mercury-agent 0.16.0 → 0.16.2

package/docs/goals/football-reporter-profile/decisions.md

@@ -0,0 +1,479 @@
+# Decisions: Football Reporter Profile
+
+**Goal**: [football-reporter-profile](goal.md)
+
+> Inherits every decision in
+> [`whatsapp-bot-hardening/decisions.md`](../whatsapp-bot-hardening/decisions.md),
+> in particular **`whatsapp-bot-hardening` D-014** (one instruction, one layer;
+> countable rules). Bare `D-0NN` ids below always mean *this* goal's sequence.
+> No conflict with that goal's tech choices — this goal adds no new storage,
+> framework or runtime.
+
+## D-001: The football profile is a *space-scoped* folder in the mercury repo
+- **Category:** architecture
+- **Decided:** `examples/profiles/football-reporter/` in this repo, applied to
+  one space by `scripts/space-profile.ts apply|check|dump`. Not a
+  `resources/profiles/` built-in, not `mercury profiles apply`, not a profile
+  in `tagula-agent-profiles`.
+- **Alternatives considered:** (a) a built-in profile under
+  `resources/profiles/` applied with `mercury profiles apply` — rejected
+  because `applyProfile` (`src/core/profiles.ts`) writes
+  `.mercury/global/AGENTS.md` and pins `member_permissions` and
+  `profile_prompt` **project-wide**, and `~/whatsapp-bot` hosts several
+  spaces; (b) `profiles/football-reporter/` in `tagula-agent-profiles`,
+  deployed by its `deploy.sh` — rejected because the football bot is the
+  personal `~/whatsapp-bot`, feed-watch is already a mercury backlog item, and
+  tagula's loader is `deploy.sh` copying files by hand, which
+  `~/sync-mercury.sh` can do equally well; (c) extend mercury with
+  `mercury profiles apply --space <id>` — real upstream value but scope creep
+  for this goal.
+- **Reasoning:** the value is one owner for the rules plus a drift check, not
+  the profile machinery. A folder + script gives both today with zero core
+  change. `examples/` mirrors the bundled-but-optional convention of
+  `examples/extensions/`.
+- **Revisit if:** upstream adds per-space profile application — then the
+  folder becomes a manifest and the script goes away.
+- **Used by:** football-reporter-space-profile, daily-article-redesign
+
+## D-002: Profile contents — the shape of a tagula profile, none of the plumbing
+- **Category:** scope
+- **Decided:** the folder holds `README.md` (the layer map), `AGENTS.md`
+  (the editorial standard, Hebrew), `tasks/*.md` (task prompt sources),
+  `config.yaml` (per-space extension config: feed-watch, longview) and
+  nothing else. Skipped: `member_permissions`, `capabilities`, engines, config
+  registries, Drive, trial gates, `env`.
+- **Alternatives considered:** full `mercury-profile.yaml` manifest for
+  parity with tagula; rejected — none of its fields apply to a single personal
+  space.
+- **Reasoning:** "without the commercial bullshit" (owner). Every file in the
+  folder must be something the bot's behaviour observably depends on.
+- **Revisit if:** a second space wants the same profile — then a manifest with
+  `{SPACE}` substitution earns its place.
+- **Used by:** football-reporter-space-profile
+
+## D-003: Layer placement for football (applies `whatsapp-bot-hardening` D-014)
+- **Category:** pattern
+- **Decided:** editorial standard (role, sourcing, freshness, format, silence,
+  honesty about writes) → `spaces/football-friends/AGENTS.md`; *procedure*
+  (when to run, what artefact to produce, where to publish, how to dedupe) →
+  the task prompt; **zero editorial** preferences — `club-coverage` moves into
+  `AGENTS.md` and is deleted, `push-rules` is already gone; the three
+  *behaviour* preferences the live space carries (`language`, `timezone`,
+  `tone` — read 2026-08-20) are not editorial and stay; global character and
+  the baked prompt untouched. Every quantity is countable ("0 emoji",
+  "1–4 lines", "1 named source", "date within 24 h").
+- **Alternatives considered:** keep `club-coverage` as the member-visible
+  knob; rejected — it is the layer that produced the F1 contradictions and
+  its 500-char cap is what pushed rules into task prompts.
+- **Reasoning:** audit F1: "a preference may reinforce a task prompt, but must
+  never state a different threshold than one" — the safest way to satisfy that
+  is to have none. `AGENTS.md` is read per run, no rebuild.
