TDD: org — roles, chains, ownership-as-data (draft)

docs/features/org/handoff-2026-08-25.md

@@ -0,0 +1,211 @@
+# Handoff — org reconciliation, 2026-08-25
+
+> Paste this whole file to a fresh analyst session. It is written to be read
+> cold. Modify nothing; this is a read-and-decide pass.
+
+## Where this stands in one paragraph
+
+The org runtime **shipped** — `cmd/org` and `cmd/org-mcp` are on workbench
+`main` (#262, #263). Charter, attach, assign, claim, yield, checkpoint, boot,
+status, blob, log, plus two hook scripts. The re-entry surface a previous brief
+called "the closest-to-buildable gap" is built. What it is *not* is **wired**:
+`~/.claude/settings.json` still has only `PreToolUse` and `PostToolUse`, and
+zero mentions of org. Meanwhile a second, independently-built Stop hook is
+sitting in another repo waiting for the same slot.
+
+## 1. The live collision — start here
+
+**Two Stop hooks exist, neither is wired, and they do overlapping jobs.** They
+were built days apart by different sessions and neither references the other.
+
+| | `workbench/cmd/org/hooks/stop-mark.sh` | `hooks/scripts/stop-discharge.sh` (hooks PR #42) |
+|---|---|---|
+| writes | a **mark** to the role's chain | a **discharge note** to the dossier tasks the session touched |
+| content | session id, turn count, transcript path | turns, files written, PRs touched |
+| resolution | `cwd` → `roles.map` longest-prefix | the session called `task_update`, which named the task |
+| model involved | no — explicitly refuses | no — same argument, independently reached |
+| on a bare mark | fold renders tip **Degraded** | n/a |
+| wired | no | no |
+
+Read both headers. They reach the *same* conclusion in different words —
+`stop-mark.sh`: *"a record the working agent is required to write is a verb
+wearing a costume"*; `stop-discharge.sh`: claiming at the END is a report,
+claiming at the start is a prediction and agents skip predictions. Two
+implementations converging on a design principle is the strongest signal in
+this corpus. It is also a duplication that will double-write if both are wired.
+
+The matching read side, `cmd/org/hooks/sessionstart-boot.sh`, injects a
+byte-capped `org boot` index (default 2048 bytes: charter, held work,
+obligations, liveness, last incarnation's final word — pointers, not a dump).
+That is the SessionStart read `drive:docs/features/discharge/spec.md` (PR #46)
+designed and never built.
+
+**The question to settle first:** are these one mechanism or two? If one, which
+survives, and what happens to the other repo's PR? If two, state the property
+each provides that the other cannot, in a sentence that would survive review.
+
+## 2. What shipped or moved recently
+
+Merged to workbench `main`:
+
+- **#262** `feat(org): the Baton home — a runtime for role continuity chains`
+- **#263** `feat(org): platform seams — JSON receipts, presented identity, org-mcp, operator context.d`
+- Also merged: #247, #253, #257, #259.
+
+Open, relevant: **wb#265 `feat/org-sweep`** — check what it is before proposing
+anything adjacent.
+
+## 3. What the previous session produced
+
+Four PRs, all CI-green, none merged:
+
+- **hooks#42** `feat/stop-discharge` — the Stop hook above. Has had one review round.
+- **hooks#43** `feat/discharge-sweep` (stacked on #42) — finds sessions that owed a
+  discharge and never paid, and can backfill. **Measured on the real corpus:
+  21 sessions owed a discharge on 49 tasks, 0 recorded.** That 0% is because
+  #42 is not wired.
+- **drive#47** `feat/watch` — a resident watcher tier. Findings in SQLite, one
+  writer, deliberately unable to act. One watcher over `gate next -json -live`
+  (parks, expiring grants, repos with no grant). Verified against live gate
+  state: six parks found on the first pass, zero re-notifications on the second.
