Passage-planning session, 2026-08-03: v0 met a human

3 August 2026 · 12 min read · John Kershaw and Claude

A chronicle of the first live human-ratified run of the experimental Passage Planner prompt (LIN-1811), in the tradition of collective-session-2026-06-12.md and flight-companion-session-2026-08-02.md. Every code and contract claim below was re-verified against this repository at the time of writing. Three commits have landed on main since this chronicle's task was filed: cfaee25c (the v0 prompt text session #1 ran), an unrelated KPI-windowing fix, and 31fcf36b — the v0.1 revision, merged at 09:57:58Z, which rewrote the very prompt document this chronicle cites, from 294 lines to 162. Line references are therefore pinned explicitly: defect anchors cite v0 at cfaee25c, where those defects lived and where they can still be read; surviving text is cited at current HEAD. All timestamps are UTC. No transcript of either planning session was retained — both were hand-run claude CLI sessions (a proxy-page bootstrap token plus the pasted prompt doc), which is why neither appears in the dispatch history. This chronicle rests on the flight companion's contemporaneous session notes plus the durable Linear/GitHub artifacts both sessions left behind; every sentence below traces to one or the other, named as it goes.

Cast

John — the human. Ran both sessions, ratified the passage, chose not to exercise the challenge loop.
Planning session #1 — a claude CLI session running prompt v0 (docs/passage-planner-prompt.md at cfaee25c).
Planning session #2 — a claude CLI session running the v0.1 draft (the text attached to LIN-1850's first comment; unmerged at the time it ran, landed since as 31fcf36b — see "What this session changed").
The flight companion — observer, feedback scribe, and the source of every session-interior detail in this chronicle. Supervised from outside both sessions; John pasted key artifacts into their shared channel as the sessions ran.

Boot

Prompt v0 landed at 08:23:42Z as PR #1063 (cfaee25c), closing out LIN-1841. John began session #1 shortly after, per the LIN-1842 recipe: mint a bootstrap token from the proxy page, paste docs/passage-planner-prompt.md into a fresh claude CLI session. Because the session was hand-run rather than dispatched, it left no row in the dispatch feed and no transcript — everything about its interior comes from the flight companion's contemporaneous notes.

Legs

Session #1's orientation reads are not independently recoverable (the overview that would have shown them was not retained), but the same reads resurface, quoted, in session #2's writeout and reproduce today:

North star — GET /north-star reads reading.state: "fresh" and roadmap.state: "fresh" off a report generated 2026-08-02T19:44:55Z. The roadmap narrative contains, word for word, "LIN-1625 is seven days stale and blocks seven tasks—the longest listed chain" and flags Product capability work as drift against "finish transitions before starting capabilities."
Periodicals — GET /periodicals returns 15 templates: 4 due (Documentation Review, 23 days since last run; Code Quality Review, 23 days; Drift & Coherence Review, 22 days; Recent Headwinds, 24 days), 1 recent (Design & Interface Review, 5 days), and 10 never. Per routes/proxy.js:1525-1543, never means no run recorded inside min(30-day horizon, retention) — a bounded claim, not "has never run."
Stack digest — GET /stack?limit=15&view=digest returns each task's why[] reasons verbatim. LIN-1625 carries ["unblocks 6", "critical path 3"]; LIN-1666 carries ["unblocks 1", "critical path 2"]; LIN-1557 and LIN-1558 carry [] (the digest gave no ranking reason for either); the seven stack-digest anchors that became Leg 2 (LIN-1694, LIN-1731, LIN-1594, LIN-1821, LIN-1458, LIN-1848, LIN-1389) all carry ["bug"]. A live re-read today reproduces every one of these values, though it also surfaces tasks that were never anchors — the digest is a live ranked read, not a replayable snapshot, so this corroborates the quoted values without reproducing the session's exact membership.

Dialogue

Session #1 surfaced seven human-interface defects, distilled and posted as comment 4f3d5cce on LIN-1842 at 09:19:22Z:

Serial ratification produced rubber-stamping. One-leg-at-a-time approval (docs/passage-planner-prompt.md:94 at cfaee25c, "Human dialogue and ratification, one leg at a time") gave John no whole-plan context; per the flight companion's notes he was click-approving without reading ("honestly I'm just clicking approve without reading").
Contract-formal tone throughout, with no register guidance in the doc.
A timid, unexplained 4-task budget default, with no sizing conversation in the prompt.
No synthesis-first proposal — the full total → legs → path view only appeared when John explicitly asked for it.
Orientation invisible to the human — the north-star/periodical/digest reads happened but were never shown before John was asked to ratify.
A false ratification record — the session's overview claimed "All four legs are ratified" off the hollow per-leg clicks, with no way to distinguish a genuine yes from a fatigued one.
Machine-legible, human-illegible output — docs/passage-planner-prompt.md:237 at cfaee25c ("Write every leg so it survives runner handoff") engineers every leg to survive brief-distillation but says nothing about human legibility. Per the flight companion's notes, John read session #1's 30-task overview attentively and concluded the periodicals leg was missing; it was present as Leg 4 — the wall of text simply didn't show it.