+- **Revisit if:** a member-editable knob is wanted — then one preference with
+  a single countable value, never a rule that exists elsewhere.
+- **Used by:** football-reporter-space-profile, feed-watch
+
+## D-004: Freshness is enforced by a date, not an adjective
+- **Category:** pattern
+- **Decided:** every reported item carries the publication date the agent
+  read from the source; an item without a date or outside the window (scan:
+  since the last update; article: 24 h) is dropped; a transfer is reported as
+  live only after checking the player's current club. Feed-watch supplies
+  `publishedAt` per digest item; the harness asserts a date token per item.
+- **Alternatives considered:** "only current news" wording; rejected — audit
+  F4 showed a months-old real article satisfies every attribution rule as
+  written.
+- **Reasoning:** borrowed from tagula's date-grounding rule ("without
+  reading the engine you have no clock"): a date is checkable, "is this
+  current?" is not.
+- **Revisit if:** sources stop carrying dates — then the host timestamps
+  first-seen and that becomes the date.
+- **Used by:** football-reporter-space-profile, feed-watch,
+  football-reporter-fixture-harness
+
+## D-005: Keep Telegraph, make the summary fuller, and redesign the article itself (decided 2026-08-20)
+- **Category:** scope
+- **Decided:** the 09:00 article stays a Telegraph page, but (a) the chat
+  message becomes a fuller summary — headline plus the key facts with sources,
+  cap 900 chars for this space — and (b) the page is redesigned to read and
+  look like a top-tier sports article: headline, standfirst, dateline,
+  per-club prose sections, at most one real pull-quote, a "certain / in
+  doubt" box, a sources footer; accurate, concise, intriguing through the
+  facts, designed for readability, and with **no clickbait or commercial
+  dressing** (a forbidden-word list in the prompt, asserted by the harness).
+- **Alternatives considered:** move the article in-chat and disable
+  `longview` for the space (the 2026-08-20 proposal); status quo.
+- **Reasoning:** owner's call. The group's complaint (audit F3) was about
+  teasing and quality, which the fuller summary and the redesign address
+  directly; a page still fits an article's length better than a chat
+  bubble, and the in-chat full answer (#1905) shape is kept as the *summary*.
+- **Revisit if:** two weeks of reactions still read the link as a tease —
+  then in-chat (the proposal) is the next step, one config value away.
+- **Used by:** daily-article-redesign
+
+## D-010: The chat summary is author-supplied by the same run, not a second model call
+- **Category:** architecture
+- **Decided:** `longview` honours a leading `[longview:summary] … [/longview:summary]`
+  block in the reply: used verbatim as the chat text (≤ `summary_max_chars`,
+  new per-space key, default 300 = today's `BLURB_MAX_CHARS`), stripped from
+  the published page, `summarize()` skipped. No block → today's path.
+- **Alternatives considered:** raise the blurb cap and reword
+  `prompts/summarize.md` only (a second prompt to drift from `AGENTS.md`,
+  a second model call per article, and a summary that has not read the
+  sources); a host-side extractive summary (loses the editorial split).
+- **Reasoning:** one owner for the shape (D-003) — the run that verified the
+  sources writes both the page and the summary under the same `AGENTS.md`;
+  deterministic, cheaper, and behaviour-preserving for every other space.
+- **Revisit if:** the model misses the block often enough that the fallback
+  blurb is what the group usually sees — the harness reports the rate.
+- **Used by:** daily-article-redesign
+
+## D-011: Source policy — verified feeds, journalist-over-masthead tiers, a five-step grade ladder
+- **Category:** scope
+- **Decided:** (1) feed-watch's initial `sources` for the space are the feeds
+  verified on 2026-08-20 in `docs/notes/football-sources-and-style-research.md`
+  §1 (BBC football, Sky 11095, ESPN soccer, Marca per club, Daily Briefing,
+  ONE, Walla 156/316, ynet sport) plus one Google News RSS query per Hebrew
+  term and one per English term with `when:1d`; feeds that were blocked from
+  the planning machine (MEN, manutd.com, Guardian, Athletic, Transfermarkt)
+  join only after a fetch test from WSL. (2) `AGENTS.md` carries the source
+  tiers: official channel > named Tier-1 journalist > serious outlet; the
+  Tier-4/5 list (Sun, Mirror, Star, Express, Mail, 90min, Caught Offside,
+  Football Insider, TeamTalk, Sport Witness, Don Balón, Fichajes, El
+  Chiringuito, Bleacher Report, aggregators) is never cited; the journalist
+  outranks the masthead. (3) The certainty vocabulary becomes the five-step
+  ladder רשמי / סגור (here we go) / הסכמה / במשא ומתן / שמועה, each tied to
+  the wording the source used; hedged filler ("לפי דיווחים", "ייתכן ש") is
+  banned — *either we know it and report it, or we don't report it.*
+- **Alternatives considered:** keep the live prompt's free-text source list
+  and four grades; Brave search as the discovery layer (today's cost and
+  staleness problem).