+- **wb#245** `docs/org-tdd` — carries `vision.md`, `store-decision.md`,
+  `p0-findings.md`, `reconciliation-brief.md`. **Do not merge until the
+  operator locks it.**
+
+`reconciliation-brief.md` on that branch is the previous handoff. **Its §5 is
+now stale** — it treats re-entry as unbuilt. Its §2 (the metaphor is not the
+mechanism) still holds and is the governing constraint; see below.
+
+## 4. The governing constraint
+
+From the operator, and it survives everything above:
+
+> Why does the distinction of an org even matter at all? Previously I was doing
+> "driver" and "worker" agents, but it's all the same. Ownership / assignment,
+> that's all that matters. Yes, org gives us hierarchy — but that is usually in
+> effect to something like permissions or something else, and that can be
+> represented any way.
+
+An org chart is a **proxy** for mechanical properties. Build the properties.
+
+| org term | the property it proxies for |
+|---|---|
+| role / lead / IC | a durable assignee identity that outlives a session |
+| **hierarchy** | **authority scoping — who may do what, where** |
+| reporting line | where a conclusion must land to be found by the next reader |
+| org chart / tree | possibly nothing — at 2–5 assignees under one operator, depth is 1 |
+
+**Standing instruction:** wherever a source document uses an org word,
+substitute the property and re-ask the question. If the question dissolves,
+that is a finding. `vision.md` §4.5 (ownership is a tree), §4.6 (three things
+bubble) and §10.1 (fold per node kind — called *"the real design work behind
+d3"*) all exist to compress a hierarchy. Check whether any survives at depth 1.
+
+The operator's clarification, which matters: this is **not** a claim that any
+existing implementation solved it. Crown nothing by default.
+
+## 5. Ground truth — earlier passes got these wrong
+
+- **`~/dev/gate` is ARCHIVED** (last commit: `docs: archive banner — gate
+  migrated into workbench as cmd/gate`, 2026-07-17). Live code is
+  `workbench/cmd/gate` + `contracts/gateauthorization`. Only `~/dev/gate/state/`
+  is live — the `GATE_STATE` journal, ~5000 records.
+- **switchboard is `~/dev/agents-as-processes-gleam`.** Gate C2 measured that
+  residency does **not** buy crash recovery — it buys *serialized ownership*.
+  Measured; do not re-derive.
+- **`hack-branchroom` is the ANCESTOR of `contracts/org`**, not a rival. See
+  the dossier note on task `p1-t3-reduce` in project `org`.
+- **`braid` and `reprise` are dead** on their own kill conditions
+  (`~/dev/bakeoff/haskell-08-10/scorecard.md`, `haskell-dsl-08-10/`).
+
+## 6. Two corrections worth learning from
+
+Both were made by the previous session; both are instructive about how this
+corpus lies to you.
+
+**The notes error.** It claimed twice that `task_update`'s notes could not be
+read back — first that no CLI or MCP path returned them, then, narrowed, that
+only `task_get` did. Both wrong: `dossier task_list` returns a structured
+`notes` array. The mechanism is the lesson — `Task.notes` is
+`#[serde(default, skip_serializing_if = "Vec::is_empty")]`
+(`~/dev/dossier/src/domain.rs:184`), so the key is **absent** from any task with
+no notes. The check ran `jq '.[0] | keys'` and generalised from one row. Most
+rows have none. A conditionally-serialised field and a nonexistent one are
+identical from a single sample. `store-decision.md` finding 2 is now withdrawn;
+hooks#43 was rewritten to read through the CLI (and got 5× faster: 15s → 3.1s).
