fix(config): coerce list/dict literals in 'hermes config set'

[cadamec]

What does this PR do?

hermes config set stores list/dict literals as raw strings, silently defeating every list-gated reader.

The bug

set_config_value (hermes_cli/config.py) only coerces bool/int/float. A list literal like

hermes config set platform_toolsets.discord '["clarify","file"]'

is written to config.yaml as the string '["clarify","file"]'. Every reader gated on isinstance(..., list) — _get_platform_tools, _get_enabled_set, _get_disabled_set — silently ignores it and falls back to its default. The setting looks saved (the echo says ✓, it's in the YAML) but never takes effect. This is the config half of #57063, which reported a deployment running for weeks with a zero-tool agent because platform_toolsets had been set this way.

The fix

Values whose first non-whitespace char is [ or { are parsed with yaml.safe_load. Only real list/dict results are accepted; anything else (malformed YAML, a string like "[unterminated" that happens to parse to a scalar) warns on stderr and keeps the legacy string behavior. Bool/int/float coercion and string-typed defaults (approvals.mode: off staying a string) are untouched.

Tests

tests/hermes_cli/test_config_set_list_values.py:

• list literals stored as real lists (double-quoted, single-quoted, YAML flow style)
• dict literals stored as real dicts
• malformed literals warn and keep legacy string behavior (7 parametrized cases)
• bool/int/float coercion preserved

Affected suites: 95 passed (test_config_set_list_values.py + test_set_config_value.py).

Related

• fix(toolsets): platform plugins without a static bundle silently get zero tools; config set stores list literals as strings #57063 (the other half — 8 static platform bundles — is obsolete on current main since 52d9e57's dynamic toolset generation; empirical evidence posted on that PR)
• fix(plugins): pre-register deferred platform client tools at discovery; coerce config set list/dict literals #79432 (was bundled with this fix; split per triage feedback so the a2a client-tool fix can be evaluated on its own)

[alt-glitch]

This was generated by AI during triage.

Duplicate of #57063: both patches correct hermes config set persisting list/dict literals as strings, which silently defeats list-gated configuration readers.

[cadamec]

Fair flag, but this isn't a clean duplicate of #57063 — and my own empirical check on #57063 (Aug 5) shows why.

#57063 bundles two distinct failure modes, and the primary one no longer reproduces on current main. Commit 52d9e57 ("feat: dynamic toolset generation for plugin platforms") already fixed the "eight platforms resolve to zero tools" case — I verified all eight now resolve to the full 61-tool core toolset after discovery, in any process. That half of #57063 is dead.

What still reproduces live are the two remaining halves, and #79479 targets one of them: hermes config set persisting list/dict literals as strings (config.py:4903-4914, bool/int/float-only coercion), which silently defeats every list-gated reader. #57063 is also sitting CONFLICTING (stale) and its body still describes the now-obsolete zero-tools mode.

So rather than a duplicate of a healthy PR, this is the current, rebased, tested fix for the one live half of #57063 — a successor, not a clone. The a2a half is tracked separately in #78050. Happy to rebase/retarget or close #57063-scope accordingly; goal is the fix landing on current main.

[cadamec]

Correction: the triage bot was right and my earlier reply was wrong. I checked #57063's diff (not just its body/merge state) and it already contains this exact fix — same yaml.safe_load + isinstance(parsed, (list, dict)) approach in hermes_cli/config.py, same test file. I duplicated it without knowing.

Closing this as a duplicate. The real gap is that #57063 is stale (CONFLICTING with main) and its body still describes the 8-bundle half that 52d9e57 already made obsolete — I'm going to rebase it onto current main and narrow it to the live config half, preserving the original author's work per the contribution rubric.

[cadamec]

Closing as duplicate of #57063 (see correction above).