+- **Reasoning:** the research note: three independent tier guides agree on
+  the shape; Romano's own definition of "here we go" and the BBC's
+  "understands" ruling give the grade ladder and the no-hedging rule an
+  external anchor; every ✅ feed was fetched and shown to carry `pubDate`,
+  which D-004 depends on.
+- **Revisit if:** a ✅ feed goes stale or changes shape (feed-watch's
+  `source_status` backoff will show it), or a new Tier-1 journalist emerges
+  for one of the seven scopes — the list is config, not code.
+- **Used by:** football-reporter-space-profile, feed-watch,
+  daily-article-redesign, football-reporter-fixture-harness
+
+## D-006: Scheduled research runs post from the primary model leg or not at all
+- **Category:** trade-off
+- **Decided:** a scheduled task may opt out of the fallback chain
+  (primary-or-skip); a skipped run is recorded on the task row and in the
+  ledger, not posted to the group. Marking leg-2 output in the message is the
+  fallback design if opt-out proves impossible.
+- **Alternatives considered:** (2) mark fallback output in the message;
+  (3) log a warning only (audit F2's ordered options).
+- **Reasoning:** "a missed scan is strictly better than a fabricated one";
+  `mercury.yaml` itself documents leg-2 as bad at exactly the verification
+  job the scan is.
+- **Revisit if:** the primary leg's availability drops enough that skips
+  become the norm — then the chain needs a second research-grade leg.
+- **Used by:** scheduled-task-model-leg-policy
+
+## D-007: Bugs run on a parallel track, never inside a story
+- **Category:** pattern
+- **Decided:** defects found before or during the goal are filed with
+  `/f-bug-report` and fixed with `/f-bug-fix`; a story never absorbs a fix.
+  Filed at planning: F5 (reply-to-member triggers the bot), F7a (HTML tags not
+  stripped), F7d (`_(responded in Ns)_` not stored), F7e (bot self-name is a
+  phone number). F7c (emoji) is prompt compliance → M1.1 + harness; F6 →
+  M1.1 honesty rule; F4/F8 → M1.1/M2.1.
+- **Alternatives considered:** one "small fixes" story; rejected — it hides
+  root causes and breaks the post-mortem trail.
+- **Reasoning:** owner's instruction; matches the refactor invariant in
+  `CLAUDE.md` (file, don't fix in place).
+- **Revisit if:** never.
+- **Used by:** football-regression-suite
+
+## D-008: Proof of quality = deterministic checks + fixtures + a ledger; measure before gating
+- **Category:** pattern
+- **Decided:** (1) any rule that can be a host-side gate or a script is one
+  (output lint: HTML, emoji, line count, source + date tokens); (2) prompts are
+  exercised against dated fixtures in the private `main` space, never the
+  group; (3) every scheduled run writes a ledger row (leg, cost, outcome,
+  item count); (4) numbers are reported for two weeks before any of them
+  becomes a threshold.
+- **Alternatives considered:** rely on the re-audit alone; rejected — it is
+  a 128-row read that happens once.
+- **Reasoning:** tagula's "measure min/median/max over 16+ seeds before
+  writing a number an agent will branch on" and the `normalizeChatMarkdown`
+  precedent ("a prose rule relies on the model complying").
+- **Revisit if:** the ledger shows the lint never fires for a month — then
+  the lint can be relaxed to a warning.
+- **Used by:** football-reporter-fixture-harness, scheduled-run-ledger,
+  football-regression-suite
+
+## D-009: feed-watch is absorbed as written; only its bundled ops move
+- **Category:** scope
+- **Decided:** `docs/backlog/feed-watch.md` becomes M2.1 unchanged in
+  architecture. Its "bundled space work" (the football `AGENTS.md`, the
+  workspace cleanup, the task-prompt rewrite) is owned by M1.1 and must be
+  live **before** feed-watch is enabled for the space (its own ops order
+  already requires this).
