feat(sl-hunting): v5j + SLH-019; CI runs the SL Hunting tests; coverage floor 73%

[DoRmAmMu1997]

Same read, one basket against three

Both books read 23 Sep identically: sellers seated from yesterday's fall, a FLAT open (+0.099%, 23,329.00 → 23,352.15), so BUY to hunt them.

IH bought all three indices at the open, held through a dip where NIFTY and SENSEX kept rejecting while BankNIFTY climbed, and booked his target.

The agent opened three baskets:

#	open	entry	stop	lots	exit	NIFTY	BNF	basket
1	09:32	23,371.75	23,360.00	3	discretionary	−750.75	−423.00	−₹1,173.75
2	09:59	23,368.95	23,349.00	1	discretionary	+149.50	+721.50	+₹871.00
3	10:17	23,368.25	23,347.00	1	discretionary	+786.50	+295.50	+₹1,082.00

Day +₹779.25.

The reason trade 1 gave was false when it filled

Its exit said "booking the still-positive basket (+420)". Both legs lost at the fill: −₹1,173.75.

position_state is a point-in-time mark, read inside a turn whose median was 33–37 s across 21–23 Sep (one in ten over 45–53 s). Across the 31 exits that quoted a basket figure:

• median gap between cited and realised: ₹496.50
• no systematic bias (mean −₹14.63)
• four cited a profit and realised a loss — never the reverse

The one-way flips have a mechanism: an exit is usually triggered by price turning against you, and half a minute later you fill further into that turn.

IH said the rule out loud

"Look at the trade less. Just focus on the chart… the premiums are moving up and down a lot, so the chart looks better than the premiums right now."

His own line for being wrong was a price — "if it breaks down the closing price, the trade can go wrong."

Measured before writing it

All 130 discretionary exits walked forward (one print a minute) against holding the original bracket.

Holding the bracket is refused. Exiting made +404.80 points vs +213.25 for holding; exiting won 84–46.

But the basis of the exit matters:

exit reason leans on	n	exit − hold
price action only	94	+241.25
cites a rupee figure	36	−49.70

Rupee-citing exits are mostly profit bookings, so that could be winners-vs-losers in disguise. Controlled for outcome it holds in both:

class	price action	rupee figure
winners (per exit)	+3.28	+0.92
losers (per exit)	+2.08	−7.38

Cutting a loser on the rupee figure did worse than letting the stop fire.

v5j

Sits directly after v4f and tells the agent how to read v4f's trigger rather than replacing it: "the move has stopped paying" is a fact about price, and the rupee figure is a late, noisy proxy for it.

• The check: strike the rupee figure out of the exit reason; what's left must still name a price. If all you can say is a number from position_state, you don't yet have an exit reason.
• Not a licence to exit less — stated outright, so the refused hold-the-bracket gate can't be re-derived from it.
• Limits in the rule: one print a minute, regex classification, only ten rupee-citing losers.

A correction

While comparing I first priced trade 1 "held to trade 2's exit" at +₹5,755.50 on the same strikes. That was wrong — NIFTY fell to 23,349.55, through trade 1's 23,360 stop, so holding would have been stopped at −11.75 points, worse than its −7.25 exit. The exit was right; the reason it gave was not.

The real difference from IH is where the line sat:

level	price
agent's stop	23,360.00
session low	23,349.55
IH's line — prior close	23,329.00

(Prior close from CPR AI's previous_day block; its pivot matches the agent's own 23,367.92.) Not encoded — a stop-distance floor was already priced negative at every level on 21 Sep.

Negative-tested 10 ways, all 10 caught, plus a passing control.

---

Addendum: SLH-019 — the mark is dated, and an EXIT reports what it booked

You approved both code follow-ups, so they ship here.

position_state is dated. With a position open it carries as_of — the moment its P&L was marked. The tool description now says what that means: a point-in-time mark, and one turn takes about half a minute, so any order fills later and at a different price.