Neither of those two section headings survives at HEAD: both were the defects v0.1 removed. They are cited at cfaee25c because that is where the defects are readable, not because they describe the current prompt.

John's decision, per the flight companion's notes: iterate the prompt text directly rather than route a fix through the normal pipeline. The companion drafted a v0.1 revision addressing all seven findings while preserving every v0 hard invariant (task-counts-only budgets, the reserved maintenance leg drawn from the shared pool, related-type anchors with an inline mapping, description/append discipline, the write gates, witness-before-runner). That draft is attached as the first comment (44f45b69) on LIN-1850, filed 09:34:02Z.

Session #2 ran the v0.1 draft. Per the flight companion's notes: orientation was shown before ratification, sizing was negotiated (the session proposed a 20–30 range and 26 was set at the midpoint), the full four-leg passage was proposed at once as a table plus per-leg detail, and evidence gaps were stated honestly in the proposal itself — including that the stack digest counted six downstream unblocks for LIN-1625 without naming them, and that the ten never-state periodicals were deliberately excluded from Leg 4 with that reasoning given. John's assessment, per the notes: "much stronger."

The 20–30 range's origin is contradictorily attributed. The flight companion's notes record the session proposing the range with John setting 26 at the midpoint; LIN-1851's own voyage-log comment instead states "John's chosen range was 20–30." Both cannot be the source, and nothing in this session resolves which — the disagreement is recorded here rather than picked.

John ratified the whole proposal in one motion — a single yes to the full 4-leg proposal as presented, with no partial per-leg approvals, per LIN-1851's voyage-log comment (dc7660d5, 09:37:51Z). Per the flight companion's notes, this was an explicit choice: John was validating the planning experience, not flying a passage, and said he lacked run-budget for the passage that day — so the challenge/negotiate loop was never exercised (Step 4, "Negotiate the whole, ratify once", at docs/passage-planner-prompt.md:103 at HEAD). The central fix session #1 lacked is therefore validated only in its propose-and-ratify half; whether a real challenge round works as intended is untested by this session — and that untested half has since shipped, which is worth knowing.

Writeout

LIN-1851 ("Passage — Cost chain, reliability bugs, Product close-out, maintenance (26 tasks)") was created 09:37:16Z: 4 legs plus 3 tasks of unallocated slack, budgets of 7 + 8 + 4 + 4 + 3 slack (summing to 26, matching the stated pool). A live check of GET /relations/LIN-1851 today confirms exactly 12 related-type relations, no blocks/blocked-by anchors, matching the description's inline leg↔anchor mapping exactly: Leg 1 → LIN-1625 (1); Leg 2 → LIN-1694, LIN-1731, LIN-1594, LIN-1821, LIN-1458, LIN-1848, LIN-1389, LIN-1815 (8; the eighth, LIN-1815, was named by the north-star reading rather than the stack digest); Leg 3 → LIN-1666, LIN-1557, LIN-1558 (3). A voyage-log comment (dc7660d5) opened at 09:37:51Z, recording the ratification and the planning-time evidence above. Append-discipline for future edits ("post-creation edits to this description go through description/append or description/replace, never whole-body PATCH") is recorded in-description.

The writeout deviates from the strict block prescribed by the prompt (### Leg: <name> with **Anchors:** / **Intent:** / **Budget:** / **Making port:** / **Wind down if:** lines): LIN-1851's description instead uses ## Leg N — <name> headings with bolded bullets. That block is one of the pieces v0.1 carried over untouched — it is byte-identical between v0 (:253 at cfaee25c) and the current HEAD text (docs/passage-planner-prompt.md:122), so the deviation stands against the shipping prompt, not just against the version session #2 ran. This is left uncorrected deliberately — LIN-1844, the cold-read witness, is the task that measures the format actually written, not this chronicle.

Gate

The prompt makes the ordering load-bearing: chronicle, then a cold-read witness (a fresh context-free session reading only GET /brief/LIN-1851 and stating each leg back), and only then may any runner be dispatched against a ratified leg. This survived the revision — it was section 9 ("The acceptance bar") at docs/passage-planner-prompt.md:265 in v0, and is Step 6 ("Stop at the gate") at :140 at current HEAD, which states it as "witness before runner, no exceptions, because runner activity would contaminate the context-free read." A live check of GET /dispatch?issueIdentifier=LIN-1851 and GET /dispatch?issueIdentifier=LIN-1812 today both return zero items — no runner dispatch exists against the passage task or its gated build-order sibling across the whole window since LIN-1851 was created, and LIN-1812 itself remains Backlog. The gate has held for the entirety of LIN-1851's life to date. This chronicle is the first of the two required steps; LIN-1844 (cold-read witness) is next, and LIN-1845 links the resulting evidence back to LIN-1811.