+- **Alternatives considered:** re-plan feed-watch inside the profile;
+  rejected — the spec is complete and its prerequisite
+  ([PR #41](https://github.com/Avishai-Tsabari/mercury/pull/41) `--timezone`)
+  is on `dev`.
+- **Reasoning:** don't re-derive a finished plan.
+- **Revisit if:** M1 changes the article's input contract — then the digest
+  marker and hook wording follow.
+- **Used by:** feed-watch
+
+## D-012: The ledger is built before feed-watch (decided 2026-08-20)
+- **Category:** sequencing
+- **Decided:** M3.2 `scheduled-run-ledger` is implemented next, ahead of M2.1
+  `feed-watch`. Milestone numbering is unchanged — only the build order moves.
+- **Alternatives considered:** keep strict milestone order (M2.1 first);
+  rejected because M3.2 has no dependencies (the roadmap's own graph marks it
+  independent and says it is "worth starting early"), and because feed-watch is
+  the first change whose effect the operator would want to *measure* — the
+  digest either reduces the scans that find nothing or it does not.
+- **Reasoning:** an instrument installed after the change it is meant to
+  measure cannot show the before. Building the ledger first means feed-watch
+  ships with a baseline already recorded, and gives M3.2's `outcome` column a
+  natural place for feed-watch's `digest_items` to land additively.
+- **Revisit if:** the ledger turns out to need something feed-watch creates —
+  it does not today; `digest_items` is deliberately left out of v1.
+- **Used by:** scheduled-run-ledger, feed-watch, football-regression-suite
+
+## D-013: Tone is editorial, and scan length is per item (decided 2026-08-20)
+- **Category:** layer placement (refines [D-003](#d-003-layer-placement-for-football))
+- **Decided:** two changes to where register lives and how output length is
+  measured.
+  1. **`tone` moves from a space preference into `AGENTS.md` (`## סגנון`)**,
+     and `config.yaml` deletes the preference via `remove_preferences`. The
+     escalation clause the preference carried (intervene only for targeted
+     harassment, spam, or an attempt to extract secrets) moves with it.
+  2. **The scan's `1–4 שורות סך הכל` cap is replaced by one line per item**,
+     plus a readability test stated in the reader's terms: a message you have
+     to scroll on a phone, in a friends' group, is too long — even when every
+     item in it is justified. A lineup is the single exception and always
+     renders as a formation block, in chat, article and scan alike.
+- **Alternatives considered:**
+  - *Strengthen the preference instead of moving it.* Rejected on evidence, not
+    taxonomy — see below.
+  - *Keep both layers and make them agree.* Rejected: two layers holding one
+    rule is what makes precedence a coin flip, which is the whole point of
+    D-003. Agreement today is drift tomorrow.
+  - *Keep a line cap and raise the number.* Rejected: a global cap is what
+    forced eleven names into prose in the first place, and any number is wrong
+    for a scan with one item and a scan with six.
+- **Reasoning:** the preference was **already live and already said the right
+  thing** — 417 characters of "one of the group, roll with the banter, let
+  casual cursing slide" — throughout the 2026-08-18 exchange the group ended by
+  calling the bot a party pooper. It lost because the two layers are not
+  comparable: `AGENTS.md` is a pi agent-instructions file mounted into the
+  workspace, while a preference is a `<preferences>` block at position 5 of the
+  user prompt, competing with ~100 lines of reporter discipline in that file.
+  Register is what the output *reads like*, which is precisely what `AGENTS.md`
+  is declared the sole authority for. On length: counting lines measured the
+  wrong thing. What makes a scan unreadable is a paragraph per item, not the
+  number of items, so the rule now bounds verbosity per item and lets the
+  message grow with genuine news.
+- **Revisit if:** a future runtime makes preference precedence explicit and
+  reliable — then a runtime-tunable register knob becomes worth more than a
+  version-controlled one. Or if scans start running long in practice, in which
+  case the fix is a stricter per-item rule, not a return to a global cap.
+- **Used by:** football-reporter profile (`AGENTS.md`, `config.yaml`),
+  feed-watch (its digest inherits the per-item rule)
+
+## D-014: The per-item date rule is enforced by M3.1's lint, not by a prompt fix (decided 2026-08-21)
+- **Category:** proof of quality (refines [D-004](#d-004-freshness-is-enforced-by-a-date-not-an-adjective))
+- **Decided:** the standard's requirement that every item carry a publication
+  date stays exactly as written in `AGENTS.md`. Nothing in the prompts is
+  changed to chase it. The **M3.1 lint becomes the enforcement**, and it must
+  cover **both** surfaces — the scan reply *and* the article's
+  `[longview:summary]` block — not the scan alone.
