iksir nestles at hum, drops the agent runtime

[adiled]

Iksīr no longer drives OpenCode. It connects to a humd as a forager bee over the thrum socket and prompts whichever hive al-Kimyawi has kindled. The choice of what burns in the furnace was never Iksīr's to make.

The ~70% that is genuinely Iksīr — Linear, GitHub, istihal branch surgery, mumayyiz triage, quiet hours, code-intel — never moved.

The substrate

New src/hum/:

file	role
thrum.ts	NDJSON unix-socket client — hello, reconnect, sid dispatch
identity.ts	stable fbee_<hex> hid from a persisted ed25519 seed
client.ts	AmilHum — the same surface the khuddām always called
thrum.test.ts	5 tests against a fake humd

The tongue, translated once and nowhere else:

khalaqaJalsa      -> mint a sid locally (jalasat were always ours)
sendPrompt        -> chi:"prompt" -> chunk* -> finish
abortSession      -> chi:"cancel"
mahaqaJalsa       -> chi:"cleanup"
summarizeSession  -> chi:"curate"   (no provider, no hardcoded model)
replyToQuestion   -> chi:"release-permit"
question.asked    <- chi:"permission-ask"

Two laws, taken from humd's source

Both load-bearing, both tested:

1. Manifests are volatile. humd clears them on restart and prunes on disconnect, so the hello rides every connection, not just the first. A silent reconnect without it leaves Iksīr nestled but unrouteable.
2. Iksīr must never declare bee:["worker"]. humd re-broadcasts the output tones of any bee whose manifest carries "worker" onto the sid sigil — and Iksīr is the bee that claimed that sigil. It would hear its own voice returned.

Instruments

src/mcp/ became src/alat/ — Ālāt al-Iksīr (آلات الإكسير). There was an MCP server there because OpenCode needed telling what Iksīr could do; nothing needs telling now. The 23 instruments ride in the hello, humd merges every forager's into the foragerTools it hands each worker, and a nida returns by name.

Gone with it: the JSON-RPC dispatch, http-transport.ts, serve.ts, the iksir-mcp systemd unit and the Requires= chaining the daemon behind it. The instrument bodies are unchanged and each still inscribes its hadath into the ahdath table, so Munaffidh keeps draining a durable journal. The sijill is the record; the thrum is the road.

Persistence

Restoring state asked the amil whether each persisted jalsa still existed — sensible against OpenCode, which owned sessions. Not now: AmilHum answers from a map that is empty at boot, so every row was skipped and katib restored nothing on any restart. The source of truth moved and the permission check stayed behind.

hammalaHala now restores unconditionally and hands each vessel back via istaadaJalsa. nestId rides in the existing hala_mufassala JSON, so a resumed murshid carries its resume token — no schema migration. Two sibling checks went with it, including a title-search recovery path that rebuilt branch names from scratch rather than the persisted ones.

Raqib no longer announces a compaction it cannot confirm. It records the count when it asks, checks later whether the vessel actually shrank, and says so once when it did not.

Verified against a live humd

0.31.18, thrum 0.7.0:

thrum.hello                     proto_version=0.7.0
bee.hid.registered              hid=fbee_86af5280f681
forager.capabilities.registered provides=["session"]
forager.tools.registered        count=23
manifest.evict.disconnect

bee.hid.registered rather than bee.hid.invalid is the one thing a fake socket cannot tell you — a malformed hid breaks reconnect dedup silently and leaks a manifest every time.

144 tests pass.

Known gaps

• Not exercised beyond the handshake. No prompting, tool-call routing, or murshid work against a real nest. The manifest handling was read from 0.32.0 source; this ran against 0.31.18.
• chi:"curate" is a no-op in humd today — it sits in the thrum.recv.todo arm. Raqib's compaction request is accepted and dropped, which is why raqib now measures instead of trusting. Upstream: feat(curate): wire chi:"curate" end to end hum#54.
• chi:"release-permit" likewise, and nothing emits permission-ask, so Saail is unreachable — its permission-ask body shape is inferred from the chi contract rather than observed.
• A worker that curates itself raises no tone Iksīr hears, so only Raqib-initiated curations are noticed.
• Three pre-existing cli.ts JalsaSijill type errors, untouched — they predate this branch.
• Two test files carried uncommitted edits from before this work; the rename had to reach them, so they are folded into the first commit.

Config: opencode.server → hum.{miqbas,namudhaj}, both optional. The socket is discovered (HUM_THRUM_SOCK → HUM_SOCKET → humd's runtime.json → state dir) and the model is chosen by kindling a hive.

What Iksīr asks of al-Kimyawi is one line, and it belongs in the README as a prerequisite: a nestled worker, plus something in the nest providing fs. A murshid without fs can contemplate the runūz but never inscribe them.