An accepted EXIT reports what it actually booked:

field	what it is
realised_pnl	change in the worker's realized_pnl across the exit — both legs book there, so it's right for NIFTY, BNF and BOTH, and excludes the day's earlier trades
open_legs_after	any leg still open afterwards
exited_at	when the exit returned
mark_you_read	the position_state figure read earlier this turn, and how many seconds before the exit

23 Sep replayed through the order tool: +420 read at 09:50:06 → filled 09:50:34, seconds_before_exit 28.0, realised_pnl −1,173.75.

Honest in the two cases that matter: an unconfirmed live exit books nothing and keeps its leg open, so it reports 0.0 and names the leg; a worker with no numeric realized_pnl gets no figure at all rather than an invented zero. The order tool calls this under the same lock the worker's stop/target/square-off paths hold, so nothing books between the two readings.

What it can't do: the EXIT's reason is written before the fill and is the permanent journal record, so it can't be corrected. SLH-019 corrects the final decision reasoning (the decisions log). v5j now names the new fields and says the realised figure is the one to state.

Tests: 10 on fakes in the agent suite, plus 2 on the real SLHuntingAIWorker — a BOTH exit with NIFTY down and BankNIFTY up reports what both legs booked; NIFTY-only then BNF-only report each leg alone. Negative-tested 14 ways, all 14 caught; two of them (a one-leg exit reporting nothing booked) are caught only by the real-worker tests.

Found on the way: CI never ran the SL Hunting worker tests — now fixed here (third commit)

The master imports the SL Hunting modules only behind SL_HUNTING_ENABLED, which CI never set, so CI reported OK (skipped=62) while skipping every SL Hunting worker test — mirror, basket P&L, one-leg exits, cooldown, exit routing, and SLH-019's two real-worker tests. They only ran on your box, because its .env switches the agent on.

Gates: master 613 OK (skipped=60) as CI runs it, and 613 OK with SL_HUNTING_ENABLED=true · 28 market-data-health · 1594 pytest · ruff · mypy (80 files) · compileall · bandit.

---

Third commit: CI now runs the SL Hunting worker tests

	before (1ca9b80)	after (e3bfd9b)
master tests run	613	617
skipped	62	2
CI coverage	72.1%	73.8%

Both Python 3.12 and 3.13, on Linux — where these ~60 tests had never run before. All passed first time. The 2 remaining skips are the Shoonya/Kotak execution-layer tests, whose SDKs the quality job doesn't install by design.

The fix is in the test loader only. The suite already loads the master inside patch.dict(os.environ, {...}); SL_HUNTING_ENABLED=true is added there. The master reads the flag once at import, patch.dict restores the environment the moment the load ends, and the production default stays off. Because the master loads .env with override=False, CI, worktrees and your box now behave identically.

What the import needs — verified, not assumed. With claude_agent_sdk blocked, all five SL Hunting modules still import; with pydantic blocked, they fail. pydantic is pinned in requirements-ai.txt, which CI already installs, so the workflow is unchanged.

Truthful skip messages. Five skip sites — three of which blamed "missing" packages that were installed — now share one helper naming the real cause: master failed to load, flag off, or import failed.

Guards, because a silent skip is how this went unnoticed:

• the worker must load whenever pydantic is installed — so the original bug now fails the suite instead of reporting OK;
• the flag must not outlive the load;
• the skip reason must tell "flag off" from "import failed";
• the master's production default must stay False — pinning the tempting wrong fix.

Negative-tested 5 ways, all 5 caught, each by the test meant for it. Dropping the flag reproduces the original bug exactly (skipped=60) and now fails.

Coverage floor — raised to 73% in the fourth commit, below.

Gates (this commit): master 617 OK, 0 skipped under CI conditions · 28 market-data-health · 1594 pytest · coverage policy passed · ruff · mypy (80 files) · compileall · bandit.