+
+**The portability error.** Every hooks#43 test passed locally and every one
+failed on CI. Cause: `mktemp -t discharge-sweep`. BSD accepts a bare prefix
+after `-t`; GNU reads it as a template and refuses it for having no `X`s. Same
+family as the absent-`timeout` trap already documented in that repo — a
+coreutils divergence that makes a script macOS-only while it looks portable and
+passes every local run. **When reviewing anything bash here, check BSD/GNU
+divergence explicitly.**
+
+## 7. The corpus
+
+Implementations touching ownership / authority / evidence:
+
+| repo / package | primitive |
+|---|---|
+| `workbench/cmd/org` + `contracts/org` | role chain, CAS-on-tip, charter/attach/claim/yield/checkpoint/boot, two hook scripts |
+| `workbench/cmd/gate` + `contracts/gateauthorization` | merge authorization; operator-minted grants, tier ceilings, TTLs |
+| `~/dev/drive` | `Scope`/`attach`/`link`/`release`; scope-bound capability, successor re-mint, orphan adoption; liveness derived |
+| `~/dev/bailiff` | `Warrant` — a capability an agent *holds*: scoped, use-capped, expiring, operator-revocable |
+| `~/dev/hack-mandate` | signed delegation pinned to an exact revision; a child may only narrow |
+| `~/dev/hack-obligation` · `~/dev/hack-proofline` · `~/dev/hack-branchroom` | evidence frontier · identity lineage · causal fork (ancestor of contracts/org) |
+| `~/dev/warrant` | `Reduce` over a journal; refuses to advance without evidence bound to the subject |
+| `~/dev/parley` | protocol kernel — Haskell compiles, Gleam enforces, Lean proves |
+| `~/dev/agents-as-processes-gleam` | switchboard |
+| `~/dev/hooks` | the bash hook layer; PRs #42/#43 |
+
+## 8. Use the instruments before reasoning
+
+**`~/dev/warrant/cmd/gate-observe`** replays a foreign evidence journal against
+a definition written afterwards, with no adoption by the observed system:
+
+```sh
+go run ./cmd/gate-observe ~/dev/gate/state/log.jsonl
+```
+
+Against gate's live log (415 runs, 5010 records, see
+`~/dev/warrant/docs/gate-history.md`): **`escalate` is 57% of all verdicts**
+(1218 of 2145, vs 809 pass and 118 block) — the most common real outcome was
+unrepresentable in the model, which grew a third verdict. One verifier's
+refusal path has never executed in 415 runs.
+
+Research already done, do not repeat: `agents-as-processes-gleam/docs/evidence/`
+(8 pre-registered docs, Gate A→C2), `~/dev/workbench-laws-lean/` (Lean 4 model
+of gate's verdict laws, explicitly not production authority),
+`~/dev/bakeoff/agent-substrates-08-21/{scorecard.md,org-compute-synthesis.md}`.
+
+## 9. Constraints
+
+- **Cite `file:line`.** A claim without a citation is noise.
+- **Mark every finding `measured` or `reasoned`.**
+- **Code beats docs.** These docs overclaim; say so when they differ.
+- **Crown nothing by default** — not org, not drive, not for being furthest along.
+- **No new store, no new repo, no rewrite.** Five stores have already died here.
+- Measured failure modes in this portfolio: scope errors 43%, stale claims 23%,
+  confabulation 8%. Re-verify anything that sounds settled, **including §5.**
+- Read-only.
+
+## 10. Output
+
+At most 2000 words.
+
+- **(a) The Stop-hook verdict (§1)** — one mechanism or two? Which survives,
+  and what happens to the loser's PR? This is the deliverable; everything else
+  is supporting.
+- **(b) Wiring recommendation** — should `sessionstart-boot.sh` and a Stop hook
+  go into `~/.claude/settings.json` now? What breaks if they do? Note the
+  injection cost: bytes injected at SessionStart enlarge every cached turn
+  thereafter, not just the first.
+- **(c) Property table** — for each property in §4, which implementations
+  provide it, and which should win.
+- **(d) What dissolves** — design work that stops being necessary once the org
+  metaphor is substituted away. Name sections and files.
+- **(e) Subtraction list** — what to archive or fold, with the argument.
+- **(f) Next three build steps, in order**, each with what it makes true, why it
+  precedes the others, and its kill condition.
+- **(g) The single riskiest assumption** in (f).