+- **Evidence at decision time:** two independent live observations, same
+  `AGENTS.md`, both missing dates.
+  1. Scan, 2026-08-20 16:31 — 0 of 3 items dated; item 1 used היום, the
+     adjective D-004 exists to forbid.
+  2. Article summary, 2026-08-21 09:00 (ledger run #2) — 0 date tokens across
+     4 lines, while the article *body* of the same run carried 5.
+- **Alternatives considered:**
+  - *File it as a bug on the parallel track (D-007).* Rejected: a bug fix here
+    would be a prompt edit, and a prompt edit is precisely the unverifiable
+    change this milestone exists to stop making. The defect is real but the
+    remedy is a check.
+  - *Tighten the task prompts now, as an M1 follow-up.* Rejected for the same
+    reason, plus it would re-open a milestone whose stories are done.
+- **Reasoning:** the rule is stated once, in the layer D-003 made its owner, and
+  it is still ignored — so the gap is not where the rule lives but that nothing
+  measures it. A third prompt rewrite would be the third unmeasured attempt.
+  The body-versus-summary split in run #2 is the sharpest evidence yet that this
+  is procedure drift rather than a capability limit: the same run, on the same
+  model, dated the long form and not the short one.
+- **Revisit if:** the lint lands and the rule still fails on fixtures — that
+  would move it from procedure drift to a capability limit and make a prompt or
+  model change the honest answer.
+- **Used by:** football-reporter-fixture-harness (M3.1), football-regression-suite (M4.1)
+
+## D-015: M2.2 is deferred until after M3.1, not dropped (decided 2026-08-21)
+- **Category:** sequencing (refines [D-006](#d-006-scheduled-research-runs-post-from-the-primary-model-leg-or-not-at-all))
+- **Decided:** `scheduled-task-model-leg-policy` stays in Milestone 2 and stays
+  unstarted. Its architecture is deliberately left unfilled. Revisit at the M2
+  checkpoint, with M3.1 in hand.
+- **Alternatives considered:**
+  - *Drop it.* Rejected: the Opus pin that pre-empts it is a per-space DB row
+    that one hand `DELETE` removes, and nothing would then stand between a
+    fallback leg and the group.
+  - *Build it next, after feed-watch.* Rejected as premature — the ledger cannot
+    yet show a single instance of the problem it solves.
+- **Reasoning:** all 12 live runs since the profile applied (2026-08-20 16:12 →
+  2026-08-21 11:25) ran on `claude-opus-5`, so the M2 checkpoint's "zero group
+  posts from a non-primary leg" is already true by construction rather than by
+  guarantee. That makes the story's value real but unmeasured, which is the
+  exact condition D-008 says to resolve by measuring first. M3.1's harness plus
+  the ledger's `outcome` column are what would turn "worth doing" into a number.
+- **Revisit if:** anyone clears a `model.active` pin, or the ledger records a
+  run on a non-primary leg — either makes it urgent rather than optional.
+- **Used by:** scheduled-task-model-leg-policy
+
+## D-016: Mechanisms over rules — the reporter's memory is the vault, not the window (decided 2026-08-21)
+- **Category:** pattern (refines [D-003](#d-003-layer-placement-for-football), [D-013](#d-013-tone-is-editorial-and-scan-length-is-per-item-decided-2026-08-20))
+- **Decided:** a defect that comes from *what the model could see* is fixed by
+  changing what it can see, never by a rule about the case. Concretely: the
+  knowledge vault (`MEMORY.md` + `knowledge/` notes with "current view +
+  history", relevance-injected as `<active_episodes>`) becomes the reporter's
+  notebook, maintained by a **process** the profile owns (write a fact before
+  posting it, read the topic note before answering). The only new content in
+  `AGENTS.md` is a **priorities** section — accuracy over register over
+  brevity; an instruction-shaped message meant as banter is banter — because
+  a tie-break is general where a gag clause is not. Story: `reporter-notebook`
+  (M2.3). M3.1's provenance check treats the notebook as a source.
