record instead of
falling through a case arm. Codex's finalize hook gains the home the
driver hands it and merges the emitted registrations into the
published ~/.codex/config.toml, replacing the separate registration
step; a bare CLI run of the translator still has no home and leaves
the published config alone.
The replaced entrypoint shell (prepare_unified_agents,
register_codex_agents, CODEX_CONFIG_FILE) is deleted, and the text
assertions over it become behavior tests: tests/test_harness_agents.py
stages source trees and reads what landed -- per-harness destinations,
source precedence, the opencode fallback destination, the grok
waiver, both degrade-to-warning paths, and the registration meeting
the published config that publish_config wrote.
The driver's third root phase installs skills and commands into each harness's native locations, one install_assets hook call per source layer -- dotagents user, org, shared assets, then the workspace overlay -- with the destinations resolved from the spec's skills_dest/commands_dest. The default hook copies with per-entry replacement (top-level dot entries skipped, symlinks recreated, wholesale replacement of whatever stands at the destination name); codex overrides it to translate the layer's portable commands into skill packages before the skills copy, so a layer's own skill package outranks its same-named command while later layers still replace earlier ones. A failed codex translation degrades to a warning; a failed copy stops the container. The replaced entrypoint shell (copy_dir_entries, translate_codex_commands, copy_shared_assets) is deleted, and the text assertions over it become behavior tests: tests/test_harness_assets.py stages layer trees and reads what landed -- layer precedence, wholesale replacement, entry selection, absent layers, per-harness destinations, the codex translate-then-copy ordering, and recorded whole-tree installs for claude and codex.
The phase-order tests mock every driver phase they are not asserting, so they no longer run the real asset install against claude's pinned live config destination from inside the suite. New tests pin the guarantees the code already made without a witness: a failed copy propagates out of the phase (and is not taken for a translation failure on the codex path), a destination symlink to a live directory is severed rather than followed, and a registration the published codex config already carries outranks the generated one. The merge_toml import moves inside the registration's try, so a python without tomllib fails the registration with its own warning instead of the translation's. Docstrings that named the entrypoint as the actor -- the command-translation shim, the Context and HarnessSpec contracts -- now name the driver.
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
1 participant
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.
Agent translation and the skills/commands installs were the entrypoint's remaining big
caseblocks. The container driver now owns both: the translator runs with the harness's emitter into its declared agents destination, thenfinalize_agents— codex registers the emitted agents in~/.codex/config.toml— and asset layers install in the canonical order (dotagents user, org, shared, workspace) with per-entry replacement, through aninstall_assetshook. The default hook copies skills then commands; codex's override translates commands to skills before the skills copy, per layer, as the shell did. Grok's waived agents destination skips translation exactly as before.The replaced shell is deleted with its text assertions ported to behavior tests in the same commits. Output trees — paths, modes, symlink targets, content — are byte-identical to the pre-change shell across full, sparse, empty, and collision layer stagings for all four harnesses.