Herdr team coordination policy for AI coding agents
93%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
Use Google's native Antigravity CLI, agy, with its existing Google subscription
sign-in. The adapter adds no OAuth client, credential store or API proxy.
The CLI's own authentication contract is at
https://antigravity.google/docs/install?tab=cli.
agy on the Herdr controller host and sign in using the CLI.
Keep the shipped composer_glyph: "> " setting; configuration validation
refuses an omitted glyph before measurement, reservation or launch. Startup,
clear and relaunch still require the existing ANSI and empty-composer proof.agy models and agy -p /usage --output-format json on that account.
Missing access or subscription quota remains unavailable, not inferred.antigravity worker kind from
skills/herdr-foreman/antigravity.example.json into the existing operator-owned
schema-7-or-later config. Preserve existing providers, preferences, judge and
foreman. Model IDs are examples; use the account's actual available IDs.
A 0.16 profile with effort on Claude rows or Antigravity judgment rows is
refused; remove those fields and rows.kind: agy is the native Herdr start kind. Configured tiers use --model;
exact process argv and explicit --conversation UUID resume use the existing
launch verifier. Claude models take no --effort: CLI 1.4.7 answered an
unsupported effort by running Gemini instead. Omit effort on those rows; they
record none, and the effort they actually run at stays unknown. No Antigravity
row therefore fills a judgment round, and high-risk work refuses rather than
inventing an effort step. Other models still name an explicit effort. The
supported grammar and judgment pins live in skills/herdr-foreman/foreman/tiers.py.
A model appearing in a catalog does not prove access, qualification or
successful delivery.
Argv only requests a model. Before any task input, the worker's live footer
must show the requested model's captured native label. Effort remains unknown;
the completed turn also needs native conversation evidence of that same model
selection before acceptance. A substituted, missing or
uncaptured label refuses with nothing sent. Captured labels live in
skills/herdr-foreman/foreman/antigravity_native.py; a model or effort
without one, including every Gemini row today, stays unadmitted until a
native launch of that exact pair is captured.
No permanent pane is reserved. Assignment-scoped workers use the existing launch, enrollment, report, verification and cleanup owners. Keep the existing Claude/Codex/Grok foreman; Antigravity foreman reset and Stop ownership are not enabled by this worker adapter. Keep the observer on a supported Claude, Codex or Grok session too; Antigravity observer registration is rejected before cache writes. Antigravity workers remain visible to the observer. Remote-account measurement needs its native controller adapter; a local Google account cannot establish remote capacity.
skills/herdr-foreman/foreman/antigravity_native.py owns the native usage command,
structured quota parser and conversation transcript contract. Quota lookup uses
the CLI slash command without a model turn and removes API-key/ADC overrides
from that lookup. Workers must use the same subscription sign-in; do not configure
API authentication for this profile. Provider failures stay local to its
measurement record; other eligible providers continue.
Each native quota group remains separate in the snapshot. Current planning conservatively uses the limiting group across the profile. It does not infer that a model family grants a separate account or usable extra headroom. Existing account-bound billing evidence and window-group rules remain unchanged.
Herdr's installed antigravity-cli integration reports agent: agy with source
herdr:antigravity_cli. Native delivery reads only that conversation's official
brain/<conversationId>/.system_generated/logs/transcript.jsonl, the latest
completed model response and the exact bare report marker. Stable current pane,
session and source readback remain required. Arbitrary archived JSONL cannot
prove an Antigravity session; unsupported source formats remain unconfirmed.
Recovery requires the exact dispatched request and the dispatched tier's captured
native model label. Missing, unsupported or contradictory native evidence stays
unconfirmed, including follow-up turns. The envelope grammar and model-selection
proof are owned by TRAILING_BLOCKS, user_request and model_selection in
skills/herdr-foreman/foreman/antigravity_native.py; read those definitions when
diagnosing a parser refusal instead of reconstructing its predicates here.
Tessl packages include the native Antigravity SessionStart hook using Google's
conversationId/transcriptPath input and injectSteps output. Both Tessl and
ACR carry the same runtime adapter and scoped worker protocol. ACR's supported
native host layouts remain unchanged; workers receive their scoped protocol in
the existing verified assignment even where no startup hook is rendered.
.tessl-plugin
rules
skills
herdr-foreman
classify
foreman
references
specialists
templates
herdr-observer
herdr-optimizer
herdr-standup