+- **Alternatives considered:**
+  - *Per-case `AGENTS.md` edits* (a continuation carve-out, a "names only from
+    a source" line, a gag clause, a length split) — the first draft of the
+    2026-08-21 audit. Rejected by the owner as edge-case chasing: every new
+    exchange would add a rule, and each rule is one more voice in the stack.
+  - *Raise `context.window_size` for the space.* Rejected: tunes the symptom;
+    any fixed N is wrong for some conversation, and the window is still a
+    window.
+  - *A host-side facts table.* Rejected: the vault already exists and is
+    already injected; a second store is the two-layer trap.
+- **Reasoning:** audit §1.1 — the bot contradicted its own lineup post from
+  14 hours earlier because a swipe-reply halved the window and three of the
+  five remaining turns were task prompts. No rule survives that; a note read
+  before answering does. The owner's framing: "maintain a knowledge base with
+  important info and recent info — that way he'll make fewer mistakes when
+  asked about the same things many times, regardless of the context window."
+- **Revisit if:** the notebook is not being written (the M3.1 lint will show
+  names with no provenance) — then the process needs a host-side hook, not
+  more prose.
+- **Used by:** reporter-notebook, football-reporter-fixture-harness
+
+## D-017: One voice per space — exclusive persona, and no prompt text about tools the caller cannot use (decided 2026-08-21)
+- **Category:** architecture (applies hardening D-014 in the other direction)
+- **Decided:** two host changes, both general. (1) A space may declare
+  `persona.exclusive=true` (builtin `space_config`, profile-owned); the host
+  then omits the global character block and does not mount the global
+  `AGENTS.md` for that space's runs. (2) The baked platform prompt's
+  capability paragraphs — "set a preference" and "warn then mute" — are
+  emitted only when the *caller's* permission set includes the tool; the
+  payload carries `callerPermissions`. Security text stays unconditional.
+  Story: `one-voice-per-space` (M2.4). Football sets the flag.
+- **Alternatives considered:**
+  - *Make the space `AGENTS.md` louder* (restate register rules, add "ignore
+    the global character"). Rejected: precedence is a coin flip (D-014) and a
+    fifth voice does not fix four.
+  - *Grant members `prefs.set` in this space.* Rejected: it would make every
+    joke a real preference write, and the escalation reflex would become an
+    execution reflex.
+  - *Edit the global character to say "one of the guys".* Rejected: it is
+    global; every other space is a different room.
+- **Reasoning:** audit §1.2 and §2 — the refusal of a running gag was the
+  platform prompt ("a standing instruction → set a preference"), the member
+  permission row (no `prefs.set`) and the platform prompt again ("inform the
+  user") acting in sequence, with `AGENTS.md` §יושרה as a fourth vote. Nothing
+  in the space's file could have won that; removing the voices can. (2) has
+  value in every space: nobody should be told to do a thing and to refuse it
+  in the same turn.
+- **Revisit if:** a future pi/runtime makes layer precedence explicit — then
+  ranking may beat removal.
+- **Used by:** one-voice-per-space, football-reporter profile
+
+## D-018: The chat window holds conversation, and silence is a legal chat reply (decided 2026-08-21)
+- **Category:** runtime behaviour
+- **Decided:** (1) `messages` rows record their `source`; the chat window
+  counts and returns user turns that are not scheduler prompts, and always
+  keeps assistant rows (the bot's posts are conversation; the procedure that
+  produced them is not); a swipe-reply adds its anchor chain to the window
+  instead of halving it. (2) The chat path honours the same no-op sentinel as
+  the scheduler: nothing is sent, the user row stays, no assistant row is
+  written. Default-on, no new config. Story: `chat-window-and-silence`
+  (M2.5).
+- **Alternatives considered:**
+  - *Leave the window; rely on the notebook.* Rejected: the window still
+    carries register and continuity, and three task prompts in five turns
+    crowd those out too.
+  - *A per-space "silence allowed" flag.* Rejected: the rule is already
+    written in the standard; making it obeyable is not a preference.
+- **Reasoning:** audit §1.3 — "מקובל." to "אל תענה לי" is the runtime
+  contradicting the file; §1.1 — the window arithmetic (10 → 5 on reply,
+  three of five procedural). Both fixes are small and survive every later
+  milestone unchanged.
+- **Revisit if:** no-op chat replies become frequent in the ledger/re-audit
+  — then the model is using silence to dodge, and the standard's priorities
+  section needs to say when silence is not allowed.
