Ghostlight/docs/architecture/ghostlight-play-agent-postmortem.md

Play agent postmortem

Written 2026-09-23 from the cut map, the commits, the Soul reports and an Eyes pass over all three. The cut map (docs/architecture/ghostlight-play-agent-cut.md) remains the detailed record: every cut's status, findings PA.f1 to PA.f195, and every ruling.

Summary

Ghostlight Dungeon now runs on one operational agent. It reads the world through an omniscient table view, calls kernel tools directly as the Play authority, dispatches characters for their own narrative turns, asks the player questions, and closes each turn with narration. world.speak and the story card are gone; world.play and a single player card replace them.

Seventeen cuts and about ten fix batches landed over two days. 195 findings were recorded. The kernel gained a Play authority, ruled facts, minting, retirement, affordance grants and a callable Persona turn. A loopback-only local inference port landed so the whole thing can run against a small model on the operator's own machine.

State at writing: the loop is built and verified as far as tests and a falsifier reach. Nobody has played it, and no model has ever run behind the local port. The seed path has never executed past its configuration check. All evidence is Windows; the release evidence is Idunn's Linux gate on Yggdrasil.

Two capabilities the design promises do not exist: fiction-first risk resolution and gestalt agency. They are named in Open follow-ups and carried into the target doc as gaps rather than promises.

Scope and invariants

The pivot: the previous machine was built for offline narrative generation and world simulation, which is the wrong set of compromises for play. The Persona sandwich stays, because actors should think narratively; everything else — turn dispatch, interpretation, prompting the player, negotiating the result — becomes one operational agent's job.

Eight invariants scoped the work. The kernel is the only writer and no transcript crosses turns; Personas never see structured state, and the agent never rewrites an existing character; an actor acts only through its own affordance; speech is verbatim, checked against that actor's own prose; there is one play authority; removal is retirement; one loop serves every backend; and the player sees a projection, the question, and the refusal of their own act, and nothing else.

Explicitly out of scope: the offline simulation lanes, which keep total interpretation; Njordr's evidence invariants, which the operator ruled do not bind the play agent ("if the agent can invent people, it can invent loot"); and the world compiler.

Timeline

Cut	What	Commits	Notable finding
1	Delete the autonomous drivers and owner controls	d69e9d4	−2,184 lines. A re-added tick loop passed 46 of 46 tests
2	The local inference port	182f803 + fixes	HTTP_PROXY sent the whole request, and a credential, to a proxy (PA.f16)
3	Play authority, ruled facts, minting	4c4973b	The genesis lane never ran require_ruler (PA.f26)
4	Retirement	3725ba8	A retired subject was still an audience and a raw speaker (PA.f33)
5	Grant and revoke affordances	732b1d3	Two loosenings survived every test (PA.f30, PA.f35)
6	The Persona lane and the recency marker	c4334cc	The spec's own stated rationale was impossible; Hands corrected the spec
7	The table vocabulary and inference seams	7d40393	actor_tools emitted names decode_actor_call refused (PA.f55)
8a	The play table and the turn loop	6b69270	communicate let the agent make any actor "tell" an invented fact (PA.f61)
8a-fix2		74b628d	Every mutation reported killed; eight rules unguarded
8a-fix3		754c7fe	The id spelling had four copies; changing the printer passed 735 tests and broke every dispatch (PA.f113)
8a-fix4		seven commits	Not ready for 8b: a crash window stranded a dispatched Persona (PA.f125)
8b	Wiring, routing, the Eve schema	1b20a0f, a99fdfe	ask_player wedged the world: no client could build an answer (PA.f134)
9	The play surface; world.speak deleted	322bdf6, f1732c0	The real client could not submit anything at all
10	The client round trip	d9f0159, 43b6af6	world.create had no field for brief: no client could ever create a world
11	The answer path and the gate	4155823	Two of three bridge tests failed instead of skipping without node
12–13	The store schema; the answer token	7052644, 2f2a8b6	An answer composed for one question resolved another (PA.f170)
14	Preflight fixes	58324f8, 0808c6a	The seed refusal handed the player the inference endpoint's URL (PA.f187)
15	The closing fixes	f295a3a … 1dea8f5	An unwritten invariant, stated and broken in one commit (PA.f193)
16–17	Closing the bar	63964b0 … 07aae19	The runbook's weekly check printed ok without running (PA.f194-B)

