Suggest roster-aware agent names

[orangewk]

Summary

This draft implements the onboarding half of #178: when an agent is not yet registered for a project, whoami.sh now appends roster-aware unused identity name suggestions.

The goal is to avoid defaulting humans and agents toward ambiguous tool labels like cc / codex when multiple sessions of the same type may coexist. The suggestions are identity names from a bundled pool, e.g. noether, lovelace, turing, gauss, rather than type prefixes.

Design

• Add scripts/lib/roster.sh as a read-only helper:
  • agmsg_roster_names <team> lists current team member names from the team registry.
  • agmsg_suggest_names <team> [count] walks a bundled identity-name pool deterministically and skips names already present in the roster.
• Add scripts/data/name-pool.txt as read-only data, not sourced code.
• Keep suggestions advisory. This does not guarantee uniqueness between suggest and join, and it does not change the identity model or join.sh semantics.
• Update whoami.sh only on join-needed paths:
  • not_joined=true ... suggested=<n1,n2,...>
  • suggest=true ... suggested=<n1,n2,...>
• Leave exact and multiple identity paths unchanged.
• Update agent templates and SKILL.md so agents show suggested= as optional unused identity names.

Explicitly out of scope for this draft

• Making spawn.sh <type> work without a positional name.
• Guaranteed unique allocation under registry-lock.
• Agent-name validation for arbitrary user-provided names.

Those are reasonable follow-ups, but this PR keeps #178 to an additive suggestion flow.

Verification

• bash -n scripts/lib/roster.sh scripts/whoami.sh scripts/join.sh scripts/team.sh
• Manual isolated-skill smoke tests for:
  • no-team not_joined includes suggested=
  • exact match does not include suggested=
  • same-type reuse suggest=true includes unused suggested= names
  • multiple identities do not include suggested=
  • suggested names avoid existing roster members
• git diff --check

BATS was not available in the local Git Bash environment. I added tests/test_roster.bats and extended tests/test_team.bats; CI should run them.

[orangewk]

Review (draft) — findings first

Overall this matches the agreed minimal design and reads well. One blocking item, otherwise non-blocking + nits. Reviewed the branch locally.

🔴 Blocking

B1 — empty-roster suggestion errors on bash 3.2 (macOS).
scripts/lib/roster.sh:99 (and defensively :110): agmsg_roster_contains "$candidate" "${roster[@]}". whoami.sh runs under set -euo pipefail, and on first run / a brand-new team the roster array is empty. On bash < 4.4 (macOS /bin/bash is 3.2) "${empty[@]}" under set -u raises unbound variable; bash 4.4+ is fine. Verified working on bash 5.2 (Git Bash) — it emits 5 names with an empty roster — so this is macOS-3.2-specific and not reproducible on Linux/Windows CI. agmsg explicitly targets bash 3.2 (see the lib/registry-lock.sh rationale).

Effect: the command-substitution subshell errors; whoami masks it via the assignment so it does not hard-fail, but suggested= comes back empty (plus an unbound variable line on stderr) — i.e. the feature silently no-ops on a primary platform, exactly on the first-run path #178 targets.

Fix (one line, both sites): ${roster[@]+"${roster[@]}"} (bash-3.2-safe alternate-value expansion), or an early [ "${#roster[@]}" -eq 0 ] branch. The "${pool[@]}" at :108 is already guarded non-empty at :93.

🟡 Non-blocking

• Test gap (would have caught B1): no test exercises an empty roster under set -u. test_roster.bats uses bash -c without set -u, and the whoami not_joined test pre-joins a team so FIRST_TEAM is non-empty. Add a true no-teams whoami run, or bash -c 'set -u; source …; agmsg_suggest_names <empty-team> 5'.
• Reuse-path collision scope: in the suggest=true path the names are checked only against FIRST_TEAM's roster; if the user joins a different team the "unused" property doesn't hold. Advisory, so fine — a one-line note in the template ("suggestions are relative to the first listed team") would clarify.

⚪ Nit

• Pool ordering: the pool is alphabetical, so the linear walk yields same-initial clusters (offset b → banach bell bernoulli boole brahe). Interleaving domains/initials in name-pool.txt would make contiguous picks feel more varied; determinism is unaffected.

✅ Confirmed good

• roster_names is Windows/CRLF-safe (agmsg_sqlite_mem strips CR; agmsg_sql_readfile_path does cygpath -w).
• Names are JSON keys, so roster_names is registration-shape-agnostic — no legacy-array normalization needed (correct).
• whoami additions are line-end additive; agent= / multiple= paths untouched → existing parsers safe.
• Deterministic (cksum offset, no $RANDOM) — verified identical across runs.
• Templates use __SKILL_NAME__, not a hardcoded /agmsg.
• Pool is identity names, lowercase ASCII, no tool labels — matches the #178 intent; charset is re-enforced at read time.
• Scope is the agreed minimal cut (spawn auto-pick + guaranteed-unique allocation excluded and noted in the PR body).

Fix B1 and I'd consider this mergeable.

[orangewk]

Re-review — B1 resolved ✅

Verified 835c47b locally.

• B1 fixed correctly. scripts/lib/roster.sh:99 and :110 now use ${roster[@]+"${roster[@]}"}, the canonical bash-3.2-safe empty-array expansion. Re-ran the failing condition (set -u + empty roster): exit 0, 5 names, clean stderr. bash -n clean. Correct by construction for bash < 4.4 as well.
• Regression tests added — roster: suggestions are safe with empty roster under set -u (asserts count == 5) and whoami: first run with no teams includes suggested names. The first is the solid guard for B1.

⚪ New nit (test quality, non-blocking)

tests/test_roster.bats:40 — [[ ! "$output" =~ "suggested=$" ]]. Quoting the RHS makes $ a literal, so this matches the literal string suggested=$ and passes vacuously (it does not assert the value is non-empty). Verified: both suggested=ada,… and an empty suggested= pass it. Drop the quotes (=~ suggested=$) to actually anchor end-of-line. Non-blocking — the empty-roster count == 5 test already covers non-emptiness; this assertion is just redundant-and-inert as written.

Caveat (same as before): both empty-array regression tests only bite on a bash < 4.4 leg — on bash 4.4+ the pre-fix code also passed. Worth confirming CI exercises a bash-3.2 / macOS-system-bash leg so the guard isn't latent.

B1 cleared. From my side this is mergeable once it comes out of draft; the remaining items are non-blocking (NB2 reuse-scope note, pool-ordering nit, the test nit above).