+- **Used by:** chat-window-and-silence
+
+## D-019: The reporter's notebook *is* napkin's vault — seeded by the profile, maintained nightly by napkin, written in-turn by the process (decided 2026-08-21)
+- **Category:** architecture (refines [D-016](#d-016-mechanisms-over-rules--the-reporters-memory-is-the-vault-not-the-window-decided-2026-08-21))
+- **Decided:** the football topic notes (clubs, competitions, sagas) are
+  **episodes** in napkin's own shape under `knowledge/episodes/` — the only
+  directory the host scores and injects (`buildEpisodeContext`), with
+  `type: episode`, `status`, `last_mentioned`, `mentions`, JSON `keywords`,
+  `summary`, `## Current State`, `## History`. Three writers, one file: the
+  profile **seeds** a note once (`workspace_seed`, write-if-absent — never
+  overwrites, a model-edited note is not drift); the in-turn process
+  (`## מחברת`) appends a fact **before** it is posted, so the same-day
+  follow-up sees it; napkin's nightly distillation updates `last_mentioned`,
+  `mentions` and `## Current State` from whatever was posted, and its
+  weekly consolidation ages episodes and writes `.memory-suggestions.md`.
+  `MEMORY.md` is seeded for standing facts only (coverage + the admin's
+  equal-weight ruling, cadence, register rules the group set, spelling rule,
+  open commitments) and stays model-owned. The profile may **extend** the
+  distiller with a per-space addendum (`knowledge/.napkin/distill.md`,
+  appended to the kb-distillation prompt when present — a small general
+  napkin feature) and never replaces it. No parallel schema, no host-side
+  facts table.
+- **Alternatives considered:**
+  - *M2.3 as first written* — `knowledge/clubs|competitions|transfers/` notes
+    maintained by the model alone, distillation out of scope. Rejected on
+    facts: those directories are never injected, and napkin was already
+    producing the notes (once it ran) with the right frontmatter.
+  - *Let napkin do it all, drop the in-turn process.* Rejected: distillation
+    sees chat only (what was posted, not what was fetched), and lags a day;
+    the 08-21 hallucination was a same-day follow-up.
+  - *Teach the distiller the football schema by forking its prompt.* Rejected:
+    the prompt is napkin's, shared by every space; an addendum file is
+    profile-owned and general.
+- **Reasoning:** `docs/notes/napkin-revisit-2026-08-21.md` §1–§3 — the first
+  real distillation of `football-friends` (10 episodes, 8 people, 8 references)
+  already held the 20.8 Hapoel XI, the return legs, Turiel's injury, the
+  18.8 admin ruling and a verified spelling list, all dated and sourced and
+  injectable as-is. Building a second store beside it is the two-layer trap
+  D-016 already names.
+- **Revisit if:** the nightly run grows too slow or too costly for the space
+  (today 5–11 min per day of chat on Opus, growing with vault size) — then
+  pin distillation to a cheaper leg or a tighter backfill window; or if the
+  M3.1 provenance check shows posted names that are in neither a fetched
+  body nor an episode — then the in-turn process needs a host-side hook.
+- **Implementation note (2026-08-21, when M2.3 shipped):** the seeded
+  `MEMORY.md` came out **smaller** than this decision described. The list above
+  named the cadence, the coverage topics, the spelling rule and "check where
+  the player is now" as `MEMORY.md` content, but `AGENTS.md` already owns all
+  four as rules, and the story's own non-negotiable forbids one rule in two
+  layers — both files reach the same prompt. The seed therefore carries only
+  what `AGENTS.md` does not: the group's dated decisions (the 18/08 two-source
+  ruling, the 15/08 no-correction-block rule, the unapplied 17/08 request),
+  open commitments, the verified spellings themselves (facts, not the rule
+  about checking them), and the index of standing notebooks. Also found while
+  implementing: `## Current State (as of …)` is not matched by the injector, so
+  every note napkin had written for this space was reaching the model as a
+  summary line with no facts —
+  `docs/bugs/episode-current-state-heading-suffix-drops-body.md`.
