Loft 2.4: see whether your agent read the specification first
Yard now counts what your coding agents read and edit on each branch, and shows for each stage whether the specification was consulted before the code was first edited. It ships with open-source Loft 2.4, in the plugin for Claude Code and Codex.
What Yard counts
Loft 2.4.0, released on November 17, 2026, adds consultation to Yard. As your agent works, the plugin sorts each tool call that reads or edits a file by where the file lies:
- Specification
- A file under the repository's
spec/. - Code
- Any other file in the repository, outside
.git/and.yard/. - Other
- Files under
.git/and.yard/, and files outside the repository. loftreads- A shell command that runs
loft check,find,terms,reconcile,verify,status,review, orrender, each of which reads the specification for the agent.
Each call counts toward the stage the agent is working in, and within it toward the main agent or the subagent that made it. For each stage, Yard also keeps when the specification was first consulted, by a read or a loft read, and when the code was first edited. A stage whose code was edited first counts the edits that came before.
Yard also estimates how much of the agents' context each class brought in. It takes how much the prompt grew between two model requests and shares the growth among the calls that ended between them. Where a request's token counts are missing, the estimate is marked partial.
- spec reads
loftreads- code reads
- other reads
- Specify the limit no code edited 14 spec reads, 6 loft reads, 4 code reads, 2 other reads.
- Build the refusal ✓ spec first 9 spec reads, 3 loft reads, 26 code reads, 4 other reads.
- Fix the refusal message ✗ code first, 2 edits 2 spec reads, 1 loft read, 7 code reads, 1 other read.
- Link the trace ✓ spec first 3 spec reads, 5 loft reads, 8 code reads, 1 other read.
- main agent 27 spec reads, 15 loft reads, 31 code reads, 6 other reads.
- Explore subagent 1 spec read, 14 code reads, 2 other reads.
loft reads count as the specification.Where you see it
On the Yard page, each stage carries a line in the timeline and the Gantt chart, such as spec 12 · code 26 · spec first, where the first number adds the reads of the specification to the loft reads. The Consultation view charts the counts, as drawn above. An open page shows a counted call within two seconds.
loft yard consultation prints the same sums as a table, or as JSON with --format json:
What the agents consulted on feat/pause-limit
stage spec loft code other spec edits code edits spec first
Specify the limit 14 6 4 2 5 0
Build the refusal 9 3 26 4 0 11 yes
Fix the refusal message 2 1 7 1 0 3 no (2 before)
Link the trace 3 5 8 1 0 2 yes
in all 28 15 45 8 5 16
agent: main 27 15 31 6 5 16
agent: Explore 1 0 14 2 0 0
Estimated context tokens (an estimate): spec 127700, code 214200, other 24700, unclassified 45400
Each session writes its counts to a file under .yard/<branch>/consultation/, beside its token totals, so they're committed with the work and your reviewer reads them beside the diff.
When you'd reach for it
Return to Ada from the 2.3 Update, limiting how long a customer can pause a subscription on feat/pause-limit.
Check a stage before you read its diff
Ada reviews Build the refusal. Its line reads spec 12 · code 26 · spec first: the agent read the refusal's criteria and ran loft find on them before it changed any code. She reads the diff against those criteria.
Catch a fix that skipped the specification
Ada rejected the first refusal message, and the agent opened a stage to fix it. Its line reads code first, 2 edits: the agent changed the message twice before it read the criterion the message has to meet. The counts don't say the fix is wrong. They tell Ada which change to check against the specification first.
See what subagents read
By agent, the Explore subagent read 14 files of code and one of the specification. When the main agent plans from what a subagent finds, Ada can tell it to send the subagent to spec/ as well.
See where the context went
The estimate puts about half of the branch's context on code and under a third on the specification. Across branches, it shows how much of the specification your skill and your instructions bring into the agent's context.
Answer beside the session
In Claude Code, /yard-pane opens a pane beside the session on the branch's Yard Log: the stage in hand with its active tasks, and what waits on you, oldest first. A question's options are lettered cards, with a field beneath for your own words, and approvals and plans have Approve and Reject. The first answer counts, whether it was given in the pane, in Claude Code's own dialog, or on the Yard page.
In Claude Code Desktop, the command also opens the Yard page at the session's branch in Desktop's browser pane. In the terminal, VS Code, and the mobile app, the pane links to the page.
loft yard page prints the page's address for the worktree and its branch, or says that no page server is running. Codex gives a plugin no commands of its own, so the skill $loft:yard-page has the agent run it and give you the address, which now reaches a run in Codex's sandbox too.
Also in 2.4
LOFT_YARD_HOME, set to an absolute directory, keeps the machine's Yard records there in place of~/.yard/, so aloftunder development runs its own page server beside an installed one.- Loft's specification, code, skill, and plugins are written in American English. The Codex notification ledger now writes
canceled, and still reads the spelling an olderloftwrote.
Upgrade
npm install --save-dev @ask-loft/loft@2.4.0
cargo install --locked loft-cli
Then run loft skill install, and its setup report says how to bring an installed plugin to the same version. Consultation needs the 2.4 plugin: a session of an older one counts nothing, and the Consultation view lists it as unmeasured. The changelog (github.com, opens in a new tab) lists every change in 2.4.0.
On the next branch your agent builds, read each stage's line before you read its diff.