---

Fourth commit: coverage floor 70% → 73%, docs level

At your direction, into the headroom the third commit created (CI 73.8%, identical on 3.12 and 3.13). The old 70.0 sat under a CI figure of 70.2% that was measured without the SL Hunting tests. New margin ~0.8pp — on the order of 230 statements and branches — so a PR adding materially untested code still goes red.

The floor is stated in several places, so they moved together:

file	change
pyproject.toml	fail_under 70.0 → 73.0; comment block rewritten (the 73.8% basis, why 70.2% was measured without the SL Hunting tests)
test_repository_policy.py	floor pin 70.0 → 73.0, with a ledger entry
scripts/check_coverage_thresholds.py	docstring "70% floor" → "73% floor"
CLAUDE.md / AGENTS.md	identical edits: 73% floor, CI measures 73.8%, and the master-suite note about the load-time flag
README.md, docs/lld/testing-and-ci.md	same floor and CI figure — nothing in the repo still says 70%

The local-vs-CI gap is now ~0.5 points, not ~2. CI skips 9 tests a fully-equipped machine runs (2 master, 7 pytest); local reads 74.3% vs CI's 73.8%. Most of the old 2-point gap was the SL Hunting tests CI wasn't running. The docs also no longer blame the broker SDKs alone — scikit-learn isn't in CI's install either.

Parity. AGENTS.md and CLAUDE.md were already identical below ## What this project is — enforced by test_agent_architecture_docs_stay_in_sync_…. The title and short intro above that marker differ on purpose (CLAUDE.md tells Claude which skills to load). Every edit went to both, and the parity test still passes.

Negative-checked 4 ways, all caught: reverting the floor to 70.0, quietly lowering it to 72.0, and editing either doc alone.

Gates (this commit): master 617 OK, 0 skipped · 28 market-data-health · 1594 pytest · coverage report exits 0 under fail_under = 73.0 · coverage policy passed · ruff · mypy · compileall · bandit.

---

Final addendum: the pre-open note for 24 SEP 2026

Source: Prediction For 24 SEP 2026 (4BKhLKn4zHY, uploaded 2026-09-23 21:25 IST).

The seated side flipped, and the flat branch with it. Tonight: positive momentum with only a small retracement, no big one (BankNIFTY held above 56500) — so buyers may be seated, and a flat open means SELL.

open	23 Sep	24 Sep
flat	BUY (hunt sellers)	SELL (hunt buyers)
gap down	small: BUY	SELL
gap up	BUY	BUY only if it opens above the prior high (NIFTY: especially above 23500)

The buy branch is a level, not a gap size. Above the "higher point", buyers stay confident and leave no SLs to hunt. That can disagree with the agent's own open classification (SLH-017's 0.5% threshold), so the note says outright that a computed GAP_UP verdict alone doesn't select the buy branch — and a gap up opening below the prior high is in neither branch. The test asserts that absence.

SENSEX has its expiry on 24 Sep.

One garble, resolved from the chart again. NIFTY's supports came through as "2360 2370" — neither obvious reading. The 1080p frame at 1:44 tags 23,369.60 and 23,271.75 (the second is the same drawn line as last night's), so 23370 / 23270. The resistances are last night's lines too (23,500.15 / 23,567.20), spoken as 23500 and 23570.

	resistance	support
NIFTY	23500 / 23570	23370 / 23270
BANKNIFTY	56800 / 57000	56370 / 56100
SENSEX	75100 / 75300	74650 / 74430

Negative-tested 12 ways, all 12 caught — including carrying 23 Sep's flat → BUY forward, turning the buy branch back into a gap-size rule, letting a GAP_UP verdict select it, and both natural misreadings of the garbled supports.

Gates (this commit): master 617 OK, 0 skipped · 28 market-data-health · 1594 pytest · ruff · mypy · compileall.

🤖 Generated with Claude Code