Structural delta

New: local_inference.rs, table.rs (the consumer-neutral vocabulary), play.rs (the turn loop), tools/eve_client_bridge.mjs and its committed fixture. Two dependencies: reqwest for the local port, unicode-segmentation for UAX #29 word boundaries.

Deleted: Cut 1's tick driver, Active elaborator and owner "let X act" path (−2,184 lines, and ten tests whose only subject was deleted code); Cut 9's SpeakPayload, world.speak arm, story card, speak control and current_operator_view.

Tests: the library went 466 → 622 and has been flat since the play loop's third fix batch — all later churn is Dungeon-side. The Dungeon binary went 56 → 46 (Cut 1's deletion) → 182, plus four client-bridge tests that are #[ignore]d behind an opt-in. ghostlight-persona-projection went 13 → 35.

The subtraction budget was met once, at Cut 1, and not thereafter. Every fix batch was net additive, mostly in tests. Soul audited the largest claim (Cut 13's "~150 lines") and found 29 lines of production code; the rest was comments. That is defensible, but the ledger stopped being a pressure after Cut 1 and nobody noticed until this postmortem.

What Soul caught

Of roughly 195 findings, about 150 were introduced by this pass's own earlier steps rather than inherited. That is the honest shape of a foundation change: each fix is a fresh chance to be wrong, which is why the falsifier never grades its own work.

Split authority (~14). Four copies of one id spelling, because the library's printer was private to its module (PA.f113). Two computations of the retired set and of the exercise revision (PA.f82). Two owners for round entry (PA.f127). The resume path's collision check duplicated verbatim (PA.f118).

Tests that pinned spelling or passed trivially (~13). A dependency check asserting contains("0"), always true. A test that failed one run in four and checked the wrong subject when it passed. Two refusal tests whose expected kind was a substring of their own fixture's handle.

Mutations aimed at the wrong layer (~4). Deleting an in-memory record instead of the on-disk write it stood for; testing a helper instead of its call site; every test using a fresh key, so deriving the turn id from the key survived.

Inputs built from a copy of production logic (2). The dispatch tests spelled their subject id with a fourth copy of the printer, so changing what the world view actually printed left 735 tests green while every dispatch an agent could make was refused.

Forged input accepted (~6). A forged opportunity, a forged revision, a deserialized colliding subject id, and a ruled fact riding in through world creation.

Player-visible leaks (~10). The question id rendered as a visible, editable box. A connection-refused error handing the player the inference endpoint's host and port. A subject census on a card any authenticated account could see.

Lifecycle and crash windows (~11). A fresh request key discarding an open turn; an empty answer wedging it forever; a spent round budget leaving it running with nothing said; a crash between persisting a Persona's prose and recording the call that produced it, which stranded that character for the rest of the turn.

Things the real client could never do (~5). The binding shape did not match what the client sends, so every command submitted an empty payload — including world.create, which also had no field for a value its own contract required. Nobody had ever driven the real client end to end.

Operator corrections

Eight questions were ruled by the operator directly, and one correction changed a design mid-pass. Self had suggested updating an existing character's attitude through set_persona_material; that collapses the no-puppets rule, and the ruling was reversed so the agent authors new characters and never rewrites an existing one. The operator's review — "I saw those calls in your traces, I've been watching. No objections" — arrived after the reversal rather than before it.

The operator also cut a scope error at the root. Self proposed binding the play agent to the offline lanes' evidence invariants; the ruling was "if the agent can invent people, it can invent loot", which is what made ruled facts and minting coherent rather than a hole in the model.

What made the wrong answers look right in both cases: the agent was reasoning from the library's offline discipline, where those rules are correct, without noticing that the play path has a different author and a different purpose.