What this session changed

The v0.1 revision landed. PR #1065 (LIN-1850) merged at 09:57:58Z as 31fcf36b, and LIN-1850 is Done — so within roughly forty minutes of the session that exposed them, all seven defects were fixed in the shipping prompt. docs/passage-planner-prompt.md at HEAD is v0.1: 162 lines against v0's 294, restructured from numbered sections into six conversational steps ("Orient, and SHOW your orientation" → "Size it together" → "Propose the WHOLE passage at once" → "Negotiate the whole, ratify once" → "Write it" → "Stop at the gate"), with a "How to talk" register section added. The merged text carries the validated draft's body unchanged, plus the header-notes block (design-artifact status, the LIN-1380 graduation path, the both-paths sync discipline, and the vocabulary note) that LIN-1850 required be preserved — the draft in comment 44f45b69 omitted those notes, so the merged file is the draft plus that block, not a byte-identical copy of it.

LIN-1849 (a one-click kickoff page for the planner, mirroring the Flight Companion pattern) was filed at 09:00:13Z and deliberately sequenced behind LIN-1850 so it would serve the revised text rather than v0; with v0.1 landed, that dependency is discharged.

The compression was not free, and one thread from it is still open. Shrinking the prompt by 45% traded out two witness-facing specifics that existed nowhere else in the repository. LIN-1850's independent review caught the loss and made preserving them — not necessarily restoring them to the prompt text itself — the single condition on its approval; they were re-homed rather than restored, and now live in LIN-1844's description under its own "Carried forward from planner-prompt v0 §9" heading. Separately, the same revision dropped the single word "alongside" from the maintenance-leg reservation clause (v0 :76/:116 at cfaee25c → HEAD :41/:97), leaving it undetermined whether the now-mandatory maintenance leg counts against the passage's 2–4 leg cap. That was raised as a non-blocking finding on the same review and left unfixed rather than slipped into an approval that hadn't seen it — it is tracked open as LIN-1853.

The session's overall finding: the planner's evidence and writeout machinery worked as designed on the first real run: correct anchors, correct budgets, correct quoted evidence, an honored write gate. What failed was entirely in the human-interface layer — ratification mechanics, tone, sizing, and legibility — and that is the layer v0.1 targets.

Gaps and open threads

No transcript was retained, for either session, though LIN-1842's own acceptance criteria called for one. This chronicle rests on the flight companion's contemporaneous notes plus the durable artifacts cited above, not a replayable record.
LIN-1842's tracker state lags reality: it remains Todo despite having run and produced LIN-1851, and its blocks relation to this ticket is still open.
The 20–30 range's origin is contradictorily attributed between the flight companion's notes (the session proposed it) and LIN-1851's voyage log ("John's chosen range") — see "Dialogue" above; this chronicle does not resolve it.
Session boundaries are inferred, not recorded. No start/end timestamp exists for either session. The defensible bounds: session #1 ran between 08:23:42Z (v0 merged) and 09:19:22Z (findings posted); session #2 ran between 09:19:22Z and 09:37:16Z (LIN-1851 created).
Session #1's own output is unrecoverable. Its 30-task overview, its four proposed legs, and its false "all four legs are ratified" claim survive only as the seven distilled findings in comment 4f3d5cce; the overview artifact itself was not retained.
The ratification was honest but narrow. One genuine yes to a full proposal John had actually read — but the challenge/negotiate loop was never exercised, so that half of the central fix remains unvalidated by this session. It shipped anyway, in 31fcf36b, roughly twenty minutes after the session that half-validated it: v0.1 is the live prompt with one of its two central mechanics never having been run.
Ten of fifteen periodicals read never at planning time, which is a bounded claim (no run inside the retained 30-day window), not evidence that they have never run at all; LIN-1851's own Notes section states this correctly and this chronicle mirrors that phrasing rather than the stronger, incorrect one.
The writeout format deviated from the strict ### Leg: block (see "Writeout" above) — recorded here as a fact for LIN-1844 to measure, not smoothed into "close enough." The block survived v0.1 byte-identical, so the deviation did not get revised away with the rest of the prompt.
This chronicle was overtaken by its own subject mid-writing. v0.1 merged after the first draft was written and before it landed, rewriting the document this chronicle cites and invalidating every line reference in it. The citations above are pinned to explicit commits for that reason. A chronicle of a still-moving experiment is a snapshot, not a settled record — LIN-1844 and LIN-1845 should re-verify rather than inherit.