Agent Onboarding
Installed Agent Manual
An embedded reference, discoverability paths, and project init command for coding agents.
3 milestones 8 tasks 19 agents
4 plan 8 execute 11 review
145k peak context
102k median execute
1h 52m total 6m 12s per task
44m 16s execute 34m 34s review
66 edits 650 commands
codex harness
This run put Gantry's agent orientation inside the installed binary. It built the manual corpus, exposed it through the command surface, tied dynamic sections to product data, added tests against documentation drift, and then added a project-local init command that writes the same pointer where future agents will read first.
How this walkthrough is structured
Feature
What did the run build and ship?
The current tree still carries the embedded manual renderer, command dispatch, reachability pointers, drift checks, and init behavior that the run introduced.
Build
How did Gantry structure the work?
The work was cut as a manual foundation, a discoverability and honesty pass, and a small closing init command, with each boundary staying green in the recorded run.
Milestone 1
Embedded Manual
4 tasks 11 agents 50m 46s wall time 145k peak context
The manual foundation is still visible in the tree as embedded markdown under the manual directory and as rendering code in the manual engine module. The command surface still routes the manual through the agents verb, and integration tests still exercise topic rendering and edge cases. Later commits have kept the same feature family rather than replacing it with an external documentation lookup.
This milestone carried the core manual experience before any broader discoverability work depended on it. The planner split it into authored content, command dispatch, live pattern data, and truth gates, which gave each worker a small contract to consume from the previous row. The run facts show the milestone-level gate and review stayed green, so the top-level cut did not need a recorded recovery before it was merged upward.
Manual Corpus
2 agents 15m 57s wall time 145k peak context
The manual pages remain under the embedded manual directory, and the renderer still includes them directly in the binary. The topic list in the manual engine keeps the required slugs, descriptions, and source paths in a stable order. The patterns page still carries a slot that later rendering replaces rather than a copied roster.
This sprint was asked to author the installed-user manual corpus and define the topic catalog contract that later work would route and test. It also reserved a roster slot for live pattern data without trying to solve discovery in the same worker. Its gate and review were green, which fit the boundary: content shape and embedding could be proven before command behavior arrived.
Agents Command
2 agents 4m 49s wall time 75k peak context
The current command layer still exposes agents as the manual verb and leaves the older agent spelling on the assistant diagnostic path. Topic resolution still normalizes queries, prefers exact matches, accepts unique prefixes, and renders index pages for failed lookups. The integration tests still cover the top page, exact topics, prefixes, unknown topics, ambiguous topics, and the diagnostic distinction.
This sprint took the catalog as input and made it reachable through the public agents command. Its scope was command dispatch and topic resolution: exact names, unique prefixes, unknown requests, and ambiguous requests. The recorded run shows no recovery, so the handoff from content contract to CLI behavior held cleanly.
Live Patterns
2 agents 12m 13s wall time 136k peak context
The patterns topic still renders through the same registry used by pattern discovery. Entries include their source tier, included-only marker where relevant, and when-to-use text loaded from the winning definition. The tests still prove bundled, project, and global pattern data affect the rendered page instead of relying on stale prose.
This sprint was scoped to one dynamic section inside an otherwise normal manual topic. It consumed the roster slot from the corpus and the command behavior from the prior sprint, then connected the page to pattern discovery at render time. The clean gate meant the dynamic data path could be added without destabilizing the manual command.
Truth Gates
2 agents 8m 18s wall time 102k peak context
The manual engine still has tests that reject source-tree-only references, check the topic catalog, compare the commands topic with the CLI specification, hide feature-gated text when support is absent, and verify the live pattern roster. Those checks live beside the renderer, which keeps the documentation contract close to the code that ships it.
This sprint closed the first milestone by converting manual accuracy rules into executable checks. Its brief made it consume the catalog, topic resolution, and live pattern behavior, so the worker did not need to reopen the whole feature from scratch. The run stayed green through gate and review, showing the contracts were testable at the point they were introduced.
Milestone 2
Truth Gates
3 tasks 9 agents 31m 20s wall time 132k peak context
The installed manual is now referenced from the visible entry points a cold reader is likely to hit. The help text, readiness report, run-entry guide, assistant prompt, command specification, and manual tests still carry the discoverability and truth-checking surface added here. The feature is not only prose; it is backed by assertions in the current suite.
This milestone moved from having a manual to making it easy to find and hard to let drift. Its child sprints separated reachability surfaces, command honesty, and page freshness, so no worker had to edit every pointer and every test in one pass. The recorded boundary stayed green and needed no recovery, leaving the earlier manual as a stable dependency.
Manual Reachability
2 agents 10m 18s wall time 132k peak context
The README now names the agents command as the on-machine manual, the CLI help includes the same pointer, and the doctor report closes by sending users there. The assistant prompt and command documentation also point usage questions back to the installed manual. Tests still protect the distinction between the agents manual and assistant diagnostic surfaces.
This sprint was asked to put the agents manual pointer on every installed-user path named by the plan. It also kept the assistant diagnostic naming cleanup in the same pass because both changes touch how a reader understands the command surface. The gate remained green, so broad reachability edits did not require a replanned cut.
Command Truth
2 agents 6m 12s wall time 71k peak context
The current manual tests still render the commands topic from the same usage specification exposed by help. They check available verbs, reject unavailable command spellings, and scan rendered pages across feature combinations. The stripped-build integration tests still ensure development-only and screencast-only surfaces do not leak into manual output when the binary lacks those features.
This sprint narrowed the problem to mechanical agreement between manual claims and the binary's available commands. That boundary let the worker focus on the CLI specification and feature configurations rather than rewriting manual prose. It passed without a recorded recovery, so the command-surface assertions fit the existing renderer and usage model.
Page Freshness
2 agents 5m 17s wall time 92k peak context
The manual source scan still rejects references that only make sense in the repository rather than an installed binary. The patterns tests still assert that the roster slot remains in the authored source and disappears from rendered output after live discovery fills it. They also verify that built-in entries and their descriptions remain represented in the rendered patterns page.
This sprint finished the remaining manual freshness checks after command honesty had its own row. Its scope was source-only references and the patterns topic's live roster behavior, both of which are documentation drift risks rather than command dispatch risks. The clean run facts show those checks landed without forcing the planner to redraw the milestone.
Milestone 3
Agent Init
1 tasks 5 agents 18m 50s wall time 138k peak context
The current tree still documents and dispatches the agents init subcommand. It writes a marked Gantry section into a project instruction file, preserving unrelated local content and avoiding duplicate generated blocks on later runs. The tests still cover creation, update, decline, noninteractive acceptance, idempotence, and refusal outside a repository.
This final milestone was intentionally optional and narrower than the manual and discoverability work before it. It depended on those earlier rows being complete, then added a convenience command that writes the orientation pointer into a user's own project. The recorded milestone boundary stayed green, so the closing loop did not disturb the shipped manual path.
Agent Initialization
2 agents 12m 32s wall time 138k peak context
The command still previews the proposed section, asks for confirmation unless global yes is in effect, and reports whether it created, updated, left content unchanged, or declined to write. The generated block still points future agents at the installed manual instead of copying the manual into the project. Existing project instructions are preserved around the marked Gantry section.
This sprint carried the whole optional init command because the earlier milestones had already established the manual target it should point to. Its brief focused the worker on describe-confirm-act behavior, marked block ownership, and repository-local file handling. It stayed green through the recorded gate and review, so there was no later subdivision or recovery for this final feature.