Incidents

• The workstation rebooted mid-pipeline under concurrent cargo builds, mine among several sessions'. Rule: one cargo process at a time on the workstation, in every brief, and verification moves to Yggdrasil under Idunn.
• A Hands agent amended and force-pushed over its own first push of Cut 3. Nothing was lost, because it had rebased first. Rule: never amend, never force-push; it is now in every brief and in the skill.
• A Soul pass died on an API 529. Its worktree was cleaned and the pass rerun.
• A git pull --rebase was refused while Hands had unstaged work in the shared tree. Rule: never pull in a tree where Hands is working. (Observed in session; not recorded in the cut map.)
• PowerShell map edits failed on an overload error and on mismatched anchors. Rule: Python scripts that assert each anchor occurs exactly once and preserve CRLF. (Observed in session.)
• A stray node_modules left by an agent's own debugging made a test count report success for tests that could not run on a clean checkout (PA.f162).

What worked

• Quoting the cut inline in the brief instead of pointing at a 3,900-line map. The one run that read the whole map reached ~400k tokens and produced the pass's sloppiest claims.
• Hands stopping on a real fork. Cut 9's counter had no honest derivation; Hands traced why, stopped, and reported. The ruling took one message and the cut resumed correctly.
• Soul driving the real client. Every blocker in the last third of the pass came from running the vendored browser lowering over the served surface rather than from reading it.
• Recording a finding as closed-not-fixed when the mutation proved the code unreachable. Hands built a guard, mutated it, found an existing guard already refused, reverted its own work and kept the test. The mutation decided, not the diff.

What to change in the pipeline

Three lessons are already in the Eureka brief and changelog:

1. Mutate the production call site, at the layer the rule protects — a later write can hide a deleted one (fd01545).
2. A test's inputs come from the production path, never from a helper that re-spells what production prints (2e0ea97).
3. Name the inference, not the line (181e819). One defect shape ran through five instances: a committed file exists, therefore the client works; git exited zero, therefore this is a checkout; the format check failed, therefore the row is unreadable; the schema is unknown, therefore the row cannot be read; cargo printed ok, therefore the fixture is live. Each local fact true, each inference false, each fix scoped to the reported instance. And after changing how a test reports, run the owning runbook's commands verbatim and read what they print: that surface is the one a test suite structurally cannot check.

Two more this pass earns:

4. A subtraction ledger that stops being updated stops being pressure. It held at Cut 1 and then went unexamined for sixteen cuts.
5. Soul's closing pass should be briefed for a walk-through, not only a verdict. Asking what happens when a person actually sits down and plays produced the most useful output of the pipeline, including two defects no invariant covered.

Open follow-ups

Item	Where	Why it can wait
Fiction-first risk resolution	kernel + play.rs	The kernel resolves a risky act through an affordance's weighted outcome bands, drawn deterministically. There is no assessment step, no modifiers, and the agent cannot author affordances — so an act no granted verb expresses is authored directly as a ruled consequence with no draw at all. Those two paths are indistinguishable to the player. This is the next pipeline.
Gestalt and institution agency	kernel	Institutions exist as structure — offices, grants, jurisdiction — and constrain who may act. Nothing makes them act. Advancing time rolls routines and accumulates pressure; nobody deliberates off-screen.
The seed leg	runtime.rs, vault	Never executed past its configuration check. GHOSTLIGHT_SEED_VAULT_ROOT appears in no runbook. Attempt it alone, first.
PA.f194-A's fourth exit	play.rs	Closed in Cut 17, listed because its rule now lives on retire_unreadable_row as a precondition and every future caller must satisfy it.
PA.f195-B, PA.f195-C	play.rs, runtime.rs	A dead tuple element, and a skip notice cargo swallows without --nocapture. Trivial.
state/map.yaml is not parseable YAML	state/	Latent: nothing reads it programmatically. Fixing it is a deliberate sweep, not a drive-by.
A read path for a retired row	play.rs	The sidecar is JSON with the payload in base64, so an operator can read it. Nothing imports one back.