+- **Used by:** reporter-notebook (M2.3, re-cut), football-reporter-fixture-harness (provenance source = `episodes/`), napkin (distill addendum)

package/docs/goals/football-reporter-profile/goal.md

@@ -0,0 +1,91 @@
+# Football Reporter Profile
+
+**Status**: Active
+**Slug**: football-reporter-profile
+**Created**: 2026-08-20
+**Last updated**: 2026-08-21 (roadmap extended after the 2026-08-21 audit)
+
+---
+
+## Goal Statement
+
+Redefine the `football-friends` bot's behaviour as a dedicated *top football
+reporter* profile — without the commercial plumbing the business profiles in
+`tagula-agents` carry — and use that profile as the one owner of the editorial
+rules that are today scattered across preferences, two task prompts and an empty
+space `AGENTS.md`. Then put a deterministic feed in front of the model
+(feed-watch), prove the output quality with fixtures and a run ledger, fix
+unrelated bugs as they surface, and finally re-test every defect the
+2026-08-19 48-hour audit found (F1–F8) with dedicated checks while monitoring
+for new ones.
+
+The sequence the owner chose: (1) profile + editorial standard in one layer →
+(2) feed-watch → (3) fixture harness + ledger, with a parallel bug track for
+anything unrelated, then (4) regression suite + re-audit.
+
+## Constraints
+
+- **The live bot is the beta tester.** Every change runs on `dev` in
+  `~/whatsapp-bot` before it can be promoted; nothing is "done" until it has
+  been observed in `mercury service logs -f` and in the group's behaviour.
+- **Never test in the group.** Manual checks run in the private `main` space
+  (standing rule, `docs/backlog/feed-watch.md` Non-Negotiables).
+- **Prompt-layer facts** (`docs/notes/bot-configuration-surfaces.md`,
+  D-014 in `whatsapp-bot-hardening`): preference values are capped at 500
+  chars; `spaces/<id>/AGENTS.md`, preferences and `tasks.prompt` are read per
+  run with no restart; the baked system prompt and `mrctl` need an image
+  rebuild; precedence between layers is a coin flip, so a rule may live in one
+  layer only.
+- **Mercury's built-in profiles are project-wide.** `mercury profiles apply`
+  overwrites `.mercury/global/AGENTS.md` and pins `member_permissions` for
+  every space (`src/core/profiles.ts`). `~/whatsapp-bot` hosts several spaces,
+  so the football profile must be space-scoped.
+- **Cost baseline:** $36.98 / 7 days for the space (~$5.30/day), 4 of 10 scans
+  `NO_UPDATE` at ~$0.50 each (audit F8). Report numbers; measure two weeks
+  before writing any threshold.
+- Output language is Hebrew; the group has said twice it does not want
+  teasing links for the daily article (audit F3).
+
+## Preferences
+
+- Borrow the *shape* of a tagula profile — one `AGENTS.md` owner generated
+  from source, extension + config, apply/check script, fixture harness — and
+  skip `member_permissions`, engines, config registries, Drive, trial gates
+  (required: "without the commercial bullshit").
+- Feed-watch stays as architected in `docs/backlog/feed-watch.md`; this goal
+  absorbs it as Milestone 2 rather than re-planning it (preferred).
+- Bugs unrelated to the profile go through `/f-bug-report` → `/f-bug-fix` in
+  parallel and are never folded into a story (required — keeps stories scoped
+  and behaviour-preserving).
+- Deterministic first: a rule that can be a host-side gate or a script is not
+  left as a prose rule the model must comply with (preferred; the
+  `normalizeChatMarkdown` precedent).
+
+## Target Users
+
+- **Group members of `football-friends`** — want real, fresh, sourced news in
+  chat, in Hebrew, with nothing when nothing happened. (inferred from the audit
+  transcript)
+- **The owner / space admin** — wants one place to edit the bot's editorial
+  behaviour, a bill that tracks news rather than the clock, and a way to see
+  problems without reading 128 chat rows.
+- **The operator (same person)** — wants every previously reported defect to
+  have a test that would catch its return.
+
+## Success Criteria
+
+1. Every editorial rule for `football-friends` has exactly one home, and
+   `scripts/space-profile.ts check` proves the deployed `AGENTS.md` and task
+   prompts are byte-identical to the repo source (zero drift) — `club-coverage`
+   and `push-rules` preferences are gone.
+2. No scheduled post reaches the group without a named source and a publication
+   date, and no item older than its window does — verified by the fixture
+   harness and by the two-week re-audit.
+3. Empty scans cost nothing: the space's spend is reported per run in the
+   ledger and, measured over two weeks, tracks news rather than the clock
+   (today: ~$1/day confirming nothing happened).
+4. A run that landed on a fallback model leg never posts to the group
+   (audit F2).
+5. F1–F7 each have a dedicated test or harness check; the re-audit after
+   Milestone 4 finds no recurrence, and any new defect is visible in the run
+   ledger without reading the chat.